diff --git a/packages/core/src/sync/index.ts b/packages/core/src/sync/index.ts
index f08532afa14..ec9e3f5d864 100644
--- a/packages/core/src/sync/index.ts
+++ b/packages/core/src/sync/index.ts
@@ -26,6 +26,7 @@ import { kilo } from "./providers/kilo.js";
 import { llmgateway, llmgatewayProviders } from "./providers/llmgateway.js";
 import { mergeGateway } from "./providers/merge-gateway.js";
 import { nanoGpt } from "./providers/nano-gpt.js";
+import { neosantara } from "./providers/neosantara.js";
 import { openai } from "./providers/openai.js";
 import { ofox } from "./providers/ofox.js";
 import { openrouter } from "./providers/openrouter.js";
@@ -150,6 +151,7 @@ export const providers: {
   "llmgateway-providers": SyncProvider<any>;
   "merge-gateway": SyncProvider<any>;
   "nano-gpt": SyncProvider<any>;
+  neosantara: SyncProvider<any>;
   ofox: SyncProvider<any>;
   openai: SyncProvider<any>;
   openrouter: SyncProvider<any>;
@@ -184,6 +186,7 @@ export const providers: {
   "llmgateway-providers": llmgatewayProviders,
   "merge-gateway": mergeGateway,
   "nano-gpt": nanoGpt,
+  neosantara,
   ofox,
   openai,
   openrouter,
@@ -209,6 +212,7 @@ export const groups = {
     "llmgateway-providers",
     "merge-gateway",
     "nano-gpt",
+    "neosantara",
     "ofox",
     "requesty",
     "openrouter",
diff --git a/packages/core/src/sync/providers/neosantara.ts b/packages/core/src/sync/providers/neosantara.ts
new file mode 100644
index 00000000000..8c1fc5d7a59
--- /dev/null
+++ b/packages/core/src/sync/providers/neosantara.ts
@@ -0,0 +1,316 @@
+import { z } from "zod";
+import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import { factorBaseModel, resolveModelMetadataBaseModel } from "./openrouter.js";
+
+// Neosantara serves a models.dev / LLM Gateway shaped catalog at this single public endpoint.
+const CATALOG_ENDPOINT = "https://api.neosantara.xyz/v1/catalog";
+const MIN_CONTEXT_WINDOW = 100_000;
+
+// Accepted reasoning_effort enum; efforts from the catalog are intersected with it defensively.
+const HOST_EFFORTS = ["none", "minimal", "low", "medium", "high", "xhigh", "max"] as const;
+const HOST_EFFORT_SET: ReadonlySet<string> = new Set(HOST_EFFORTS);
+
+const CatalogProvider = z
+  .object({
+    providerId: z.string().optional(),
+    vision: z.boolean().optional(),
+    tools: z.boolean().optional(),
+    reasoning: z.boolean().optional(),
+    reasoning_efforts: z.array(z.string()).optional(),
+    has_toggle: z.boolean().optional(),
+  })
+  .passthrough();
+
+// Per-token USD strings; "0" is a real price for free models (kept), unknown for cache/reasoning.
+const CatalogPricing = z
+  .object({
+    prompt: z.string().optional(),
+    completion: z.string().optional(),
+    internal_reasoning: z.string().optional(),
+    input_cache_read: z.string().optional(),
+    input_cache_write: z.string().optional(),
+  })
+  .passthrough();
+
+export const NeosantaraModel = z
+  .object({
+    id: z.string(),
+    name: z.string().optional(),
+    display_name: z.string().optional(),
+    created: z.number().optional(),
+    description: z.string().optional(),
+    family: z.string().optional(),
+    architecture: z.object({
+      input_modalities: z.array(z.string()),
+      output_modalities: z.array(z.string()),
+    }),
+    context_length: z.number().int().nonnegative().optional(),
+    pricing: CatalogPricing,
+    supported_parameters: z.array(z.string()).optional(),
+    structured_outputs: z.boolean().optional(),
+    base_model: z.string().optional(),
+    override_name: z.string().optional(),
+    providers: z.array(CatalogProvider).min(1),
+    deprecated: z.boolean().optional(),
+  })
+  .passthrough();
+
+export const NeosantaraCatalogResponse = z
+  .object({ data: z.array(NeosantaraModel) })
+  .passthrough();
+
+export type NeosantaraSourceModel = z.infer<typeof NeosantaraModel>;
+
+// One mapping per model carries this host's capability flags and reasoning efforts.
+function mapping(model: NeosantaraSourceModel) {
+  return model.providers[0];
+}
+
+// Canonical lab base models are supplied directly by the /v1/catalog endpoint, verified
+// against the local models/ directory before falling back to id-based resolution.
+export function resolveNeosantaraBaseModel(model: NeosantaraSourceModel | string) {
+  if (typeof model === "string") return resolveModelMetadataBaseModel(model);
+  if (model.base_model !== undefined) {
+    const verified = resolveModelMetadataBaseModel(model.base_model);
+    if (verified !== undefined) return verified;
+  }
+  return resolveModelMetadataBaseModel(model.id);
+}
+
+type ReasoningControls = NonNullable<SyncedFullModel["reasoning_options"]>;
+
+// Toggle controls carry a leading wire comment naming the off value (AGENTS.md requirement).
+const TOGGLE_HEADER = `# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+`;
+
+// Whether the catalog reported a caller-control surface for this reasoning model. A missing
+// `reasoning_efforts` is "unknown" (skip + report), NOT an affirmative always-on `[]`.
+function hasReasoningEfforts(model: NeosantaraSourceModel) {
+  return Array.isArray(mapping(model).reasoning_efforts);
+}
+
+// The catalog reports each model's real reasoning surface (resolved host-side from the model's
+// models.dev lab entry): [] = always-on (no caller control), ["none"] = on/off toggle, otherwise
+// the graded effort levels. When off is expressed via reasoning_effort = "none" alongside graded
+// levels, AGENTS.md strictly requires authoring only `effort` with `none` in values and NO toggle.
+export function neosantaraReasoningControls(
+  model: NeosantaraSourceModel,
+  _existing?: ExistingModel,
+): ReasoningControls {
+  const map = mapping(model);
+  const efforts = (map.reasoning_efforts ?? []).filter((effort) =>
+    HOST_EFFORT_SET.has(effort),
+  );
+  if (efforts.length === 0) return [];
+  if (efforts.length === 1 && efforts[0] === "none") return [{ type: "toggle" }];
+
+  const graded = efforts.filter((effort) => effort !== "none");
+  const values = efforts.includes("none") ? ["none", ...graded] : graded;
+  return [{ type: "effort", values: values as never }];
+}
+
+// Toggle controls carry a leading wire comment naming the off value (AGENTS.md requirement).
+// Dedicated headers document always-on relays (minimax-m2.7) and external effort baselines (muse-glimmer-30b).
+export function neosantaraReasoningHeader(
+  controls: ReasoningControls | undefined,
+  modelId?: string,
+) {
+  if (controls?.some((option) => option.type === "toggle")) {
+    return TOGGLE_HEADER;
+  }
+  if (modelId === "minimax-m2.7") {
+    return `# Always-on thinking: upstream Dahl forwards no reasoning_effort parameter.
+# Matches lab providers/minimax/models/MiniMax-M2.7.toml and peers OpenRouter/FastRouter/Cortecs.
+`;
+  }
+  if (modelId === "kimi-k2-thinking") {
+    return `# Dedicated thinking variant: always-on reasoning with no caller control.
+# Matches lab providers/moonshotai/models/kimi-k2-thinking.toml and peers OpenRouter/Vercel.
+`;
+  }
+  if (modelId === "muse-glimmer-30b") {
+    return `# Sources:
+# https://huggingface.co/meta-models/Muse-Glimmer-30B
+# Effort: reasoning_effort = low|medium|high|xhigh
+# Matches peers OpenRouter and Vercel AI Gateway.
+`;
+  }
+  return undefined;
+}
+
+// Image models output an image modality, are priced per image, and skip the token filters.
+function isImageModel(model: NeosantaraSourceModel) {
+  return model.architecture.output_modalities.includes("image");
+}
+
+// Per-token USD string -> USD per million tokens. `0` is preserved (free models are real).
+function price(value: string | undefined): number | undefined {
+  if (value === undefined) return undefined;
+  const number = Number(value);
+  return Number.isFinite(number) && number >= 0
+    ? Math.round(number * 1_000_000_000_000) / 1_000_000
+    : undefined;
+}
+
+// Cache/reasoning prices report `0` when unknown; never downgrade to a published zero.
+function nonZeroPrice(value: string | undefined): number | undefined {
+  const result = price(value);
+  return result !== undefined && result > 0 ? result : undefined;
+}
+
+const ALLOWED_MODALITIES = new Set(["text", "audio", "image", "video", "pdf"]);
+
+function modalities(values: string[]): string[] {
+  const mapped = values
+    .map((value) => value.toLowerCase())
+    .map((value) => (value === "file" ? "pdf" : value))
+    .filter((value) => ALLOWED_MODALITIES.has(value));
+  return [...new Set(mapped.length > 0 ? mapped : ["text"])];
+}
+
+// Modalities as served by a specific deployment: a mapping without vision must
+// not carry image/pdf/video input, regardless of what the model-level architecture
+// claims — attachment=false with image/pdf/video input is contradictory.
+export function deploymentModalities(model: NeosantaraSourceModel, vision: boolean | undefined) {
+  const baseInput = modalities(model.architecture.input_modalities);
+  if (vision !== false) {
+    return {
+      input: baseInput,
+      output: modalities(model.architecture.output_modalities),
+    };
+  }
+  const filtered = baseInput.filter(
+    (value) => value !== "image" && value !== "pdf" && value !== "video",
+  );
+  return {
+    input: filtered.length > 0 ? filtered : ["text"],
+    output: modalities(model.architecture.output_modalities),
+  };
+}
+
+// Image generators, or 100k+ context function-calling text models.
+export function meetsNeosantaraPublicFilter(model: NeosantaraSourceModel) {
+  return (
+    isImageModel(model) ||
+    ((model.context_length ?? 0) >= MIN_CONTEXT_WINDOW && mapping(model).tools === true)
+  );
+}
+
+export function shouldSyncNeosantaraModel(model: NeosantaraSourceModel) {
+  if (model.deprecated || resolveNeosantaraBaseModel(model) === undefined) return false;
+  if (!meetsNeosantaraPublicFilter(model)) return false;
+  if (isImageModel(model)) return true;
+  // Skip rather than throw: one missing price costs a single model, not the run.
+  if (price(model.pricing.prompt) === undefined || price(model.pricing.completion) === undefined) {
+    return false;
+  }
+  // A reasoning model with no reported effort surface is unknown, not always-on: skip it (and
+  // report it) rather than stamping `[]`. Explicit `[]` from the catalog is a real always-on set.
+  if (mapping(model).reasoning === true && !hasReasoningEfforts(model)) return false;
+  return true;
+}
+
+export function buildNeosantaraModel(
+  model: NeosantaraSourceModel,
+  existing: ExistingModel | undefined,
+): SyncedModel {
+  const baseModel = resolveNeosantaraBaseModel(model);
+  if (baseModel === undefined) {
+    throw new Error(`No canonical base model mapping for Neosantara model '${model.id}'`);
+  }
+
+  // context_length is 0 for image models; never author limit.context = 0 — inherit the base.
+  const limit = { context: model.context_length || undefined };
+  const baseModelOmit = existing?.base_model === baseModel ? existing.base_model_omit : undefined;
+
+  // Per-image pricing has no models.dev field, so cost is left unpublished and inherited.
+  if (isImageModel(model)) {
+    return factorBaseModel(
+      baseModel,
+      { status: model.deprecated ? "deprecated" : existing?.status },
+      limit,
+      baseModelOmit,
+    );
+  }
+
+  const map = mapping(model);
+  const reasoning = map.reasoning === true;
+  const inputCost = price(model.pricing.prompt);
+  const outputCost = price(model.pricing.completion);
+  if (inputCost === undefined || outputCost === undefined) {
+    throw new Error(`Missing token pricing for Neosantara model '${model.id}'`);
+  }
+
+  return factorBaseModel(
+    baseModel,
+    {
+      name: model.override_name,
+      reasoning,
+      reasoning_options: reasoning ? neosantaraReasoningControls(model, existing) : undefined,
+      interleaved: reasoning ? { field: "reasoning_content" as const } : undefined,
+      attachment: map.vision,
+      tool_call: map.tools,
+      structured_output: model.structured_outputs === false ? false : undefined,
+      modalities: deploymentModalities(model, map.vision),
+      status: model.deprecated ? "deprecated" : existing?.status,
+      limit,
+      cost: {
+        input: inputCost,
+        output: outputCost,
+        reasoning: reasoning ? nonZeroPrice(model.pricing.internal_reasoning) : undefined,
+        cache_read: nonZeroPrice(model.pricing.input_cache_read),
+        cache_write: nonZeroPrice(model.pricing.input_cache_write),
+      },
+    },
+    limit,
+    baseModelOmit,
+  );
+}
+
+async function fetchJson(url: string) {
+  const response = await fetch(url);
+  if (!response.ok) {
+    throw new Error(`Neosantara request failed: ${response.status} ${response.statusText}`);
+  }
+  return response.json();
+}
+
+export const neosantara = {
+  id: "neosantara",
+  name: "Neosantara",
+  modelsDir: "providers/neosantara/models",
+  authoritativeHeaders: true,
+  sourceID(model) {
+    if (model.deprecated) return undefined;
+    return meetsNeosantaraPublicFilter(model) ? model.id : undefined;
+  },
+  skippedNotice(ids) {
+    if (ids.length === 0) return [];
+    return [
+      `${ids.length} Neosantara models were not created because they lack a canonical \`models/\` entry to inherit, the catalog reported no usable token pricing, or a reasoning model did not report its effort surface.`,
+      `Skipped remote IDs: ${ids.map((id) => `\`${id}\``).join(", ")}`,
+      "Ensure the catalog provides a valid base_model or add a `models/<provider>/<model>.toml` entry to include them in the next sync.",
+    ];
+  },
+  async fetchModels() {
+    return fetchJson(CATALOG_ENDPOINT);
+  },
+  parseModels(raw) {
+    const data = NeosantaraCatalogResponse.parse(raw).data;
+    // An empty catalog would delete every model file; fail loudly instead of wiping the provider.
+    if (data.length === 0) {
+      throw new Error("Neosantara catalog returned no models");
+    }
+    return data;
+  },
+  translateModel(model, context) {
+    if (!shouldSyncNeosantaraModel(model)) return undefined;
+    const translated = buildNeosantaraModel(model, context.existing(model.id));
+    return {
+      id: model.id,
+      model: translated,
+      header: neosantaraReasoningHeader(translated.reasoning_options, model.id),
+    };
+  },
+} satisfies SyncProvider<NeosantaraSourceModel>;
diff --git a/packages/core/test/neosantara-sync.test.ts b/packages/core/test/neosantara-sync.test.ts
new file mode 100644
index 00000000000..4aa5f4f6b43
--- /dev/null
+++ b/packages/core/test/neosantara-sync.test.ts
@@ -0,0 +1,398 @@
+import { expect, test } from "bun:test";
+
+import {
+  buildNeosantaraModel,
+  deploymentModalities,
+  meetsNeosantaraPublicFilter,
+  neosantara,
+  neosantaraReasoningControls,
+  neosantaraReasoningHeader,
+  NeosantaraCatalogResponse,
+  resolveNeosantaraBaseModel,
+  shouldSyncNeosantaraModel,
+} from "../src/sync/providers/neosantara.js";
+
+// A /v1/catalog response in the models.dev / LLM Gateway shape the backend now emits.
+const catalogResponse = {
+  data: [
+    {
+      id: "gemini-3.7-flash",
+      name: "gemini-3.7-flash",
+      display_name: "Neosantara: Gemini 3.7 Flash",
+      created: 1_700_000_000,
+      description: "Fast reasoning model",
+      family: "google",
+      architecture: { input_modalities: ["text", "image"], output_modalities: ["text"] },
+      context_length: 1_000_000,
+      pricing: {
+        prompt: "0.000001500000",
+        completion: "0.000006000000",
+        internal_reasoning: "0",
+        input_cache_read: "0.000000150000",
+        input_cache_write: "0.000001500000",
+      },
+      supported_parameters: ["temperature", "tools", "tool_choice", "response_format", "reasoning"],
+      structured_outputs: true,
+      providers: [
+        {
+          providerId: "neosantara",
+          vision: true,
+          tools: true,
+          reasoning: true,
+          reasoning_efforts: ["none", "low", "medium", "high"],
+        },
+      ],
+      deprecated: false,
+    },
+    {
+      id: "gpt-oss-20b",
+      name: "gpt-oss-20b",
+      display_name: "Neosantara: Gpt Oss 20b",
+      created: 1_700_000_000,
+      description: "Open model, no reasoning on this host",
+      family: "openai",
+      architecture: { input_modalities: ["text"], output_modalities: ["text"] },
+      context_length: 131_072,
+      pricing: { prompt: "0.000000020000", completion: "0.000000080000" },
+      supported_parameters: ["temperature", "tools", "tool_choice"],
+      structured_outputs: true,
+      providers: [{ providerId: "neosantara", vision: false, tools: true, reasoning: false }],
+      deprecated: false,
+    },
+    {
+      id: "glm-4.7-flash",
+      name: "glm-4.7-flash",
+      display_name: "Neosantara: Glm 4.7 Flash",
+      created: 1_700_000_000,
+      description: "Free reasoning model",
+      family: "zhipuai",
+      architecture: { input_modalities: ["text"], output_modalities: ["text"] },
+      context_length: 1_000_000,
+      pricing: { prompt: "0.000000000000", completion: "0.000000000000" },
+      supported_parameters: ["temperature", "tools", "reasoning"],
+      structured_outputs: false,
+      providers: [{ providerId: "neosantara", vision: false, tools: true, reasoning: true, reasoning_efforts: ["none"] }],
+      deprecated: false,
+    },
+    {
+      id: "too-small",
+      name: "too-small",
+      display_name: "Neosantara: Too Small",
+      created: 1_700_000_000,
+      description: "Below the context floor",
+      family: "openai",
+      architecture: { input_modalities: ["text"], output_modalities: ["text"] },
+      context_length: 32_000,
+      pricing: { prompt: "0.000001000000", completion: "0.000002000000" },
+      supported_parameters: ["tools"],
+      structured_outputs: false,
+      providers: [{ providerId: "neosantara", vision: false, tools: true, reasoning: false }],
+      deprecated: false,
+    },
+    {
+      id: "no-tools",
+      name: "no-tools",
+      display_name: "Neosantara: No Tools",
+      created: 1_700_000_000,
+      description: "No function calling",
+      family: "openai",
+      architecture: { input_modalities: ["text"], output_modalities: ["text"] },
+      context_length: 128_000,
+      pricing: { prompt: "0.000001000000", completion: "0.000002000000" },
+      supported_parameters: [],
+      structured_outputs: false,
+      providers: [{ providerId: "neosantara", vision: false, tools: false, reasoning: false }],
+      deprecated: false,
+    },
+  ],
+};
+
+test("parses the single /v1/catalog endpoint", () => {
+  const parsed = NeosantaraCatalogResponse.parse(catalogResponse);
+  expect(parsed.data).toHaveLength(5);
+  expect(parsed.data[0]?.providers[0]?.reasoning_efforts).toEqual(["none", "low", "medium", "high"]);
+});
+
+test("filters to 100k+ context, function calling, image models, and known base models", () => {
+  const [gemini, oss, glm, tooSmall, noTools] = NeosantaraCatalogResponse.parse(catalogResponse).data;
+
+  expect(meetsNeosantaraPublicFilter(gemini!)).toBe(true);
+  expect(meetsNeosantaraPublicFilter(oss!)).toBe(true);
+  expect(meetsNeosantaraPublicFilter(glm!)).toBe(true);
+  expect(meetsNeosantaraPublicFilter(tooSmall!)).toBe(false); // < 100k context
+  expect(meetsNeosantaraPublicFilter(noTools!)).toBe(false); // no function calling
+
+  expect(shouldSyncNeosantaraModel(gemini!)).toBe(true);
+  expect(shouldSyncNeosantaraModel(tooSmall!)).toBe(false);
+  expect(shouldSyncNeosantaraModel(noTools!)).toBe(false);
+
+  expect(resolveNeosantaraBaseModel("gemini-3.7-flash")).toBe("google/gemini-3.7-flash");
+  expect(
+    resolveNeosantaraBaseModel({ id: "claude-4.5-sonnet", base_model: "anthropic/claude-sonnet-4-5" } as never),
+  ).toBe("anthropic/claude-sonnet-4-5");
+  expect(
+    resolveNeosantaraBaseModel({ id: "grok-code-fast", base_model: "xai/grok-4.3" } as never),
+  ).toBe("xai/grok-4.3");
+  expect(
+    resolveNeosantaraBaseModel({ id: "gemini-3.7-flash", base_model: "google/nonexistent-model" } as never),
+  ).toBe("google/gemini-3.7-flash");
+  expect(
+    resolveNeosantaraBaseModel({ id: "unknown-id", base_model: "nonexistent/fake-model" } as never),
+  ).toBeUndefined();
+  expect(resolveNeosantaraBaseModel("definitely-not-a-model")).toBeUndefined();
+});
+
+test("builds override-only models with per-million USD cost and host reasoning controls", () => {
+  const [gemini] = NeosantaraCatalogResponse.parse(catalogResponse).data;
+  const built = buildNeosantaraModel(gemini!, undefined);
+
+  expect(built).toMatchObject({
+    base_model: "google/gemini-3.7-flash",
+    // The catalog's per-model reasoning_efforts are copied verbatim (host resolves them from the
+    // model's lab entry), intersected with the host enum.
+    reasoning_options: [{ type: "effort", values: ["none", "low", "medium", "high"] }],
+    interleaved: { field: "reasoning_content" },
+    limit: { context: 1_000_000 },
+    cost: { input: 1.5, output: 6, cache_read: 0.15, cache_write: 1.5 },
+  });
+  // internal_reasoning "0" is unknown, not a published zero.
+  expect((built as { cost?: { reasoning?: number } }).cost?.reasoning).toBeUndefined();
+  // The gateway omits max_output; never author a model output limit from a runtime cap.
+  expect((built as { limit?: { output?: number } }).limit?.output).toBeUndefined();
+});
+
+test("keeps a free model's zero token price instead of skipping it", () => {
+  const glm = NeosantaraCatalogResponse.parse(catalogResponse).data.find((m) => m.id === "glm-4.7-flash")!;
+  expect(shouldSyncNeosantaraModel(glm)).toBe(true);
+
+  const built = buildNeosantaraModel(glm, undefined);
+  expect(built.cost).toMatchObject({ input: 0, output: 0 });
+  // reasoning matches the base (factored out), but its effort controls are still authored.
+  expect(built.reasoning_options).toBeDefined();
+});
+
+test("a non-reasoning model carries no reasoning controls, interleaved, or note", () => {
+  const oss = NeosantaraCatalogResponse.parse(catalogResponse).data.find((m) => m.id === "gpt-oss-20b")!;
+  const built = buildNeosantaraModel(oss, undefined);
+
+  expect(built.reasoning).toBe(false);
+  expect(built.reasoning_options).toBeUndefined();
+  expect(built.interleaved).toBeUndefined();
+
+  const translated = neosantara.translateModel(oss, {
+    existing: () => undefined,
+    authored: () => undefined,
+  });
+  expect(translated?.header ?? "").not.toContain("reasoning_effort");
+});
+
+test("copies the catalog's per-model reasoning surface verbatim (llmgateway convention)", () => {
+  const ctl = (efforts) =>
+    neosantaraReasoningControls({ providers: [{ reasoning: true, reasoning_efforts: efforts }] } as never);
+
+  // Exactly ["none"] -> on/off toggle.
+  expect(ctl(["none"])).toEqual([{ type: "toggle" }]);
+  // [] -> always-on (reasons, no caller control).
+  expect(ctl([])).toEqual([]);
+  // Graded levels are copied verbatim (the host resolved them from the model's lab entry),
+  // intersected with the host enum; out-of-enum values are dropped.
+  expect(ctl(["minimal", "low", "medium", "high"])).toEqual([
+    { type: "effort", values: ["minimal", "low", "medium", "high"] },
+  ]);
+  expect(ctl(["none", "high", "max"])).toEqual([{ type: "effort", values: ["none", "high", "max"] }]);
+  expect(ctl(["low", "high", "max", "bogus"])).toEqual([{ type: "effort", values: ["low", "high", "max"] }]);
+
+  // Graded levels with "none" author only effort with "none" in values, without toggle.
+  expect(
+    neosantaraReasoningControls({
+      providers: [{ reasoning: true, reasoning_efforts: ["low", "medium", "high", "xhigh", "max"] }],
+    } as never),
+  ).toEqual([
+    { type: "effort", values: ["low", "medium", "high", "xhigh", "max"] },
+  ]);
+
+  expect(
+    neosantaraReasoningControls({
+      providers: [{ reasoning: true, reasoning_efforts: ["none", "low", "high", "max"] }],
+    } as never),
+  ).toEqual([
+    { type: "effort", values: ["none", "low", "high", "max"] },
+  ]);
+
+  // Toggle is driven only by the catalog (has_toggle / ["none"]), not re-inferred from prior TOMLs.
+  expect(
+    neosantaraReasoningControls(
+      {
+        providers: [{ reasoning: true, reasoning_efforts: ["low", "high", "max"] }],
+      } as never,
+      { reasoning_options: [{ type: "toggle" }, { type: "effort", values: ["low", "high", "max"] }] } as never,
+    ),
+  ).toEqual([{ type: "effort", values: ["low", "high", "max"] }]);
+});
+
+test("skips a reasoning model whose effort surface is unknown (missing != always-on)", () => {
+  const source = NeosantaraCatalogResponse.parse(catalogResponse).data[0]!;
+  // Reasoning model with a real surface syncs.
+  expect(shouldSyncNeosantaraModel(source)).toBe(true);
+
+  // Same model but the catalog omits reasoning_efforts -> unknown -> skipped and reported,
+  // never stamped as always-on `[]`.
+  const unknown = { ...source, providers: [{ ...source.providers[0], reasoning_efforts: undefined }] };
+  const parsed = NeosantaraCatalogResponse.parse({ data: [unknown] }).data[0]!;
+  expect(shouldSyncNeosantaraModel(parsed)).toBe(false);
+  expect(neosantara.sourceID(parsed)).toBe(parsed.id);
+
+  // Explicit [] is different: it is an affirmative always-on set and still syncs.
+  const alwaysOn = { ...source, providers: [{ ...source.providers[0], reasoning_efforts: [] }] };
+  const parsedAlwaysOn = NeosantaraCatalogResponse.parse({ data: [alwaysOn] }).data[0]!;
+  expect(shouldSyncNeosantaraModel(parsedAlwaysOn)).toBe(true);
+  expect(buildNeosantaraModel(parsedAlwaysOn, undefined).reasoning_options).toEqual([]);
+});
+
+test("toggle models carry a wire-comment header; effort models carry none", () => {
+  const models = NeosantaraCatalogResponse.parse(catalogResponse).data;
+  const ctx = { existing: () => undefined, authored: () => undefined };
+
+  // glm-4.7-flash advertises ["none"] -> toggle -> header names the wire field.
+  const toggle = neosantara.translateModel(models.find((m) => m.id === "glm-4.7-flash")!, ctx);
+  expect(toggle?.header).toContain("reasoning_effort");
+  expect(toggle?.header).toContain("Toggle");
+  expect(toggle?.header).toContain("https://docs.neosantara.xyz/en/capability/reasoning");
+
+  // gemini-3.7-flash advertises graded effort -> no toggle header.
+  const effort = neosantara.translateModel(models[0]!, ctx);
+  expect(effort?.header ?? "").not.toContain("Toggle");
+
+  // minimax-m2.7 documents always-on reasoning.
+  expect(neosantaraReasoningHeader([], "minimax-m2.7")).toContain("Always-on thinking");
+  // kimi-k2-thinking documents dedicated thinking variant.
+  expect(neosantaraReasoningHeader([], "kimi-k2-thinking")).toContain("Dedicated thinking variant");
+  // muse-glimmer-30b documents external baseline sources.
+  expect(neosantaraReasoningHeader([{ type: "effort", values: ["low"] as never }], "muse-glimmer-30b")).toContain(
+    "https://huggingface.co/meta-models/Muse-Glimmer-30B",
+  );
+});
+
+test("syncs image-generation models on per-image pricing without token cost", () => {
+  const image = {
+    id: "gpt-image-2",
+    name: "gpt-image-2",
+    display_name: "Neosantara: Gpt Image 2",
+    created: 1_700_000_000,
+    description: "Image generation",
+    family: "openai",
+    architecture: { input_modalities: ["text"], output_modalities: ["image"] },
+    context_length: 0,
+    pricing: {},
+    supported_parameters: [],
+    structured_outputs: false,
+    providers: [{ providerId: "neosantara", vision: false, tools: false, reasoning: false }],
+    deprecated: false,
+  };
+  const parsed = NeosantaraCatalogResponse.parse({ data: [image] }).data[0]!;
+
+  expect(shouldSyncNeosantaraModel(parsed)).toBe(true);
+  const built = buildNeosantaraModel(parsed, undefined);
+  expect("base_model" in built && built.base_model).toBe("openai/gpt-image-2");
+  expect(built.cost).toBeUndefined();
+  expect(built.reasoning_options).toBeUndefined();
+  // No limit.context = 0 authored for a context-less image model.
+  expect((built as { limit?: { context?: number } }).limit?.context).toBeUndefined();
+});
+
+test("filters deprecated models and reports unmapped ids as skipped", () => {
+  const deprecated = { ...catalogResponse.data[0]!, deprecated: true };
+  expect(shouldSyncNeosantaraModel(NeosantaraCatalogResponse.parse({ data: [deprecated] }).data[0]!)).toBe(false);
+
+  const unmapped = { ...catalogResponse.data[0]!, id: "not-in-the-canonical-tree" };
+  const parsed = NeosantaraCatalogResponse.parse({ data: [unmapped] }).data[0]!;
+  expect(shouldSyncNeosantaraModel(parsed)).toBe(false);
+  expect(neosantara.sourceID(parsed)).toBe("not-in-the-canonical-tree");
+
+  const notice = neosantara.skippedNotice(["not-in-the-canonical-tree"]);
+  expect(notice.join("\n")).toContain("not-in-the-canonical-tree");
+  expect(neosantara.skippedNotice([])).toEqual([]);
+});
+
+test("an empty catalog is fatal rather than wiping every model", () => {
+  expect(() => neosantara.parseModels({ data: [] })).toThrow();
+});
+
+test("deploymentModalities strips non-text inputs when vision is false and preserves otherwise", () => {
+  const multimodal = {
+    architecture: {
+      input_modalities: ["text", "image", "video", "pdf"],
+      output_modalities: ["text"],
+    },
+  } as never;
+
+  expect(deploymentModalities(multimodal, false)).toEqual({
+    input: ["text"],
+    output: ["text"],
+  });
+
+  expect(deploymentModalities(multimodal, true)).toEqual({
+    input: ["text", "image", "video", "pdf"],
+    output: ["text"],
+  });
+
+  expect(deploymentModalities(multimodal, undefined)).toEqual({
+    input: ["text", "image", "video", "pdf"],
+    output: ["text"],
+  });
+});
+
+test("enforces text-only input and attachment=false when vision is false on multimodal base model", () => {
+  const model = {
+    id: "kimi-k3",
+    context_length: 1_048_576,
+    architecture: { input_modalities: ["text"], output_modalities: ["text"] },
+    pricing: { prompt: "0.000003000000", completion: "0.000015000000" },
+    providers: [{ providerId: "neosantara", vision: false, tools: true, reasoning: false }],
+  } as never;
+
+  const built = buildNeosantaraModel(model, undefined);
+  // moonshotai/kimi-k3 base model has attachment=true and input=["text", "image", "video"].
+  // Because vision is false on this host, attachment must be false AND input must be overridden to ["text"].
+  expect(built).toMatchObject({
+    base_model: "moonshotai/kimi-k3",
+    attachment: false,
+    modalities: { input: ["text"] },
+  });
+  // tool_call is true in both host and base model; factored out
+  expect("tool_call" in built).toBe(false);
+});
+
+test("emits narrower modalities when host deployment does not support pdf/video", () => {
+  const model = {
+    id: "claude-4.5-opus",
+    base_model: "anthropic/claude-opus-4-5",
+    context_length: 200_000,
+    architecture: { input_modalities: ["text", "image"], output_modalities: ["text"] },
+    pricing: { prompt: "0.000005000000", completion: "0.000025000000" },
+    providers: [{ providerId: "neosantara", vision: true, tools: true, reasoning: false }],
+  } as never;
+
+  const built = buildNeosantaraModel(model, undefined);
+  // anthropic/claude-opus-4-5 base model has attachment=true and input=["text", "image", "pdf"].
+  // Because vision is true, attachment is true matching base and factored out.
+  // Because host deployment only serves text+image (not pdf), narrower input modalities are authored.
+  expect("attachment" in built).toBe(false);
+  expect(built.modalities).toEqual({ input: ["text", "image"] });
+});
+
+test("omits modalities when host input modalities match base model", () => {
+  const model = {
+    id: "claude-4.5-opus",
+    base_model: "anthropic/claude-opus-4-5",
+    context_length: 200_000,
+    architecture: { input_modalities: ["text", "image", "pdf"], output_modalities: ["text"] },
+    pricing: { prompt: "0.000005000000", completion: "0.000025000000" },
+    providers: [{ providerId: "neosantara", vision: true, tools: true, reasoning: false }],
+  } as never;
+
+  const built = buildNeosantaraModel(model, undefined);
+  expect("attachment" in built).toBe(false);
+  expect("modalities" in built).toBe(false);
+});
+
diff --git a/providers/neosantara/logo.svg b/providers/neosantara/logo.svg
new file mode 100644
index 00000000000..b03dbb9aac1
--- /dev/null
+++ b/providers/neosantara/logo.svg
@@ -0,0 +1,11 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 250 250" fill="currentColor" style="shape-rendering:geometricPrecision;text-rendering:geometricPrecision;image-rendering:optimizeQuality;fill-rule:evenodd;clip-rule:evenodd">
+  <path fill="currentColor" stroke="currentColor" stroke-width=".5" d="M101.29 207.98c.47-.85 1.06-1.62 1.77-2.3 2.7-4.98 5.58-9.86 8.66-14.62 1.14-2.92 3.72-5.82 4.37-8.79 15.27-25.59 30.64-51.13 46.12-76.61.74-.98 1.23-2.69-.7-2.5-1.64.31-3.29.46-4.96.45-22.2-1.55-44.37-1.53-66.59-2.21-3.59-.16-6.51-1.65-8.77-4.46-.74-1.82-1.57-3.59-2.48-5.31-1.87-2.85-3.67-5.71-5.42-8.59-.16.1-.32.2-.48.29.49 1.24 1.01 2.48 1.55 3.73 2.1 3.71 4.23 7.4 6.41 11.06 13.67 23.69 26.09 48.41 39.08 72.52 1.29 2.72-3.71 8.56-4.69 11.29-4.18 7.94-8.79 15.6-13.84 23.01-.28.66-.45 1.34-.51 2.05-1.78.46-2.14 3.79-4.47 4.64a295 295 0 0 1-41.84 1.2c-5.18-.47-10.34-.94-15.49-1.43-8.27-13.5-16.05-27.6-22.79-41.94-2.13-7.89 9.21-22.88 12.17-30.49q.615-1.11.6-2.37c-3.6-6.58-26.23-44.79-26.21-48.61 4.24-8.95 10.87-19.14 14.52-28.05.99-2.28 2.72-3.24 5.2-2.87 5.97-.2 31.57-1.35 35.4 1.18 3.74 2.93 4.64 8.23 7.18 12.18.86 1.23 1.7 2.43 2.53 3.62q.615-.525.33-1.32c-2.24-4.15-4.43-8.34-6.58-12.55a59 59 0 0 0-2.26-2.47c-.34-2.24-.68-2.62.64-4.65 7.86-13.99 16.03-27.79 24.5-41.4 1.15-2.12 2.9-3.28 5.25-3.5 13.51.03 35.22-1.53 47.83 1.19 5.52 1.68 15.82 26.53 19.2 27.65 6.32.4 12.65.67 18.98.83 6.07.08 28.27-1.03 32.48.76 7.06 9.91 13.5 20.16 19.15 30.91.08.91-.13 1.75-.61 2.52q-8.37 15.42-17.25 30.51c1.04.47 2.09.96 3.15 1.46a901 901 0 0 1 25.6 43.57c1.57 3.51 1.35 6.89-.68 10.16a1235 1235 0 0 1-22.36 37.75c-.42.34-.91.55-1.45.62-7.43 1.25-25.9.67-31.74 2.22-9.06 13.98-18.04 28.01-26.95 42.09-.48.97-1.27 1.44-2.35 1.41-5.12.23-30.35 1.33-33.51-.42a228 228 0 0 1-19.26-29.34c-.74-.78 1.46-3.06 1.57-4.07Zm69.74-140.24c5.51-.02 10.99.31 16.46 1.02 1.31 0 1.58-.5.8-1.49-3.25-.3-6.51-.4-9.78-.31-2.17-.2-4.33-.49-6.48-.88-22.14-.87-51.81-9.52-72.69-17.85-2.03-.65-6.22-1.92-5 2.08 2.64 4.86 5.55 9.57 8.71 14.11.41.36.88.58 1.42.64 21.01.72 42.02 1.38 63.04 1.99 1.23-.06 2.41.17 3.52.69Zm-120.96 44.9c9.04 22.25 14.88 43.83 19.24 67.42.45 1.3 1.49 4.19 3.11 1.97 2.12-4.86 5.17-20.52 5.69-25.55-8.59-14.58-17.24-29.15-25.93-43.7-4.22-7.06-8.3-14.19-12.23-21.39-.82-.77-1.28.02-.94.86 3.8 6.4 7.2 13.01 10.2 19.84.22.28.51.46.86.55Zm103.6 74.01c-2.08 3.18-4.1 6.39-6.06 9.64-3.28 4.66-6.64 9.29-10.08 13.91-.45.56-.46 1.11-.05 1.66.65-.22 1.18-.6 1.59-1.16 2.11-2.93 4.12-5.93 6.02-9.02 1.01-1.34 3-2.34 2.73-4.28.44-.2.82-.48 1.15-.83 2.26-2.75 4.48-5.54 6.67-8.35 13.6-16.12 29.38-28.44 46.36-40.73 2.72-2.54 2.22-4.05-1.49-4.53-2.37-.16-17.06-1.08-18.33-.21-9.47 14.65-18.97 29.29-28.51 43.9Z" />
+  <path fill="currentColor" d="m147.61 196.29.21 1.11c.27 1.94-1.72 2.94-2.73 4.28-1.9 3.09-3.91 6.09-6.02 9.02-.41.56-.94.94-1.59 1.16-.41-.55-.4-1.1.05-1.66 3.44-4.62 6.8-9.25 10.08-13.91" />
+  <path fill="currentColor" d="M153.67 186.65c-.12.67-.24 1.34-.35 2.01.77-.15 1.55-.29 2.32-.44-2.19 2.81-4.41 5.6-6.67 8.35-.33.35-.71.63-1.15.83l-.21-1.11c1.96-3.25 3.98-6.46 6.06-9.64" />
+  <path fill="currentColor" d="M52.18 112.78c-.7-.05-1.41-.09-2.11-.14-.35-.09-.64-.27-.86-.55-3-6.83-6.4-13.44-10.2-19.84-.34-.84.12-1.63.94-.86 3.93 7.2 8.01 14.33 12.23 21.39" />
+  <path fill="currentColor" d="M172.03 66.08c2.15.39 4.31.68 6.48.88 3.27-.09 6.53.01 9.78.31.78.99.51 1.49-.8 1.49-5.47-.71-10.95-1.04-16.46-1.02.33-.55.67-1.1 1-1.66" />
+  <path fill="currentColor" stroke="currentColor" stroke-width=".5" d="M172.03 66.08c-.33.56-.67 1.11-1 1.66a7.35 7.35 0 0 0-3.52-.69c-21.02-.61-42.03-1.27-63.04-1.99-.54-.06-1.01-.28-1.42-.64-3.16-4.54-6.07-9.25-8.71-14.11-1.22-4 2.97-2.73 5-2.08 20.88 8.33 50.55 16.98 72.69 17.85ZM50.07 112.64c.7.05 1.41.09 2.11.14 8.69 14.55 17.34 29.12 25.93 43.7-.52 5.03-3.57 20.69-5.69 25.55-1.62 2.22-2.66-.67-3.11-1.97-4.36-23.59-10.2-45.17-19.24-67.42Zm105.57 75.58c-.77.15-1.55.29-2.32.44.11-.67.23-1.34.35-2.01 9.54-14.61 19.04-29.25 28.51-43.9 1.27-.87 15.96.05 18.33.21 3.71.48 4.21 1.99 1.49 4.53-16.98 12.29-32.76 24.61-46.36 40.73Z" />
+  <path fill="currentColor" fill-opacity=".365" stroke="currentColor" stroke-width=".5" stroke-opacity=".365" d="M115.16 181.93c.31.12.62.23.93.34-.65 2.97-3.23 5.87-4.37 8.79-3.08 4.76-5.96 9.64-8.66 14.62-.71.68-1.3 1.45-1.77 2.3l-.48-.99c.06-.71.23-1.39.51-2.05 5.05-7.41 9.66-15.07 13.84-23.01Z" />
+  <path fill="currentColor" fill-opacity=".502" stroke="currentColor" stroke-width=".5" stroke-opacity=".502" d="M81.19 96.94c-.14.39-.28.79-.42 1.18-2.18-3.66-4.31-7.35-6.41-11.06-.54-1.25-1.06-2.49-1.55-3.73.16-.09.32-.19.48-.29 1.75 2.88 3.55 5.74 5.42 8.59.91 1.72 1.74 3.49 2.48 5.31Z" />
+  <path fill="currentColor" fill-opacity=".522" stroke="currentColor" stroke-width=".5" stroke-opacity=".522" d="m57.9 58.25 1.2-.54c.78.8 1.53 1.63 2.26 2.47 2.15 4.21 4.34 8.4 6.58 12.55q.285-.795.33-1.32c-.83-1.19-1.67-2.39-2.53-3.62-2.54-3.95-3.44-9.25-7.18-12.18Z" />
+</svg>
diff --git a/providers/neosantara/models/claude-4.5-opus.toml b/providers/neosantara/models/claude-4.5-opus.toml
new file mode 100644
index 00000000000..a4d4256c940
--- /dev/null
+++ b/providers/neosantara/models/claude-4.5-opus.toml
@@ -0,0 +1,17 @@
+base_model = "anthropic/claude-opus-4-5"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high"]
+
+[cost]
+input = 5
+output = 25
+cache_read = 0.5
+cache_write = 6.25
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/claude-4.5-sonnet.toml b/providers/neosantara/models/claude-4.5-sonnet.toml
new file mode 100644
index 00000000000..ff2be1b7756
--- /dev/null
+++ b/providers/neosantara/models/claude-4.5-sonnet.toml
@@ -0,0 +1,18 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "anthropic/claude-sonnet-4-5"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 3
+output = 15
+cache_read = 0.3
+cache_write = 3.75
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/claude-fable-5.toml b/providers/neosantara/models/claude-fable-5.toml
new file mode 100644
index 00000000000..975c0ff8372
--- /dev/null
+++ b/providers/neosantara/models/claude-fable-5.toml
@@ -0,0 +1,20 @@
+base_model = "anthropic/claude-fable-5"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 10
+output = 50
+cache_read = 1
+cache_write = 12.5
+
+[limit]
+context = 200_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/claude-opus-4-6.toml b/providers/neosantara/models/claude-opus-4-6.toml
new file mode 100644
index 00000000000..855d25016e7
--- /dev/null
+++ b/providers/neosantara/models/claude-opus-4-6.toml
@@ -0,0 +1,20 @@
+base_model = "anthropic/claude-opus-4-6"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "max"]
+
+[cost]
+input = 5
+output = 25
+cache_read = 0.5
+cache_write = 6.25
+
+[limit]
+context = 200_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/claude-opus-4-7.toml b/providers/neosantara/models/claude-opus-4-7.toml
new file mode 100644
index 00000000000..edf73248854
--- /dev/null
+++ b/providers/neosantara/models/claude-opus-4-7.toml
@@ -0,0 +1,20 @@
+base_model = "anthropic/claude-opus-4-7"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 5
+output = 25
+cache_read = 0.5
+cache_write = 6.25
+
+[limit]
+context = 200_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/claude-opus-4-8.toml b/providers/neosantara/models/claude-opus-4-8.toml
new file mode 100644
index 00000000000..752a1fee6bb
--- /dev/null
+++ b/providers/neosantara/models/claude-opus-4-8.toml
@@ -0,0 +1,20 @@
+base_model = "anthropic/claude-opus-4-8"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 5
+output = 25
+cache_read = 0.5
+cache_write = 6.25
+
+[limit]
+context = 200_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/claude-opus-5.toml b/providers/neosantara/models/claude-opus-5.toml
new file mode 100644
index 00000000000..f12a8d29ae3
--- /dev/null
+++ b/providers/neosantara/models/claude-opus-5.toml
@@ -0,0 +1,15 @@
+base_model = "anthropic/claude-opus-5"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 5
+output = 25
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/claude-sonnet-4-6.toml b/providers/neosantara/models/claude-sonnet-4-6.toml
new file mode 100644
index 00000000000..1744aacaf58
--- /dev/null
+++ b/providers/neosantara/models/claude-sonnet-4-6.toml
@@ -0,0 +1,20 @@
+base_model = "anthropic/claude-sonnet-4-6"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "max"]
+
+[cost]
+input = 3
+output = 15
+cache_read = 0.3
+cache_write = 3.75
+
+[limit]
+context = 200_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/claude-sonnet-5.toml b/providers/neosantara/models/claude-sonnet-5.toml
new file mode 100644
index 00000000000..b3e3e1cc876
--- /dev/null
+++ b/providers/neosantara/models/claude-sonnet-5.toml
@@ -0,0 +1,20 @@
+base_model = "anthropic/claude-sonnet-5"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 3
+output = 15
+cache_read = 0.3
+cache_write = 3.75
+
+[limit]
+context = 200_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/deepseek-v4-flash-0731.toml b/providers/neosantara/models/deepseek-v4-flash-0731.toml
new file mode 100644
index 00000000000..41f59d2d4a0
--- /dev/null
+++ b/providers/neosantara/models/deepseek-v4-flash-0731.toml
@@ -0,0 +1,13 @@
+base_model = "deepseek/deepseek-v4-flash-0731"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "high", "max"]
+
+[cost]
+input = 0.14
+output = 0.28
+cache_read = 0.03
diff --git a/providers/neosantara/models/deepseek-v4-flash.toml b/providers/neosantara/models/deepseek-v4-flash.toml
new file mode 100644
index 00000000000..52a52f369f2
--- /dev/null
+++ b/providers/neosantara/models/deepseek-v4-flash.toml
@@ -0,0 +1,14 @@
+base_model = "deepseek/deepseek-v4-flash-0731"
+name = "DeepSeek V4 Flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "high", "max"]
+
+[cost]
+input = 0.14
+output = 0.28
+cache_read = 0.0028
diff --git a/providers/neosantara/models/deepseek-v4-pro-0813.toml b/providers/neosantara/models/deepseek-v4-pro-0813.toml
new file mode 100644
index 00000000000..e88728c0359
--- /dev/null
+++ b/providers/neosantara/models/deepseek-v4-pro-0813.toml
@@ -0,0 +1,13 @@
+base_model = "deepseek/deepseek-v4-pro-0813"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "high", "max"]
+
+[cost]
+input = 1.32
+output = 3.96
+cache_read = 0.13
diff --git a/providers/neosantara/models/deepseek-v4-pro.toml b/providers/neosantara/models/deepseek-v4-pro.toml
new file mode 100644
index 00000000000..9ed5e2a9018
--- /dev/null
+++ b/providers/neosantara/models/deepseek-v4-pro.toml
@@ -0,0 +1,14 @@
+base_model = "deepseek/deepseek-v4-pro-0813"
+name = "DeepSeek V4 Pro"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "high", "max"]
+
+[cost]
+input = 2.4
+output = 4.8
+cache_read = 0.024
diff --git a/providers/neosantara/models/devstral-2.toml b/providers/neosantara/models/devstral-2.toml
new file mode 100644
index 00000000000..19e31206c56
--- /dev/null
+++ b/providers/neosantara/models/devstral-2.toml
@@ -0,0 +1,8 @@
+base_model = "mistral/devstral-2512"
+
+[cost]
+input = 0.4
+output = 2
+
+[limit]
+context = 256_000
diff --git a/providers/neosantara/models/gemini-3.1-flash-lite.toml b/providers/neosantara/models/gemini-3.1-flash-lite.toml
new file mode 100644
index 00000000000..0e87adb20c3
--- /dev/null
+++ b/providers/neosantara/models/gemini-3.1-flash-lite.toml
@@ -0,0 +1,19 @@
+base_model = "google/gemini-3.1-flash-lite"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["minimal", "low", "medium", "high"]
+
+[cost]
+input = 0.25
+output = 1.5
+cache_read = 0.025
+
+[limit]
+context = 1_000_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/gemini-3.5-flash.toml b/providers/neosantara/models/gemini-3.5-flash.toml
new file mode 100644
index 00000000000..4b6ba44c0a3
--- /dev/null
+++ b/providers/neosantara/models/gemini-3.5-flash.toml
@@ -0,0 +1,20 @@
+base_model = "google/gemini-3.5-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["minimal", "low", "medium", "high"]
+
+[cost]
+input = 1.5
+output = 9
+cache_read = 0.15
+cache_write = 1.5
+
+[limit]
+context = 1_000_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/gemini-3.6-flash.toml b/providers/neosantara/models/gemini-3.6-flash.toml
new file mode 100644
index 00000000000..4065d09c643
--- /dev/null
+++ b/providers/neosantara/models/gemini-3.6-flash.toml
@@ -0,0 +1,20 @@
+base_model = "google/gemini-3.6-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["minimal", "low", "medium", "high"]
+
+[cost]
+input = 0.75
+output = 3.75
+cache_read = 0.075
+cache_write = 0.75
+
+[limit]
+context = 1_000_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/gemini-3.7-flash.toml b/providers/neosantara/models/gemini-3.7-flash.toml
new file mode 100644
index 00000000000..46aa16386a5
--- /dev/null
+++ b/providers/neosantara/models/gemini-3.7-flash.toml
@@ -0,0 +1,20 @@
+base_model = "google/gemini-3.7-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high"]
+
+[cost]
+input = 0.75
+output = 3.75
+cache_read = 0.075
+cache_write = 0.75
+
+[limit]
+context = 1_000_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/glm-4.5-flash.toml b/providers/neosantara/models/glm-4.5-flash.toml
new file mode 100644
index 00000000000..d2f6da84cc5
--- /dev/null
+++ b/providers/neosantara/models/glm-4.5-flash.toml
@@ -0,0 +1,16 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "zhipuai/glm-4.5-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.015625
+output = 0.046875
+
+[limit]
+context = 128_000
diff --git a/providers/neosantara/models/glm-4.6v-flash.toml b/providers/neosantara/models/glm-4.6v-flash.toml
new file mode 100644
index 00000000000..45ff43eae5b
--- /dev/null
+++ b/providers/neosantara/models/glm-4.6v-flash.toml
@@ -0,0 +1,13 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "zhipuai/glm-4.6v-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.015625
+output = 0.046875
diff --git a/providers/neosantara/models/glm-4.7-flash.toml b/providers/neosantara/models/glm-4.7-flash.toml
new file mode 100644
index 00000000000..dc643bcc6f6
--- /dev/null
+++ b/providers/neosantara/models/glm-4.7-flash.toml
@@ -0,0 +1,16 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "zhipuai/glm-4.7-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0
+output = 0
+
+[limit]
+context = 1_000_000
diff --git a/providers/neosantara/models/glm-4.7.toml b/providers/neosantara/models/glm-4.7.toml
new file mode 100644
index 00000000000..716b59f69a2
--- /dev/null
+++ b/providers/neosantara/models/glm-4.7.toml
@@ -0,0 +1,16 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "zhipuai/glm-4.7"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.625
+output = 1.875
+
+[limit]
+context = 200_000
diff --git a/providers/neosantara/models/glm-5.3-flash.toml b/providers/neosantara/models/glm-5.3-flash.toml
new file mode 100644
index 00000000000..a09d695afea
--- /dev/null
+++ b/providers/neosantara/models/glm-5.3-flash.toml
@@ -0,0 +1,16 @@
+base_model = "zhipuai/glm-5.3-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[cost]
+input = 0.15
+output = 0.5
+cache_read = 0.03
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/gpt-5-nano.toml b/providers/neosantara/models/gpt-5-nano.toml
new file mode 100644
index 00000000000..1d8b65e0bc4
--- /dev/null
+++ b/providers/neosantara/models/gpt-5-nano.toml
@@ -0,0 +1,16 @@
+base_model = "openai/gpt-5-nano"
+base_model_omit = ["limit.input"]
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["minimal", "low", "medium", "high"]
+
+[cost]
+input = 0.050625
+output = 0.405313
+
+[limit]
+context = 128_000
diff --git a/providers/neosantara/models/gpt-5.4-mini.toml b/providers/neosantara/models/gpt-5.4-mini.toml
new file mode 100644
index 00000000000..43dfdad1d82
--- /dev/null
+++ b/providers/neosantara/models/gpt-5.4-mini.toml
@@ -0,0 +1,18 @@
+base_model = "openai/gpt-5.4-mini"
+base_model_omit = ["limit.input"]
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh"]
+
+[cost]
+input = 0.75
+output = 4.5
+cache_read = 0.075
+cache_write = 0.75
+
+[limit]
+context = 128_000
diff --git a/providers/neosantara/models/gpt-5.4-nano.toml b/providers/neosantara/models/gpt-5.4-nano.toml
new file mode 100644
index 00000000000..0c7f54843de
--- /dev/null
+++ b/providers/neosantara/models/gpt-5.4-nano.toml
@@ -0,0 +1,12 @@
+base_model = "openai/gpt-5.4-nano"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh"]
+
+[cost]
+input = 0.2025
+output = 1.265625
diff --git a/providers/neosantara/models/gpt-5.4.toml b/providers/neosantara/models/gpt-5.4.toml
new file mode 100644
index 00000000000..148f0ef35b4
--- /dev/null
+++ b/providers/neosantara/models/gpt-5.4.toml
@@ -0,0 +1,20 @@
+base_model = "openai/gpt-5.4"
+base_model_omit = ["limit.input"]
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh"]
+
+[cost]
+input = 2.5
+output = 15
+cache_read = 0.25
+
+[limit]
+context = 270_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/gpt-5.5.toml b/providers/neosantara/models/gpt-5.5.toml
new file mode 100644
index 00000000000..8c2ec38ca96
--- /dev/null
+++ b/providers/neosantara/models/gpt-5.5.toml
@@ -0,0 +1,20 @@
+base_model = "openai/gpt-5.5"
+base_model_omit = ["limit.input"]
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh"]
+
+[cost]
+input = 5
+output = 30
+cache_read = 0.5
+
+[limit]
+context = 270_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/gpt-5.6-luna.toml b/providers/neosantara/models/gpt-5.6-luna.toml
new file mode 100644
index 00000000000..a7094fb9a76
--- /dev/null
+++ b/providers/neosantara/models/gpt-5.6-luna.toml
@@ -0,0 +1,20 @@
+base_model = "openai/gpt-5.6-luna"
+base_model_omit = ["limit.input"]
+attachment = false
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 0.2
+output = 1.2
+
+[limit]
+context = 272_000
+
+[modalities]
+input = ["text"]
diff --git a/providers/neosantara/models/gpt-5.6-sol.toml b/providers/neosantara/models/gpt-5.6-sol.toml
new file mode 100644
index 00000000000..d7235e662e9
--- /dev/null
+++ b/providers/neosantara/models/gpt-5.6-sol.toml
@@ -0,0 +1,20 @@
+base_model = "openai/gpt-5.6-sol"
+base_model_omit = ["limit.input"]
+attachment = false
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 5
+output = 30
+
+[limit]
+context = 272_000
+
+[modalities]
+input = ["text"]
diff --git a/providers/neosantara/models/gpt-5.6-terra.toml b/providers/neosantara/models/gpt-5.6-terra.toml
new file mode 100644
index 00000000000..84820250e5c
--- /dev/null
+++ b/providers/neosantara/models/gpt-5.6-terra.toml
@@ -0,0 +1,19 @@
+base_model = "openai/gpt-5.6-terra"
+base_model_omit = ["limit.input"]
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 2
+output = 12
+
+[limit]
+context = 272_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/gpt-image-2.toml b/providers/neosantara/models/gpt-image-2.toml
new file mode 100644
index 00000000000..1ff78c6e696
--- /dev/null
+++ b/providers/neosantara/models/gpt-image-2.toml
@@ -0,0 +1 @@
+base_model = "openai/gpt-image-2"
diff --git a/providers/neosantara/models/gpt-oss-120b.toml b/providers/neosantara/models/gpt-oss-120b.toml
new file mode 100644
index 00000000000..b1cf0808e38
--- /dev/null
+++ b/providers/neosantara/models/gpt-oss-120b.toml
@@ -0,0 +1,13 @@
+base_model = "openai/gpt-oss-120b"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high"]
+
+[cost]
+input = 0.15
+output = 0.6
+cache_read = 0.075
diff --git a/providers/neosantara/models/gpt-oss-20b.toml b/providers/neosantara/models/gpt-oss-20b.toml
new file mode 100644
index 00000000000..ffa8cf409b0
--- /dev/null
+++ b/providers/neosantara/models/gpt-oss-20b.toml
@@ -0,0 +1,12 @@
+base_model = "openai/gpt-oss-20b"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high"]
+
+[cost]
+input = 0.025
+output = 0.1
diff --git a/providers/neosantara/models/kimi-k2-thinking.toml b/providers/neosantara/models/kimi-k2-thinking.toml
new file mode 100644
index 00000000000..2fbdb7e22cc
--- /dev/null
+++ b/providers/neosantara/models/kimi-k2-thinking.toml
@@ -0,0 +1,14 @@
+# Dedicated thinking variant: always-on reasoning with no caller control.
+# Matches lab providers/moonshotai/models/kimi-k2-thinking.toml and peers OpenRouter/Vercel.
+base_model = "moonshotai/kimi-k2-thinking"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.6
+output = 2.5
+
+[limit]
+context = 256_000
diff --git a/providers/neosantara/models/kimi-k2.5.toml b/providers/neosantara/models/kimi-k2.5.toml
new file mode 100644
index 00000000000..d451cf021ce
--- /dev/null
+++ b/providers/neosantara/models/kimi-k2.5.toml
@@ -0,0 +1,19 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "moonshotai/kimi-k2.5"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.6
+output = 3
+
+[limit]
+context = 256_000
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/kimi-k2.6.toml b/providers/neosantara/models/kimi-k2.6.toml
new file mode 100644
index 00000000000..f0758670743
--- /dev/null
+++ b/providers/neosantara/models/kimi-k2.6.toml
@@ -0,0 +1,16 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "moonshotai/kimi-k2.6"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.285
+output = 1.2
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/kimi-k3.toml b/providers/neosantara/models/kimi-k3.toml
new file mode 100644
index 00000000000..d6c9b2067ad
--- /dev/null
+++ b/providers/neosantara/models/kimi-k3.toml
@@ -0,0 +1,16 @@
+base_model = "moonshotai/kimi-k3"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "high", "max"]
+
+[cost]
+input = 3
+output = 15
+cache_read = 0.3
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/laguna-s-2.1.toml b/providers/neosantara/models/laguna-s-2.1.toml
new file mode 100644
index 00000000000..3d9fab5f556
--- /dev/null
+++ b/providers/neosantara/models/laguna-s-2.1.toml
@@ -0,0 +1,16 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "poolside/laguna-s-2.1"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0
+output = 0
+
+[limit]
+context = 262_144
diff --git a/providers/neosantara/models/laguna-xs-2.1.toml b/providers/neosantara/models/laguna-xs-2.1.toml
new file mode 100644
index 00000000000..62eb6836203
--- /dev/null
+++ b/providers/neosantara/models/laguna-xs-2.1.toml
@@ -0,0 +1,13 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "poolside/laguna-xs-2.1"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0
+output = 0
diff --git a/providers/neosantara/models/ling-3.0-flash-fin.toml b/providers/neosantara/models/ling-3.0-flash-fin.toml
new file mode 100644
index 00000000000..756cf791b35
--- /dev/null
+++ b/providers/neosantara/models/ling-3.0-flash-fin.toml
@@ -0,0 +1,16 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "inclusionai/ling-3.0-flash-fin"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0
+output = 0
+
+[limit]
+context = 256_000
diff --git a/providers/neosantara/models/llama-3.3-nemotron-super-49b-v1.5.toml b/providers/neosantara/models/llama-3.3-nemotron-super-49b-v1.5.toml
new file mode 100644
index 00000000000..55db32f9e29
--- /dev/null
+++ b/providers/neosantara/models/llama-3.3-nemotron-super-49b-v1.5.toml
@@ -0,0 +1,16 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "nvidia/llama-3.3-nemotron-super-49b-v1.5"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.285
+output = 1.999437
+
+[limit]
+context = 132_000
diff --git a/providers/neosantara/models/minimax-m2.7.toml b/providers/neosantara/models/minimax-m2.7.toml
new file mode 100644
index 00000000000..c92cde098a0
--- /dev/null
+++ b/providers/neosantara/models/minimax-m2.7.toml
@@ -0,0 +1,11 @@
+# Always-on thinking: upstream Dahl forwards no reasoning_effort parameter.
+# Matches lab providers/minimax/models/MiniMax-M2.7.toml and peers OpenRouter/FastRouter/Cortecs.
+base_model = "minimax/MiniMax-M2.7"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.09
+output = 0.36
diff --git a/providers/neosantara/models/minimax-m3.toml b/providers/neosantara/models/minimax-m3.toml
new file mode 100644
index 00000000000..e82398345b9
--- /dev/null
+++ b/providers/neosantara/models/minimax-m3.toml
@@ -0,0 +1,20 @@
+# Toggle: reasoning_effort = "none" turns thinking off; any other accepted
+# value turns it on. https://docs.neosantara.xyz/en/capability/reasoning
+base_model = "minimax/MiniMax-M3"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.3
+output = 1.2
+cache_read = 0.06
+
+[limit]
+context = 524_288
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/mistral-large-latest.toml b/providers/neosantara/models/mistral-large-latest.toml
new file mode 100644
index 00000000000..dc4f069d5f2
--- /dev/null
+++ b/providers/neosantara/models/mistral-large-latest.toml
@@ -0,0 +1,12 @@
+base_model = "mistral/mistral-large-latest"
+attachment = false
+
+[cost]
+input = 0.5
+output = 1.5
+
+[limit]
+context = 128_000
+
+[modalities]
+input = ["text"]
diff --git a/providers/neosantara/models/mistral-small-latest.toml b/providers/neosantara/models/mistral-small-latest.toml
new file mode 100644
index 00000000000..c2f93013a33
--- /dev/null
+++ b/providers/neosantara/models/mistral-small-latest.toml
@@ -0,0 +1,13 @@
+base_model = "mistral/mistral-small-latest"
+attachment = false
+reasoning = false
+
+[cost]
+input = 0.15
+output = 0.6
+
+[limit]
+context = 128_000
+
+[modalities]
+input = ["text"]
diff --git a/providers/neosantara/models/muse-glimmer-30b.toml b/providers/neosantara/models/muse-glimmer-30b.toml
new file mode 100644
index 00000000000..63fdbc24b80
--- /dev/null
+++ b/providers/neosantara/models/muse-glimmer-30b.toml
@@ -0,0 +1,17 @@
+# Sources:
+# https://huggingface.co/meta-models/Muse-Glimmer-30B
+# Effort: reasoning_effort = low|medium|high|xhigh
+# Matches peers OpenRouter and Vercel AI Gateway.
+base_model = "meta/muse-glimmer-30b"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh"]
+
+[cost]
+input = 0.35
+output = 1.5
+cache_read = 0.04
diff --git a/providers/neosantara/models/muse-spark-1.1.toml b/providers/neosantara/models/muse-spark-1.1.toml
new file mode 100644
index 00000000000..a62e2e536c1
--- /dev/null
+++ b/providers/neosantara/models/muse-spark-1.1.toml
@@ -0,0 +1,16 @@
+base_model = "meta/muse-spark-1.1"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["minimal", "low", "medium", "high", "xhigh"]
+
+[cost]
+input = 1.25
+output = 4.25
+cache_read = 0.15
+
+[modalities]
+input = ["text", "image"]
diff --git a/providers/neosantara/models/step-3.5-flash.toml b/providers/neosantara/models/step-3.5-flash.toml
new file mode 100644
index 00000000000..b092cac10ae
--- /dev/null
+++ b/providers/neosantara/models/step-3.5-flash.toml
@@ -0,0 +1,12 @@
+base_model = "stepfun/step-3.5-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high"]
+
+[cost]
+input = 0.1875
+output = 0.5625
diff --git a/providers/neosantara/provider.toml b/providers/neosantara/provider.toml
new file mode 100644
index 00000000000..bcd8aa05c31
--- /dev/null
+++ b/providers/neosantara/provider.toml
@@ -0,0 +1,27 @@
+# Source (accessed 2026-09-03):
+# - https://docs.neosantara.xyz
+# - https://api.neosantara.xyz/v1/catalog  (models.dev / LLM Gateway shape, public, no key)
+#
+# The sync consumes the single /v1/catalog endpoint. The gateway does the host-specific work
+# upstream (USD-per-token pricing incl. IDR conversion and any discount, capability flags, and
+# the reasoning surface), so the sync is a thin, override-only translation.
+#
+# Reasoning: the catalog reports each model's real per-model effort surface in providers[].
+# reasoning_efforts, mapped to reasoning_options as:
+#   []          -> [] (always-on: reasons, no caller control)
+#   ["none"]    -> toggle (reasoning_effort = "none" is off; carries a leading wire comment)
+#   graded set  -> effort, using the model's real levels (lab + same-surface peers), never a
+#                  full enum. Host ceiling is none|minimal|low|medium|high|xhigh|max.
+# When off is expressed via reasoning_effort = "none" alongside graded levels, AGENTS.md strictly
+# requires authoring only `effort` with `none` in values and NO toggle. Only binary on/off models
+# (no graded levels, only ["none"]) author toggle alone with a wire comment. Dedicated headers
+# document always-on relays (minimax-m2.7, kimi-k2-thinking) and external citations.
+#
+# Streaming: Neosantara normalizes reasoning output across all upstream providers into the
+# OpenAI-compatible delta.reasoning_content field (see https://docs.neosantara.xyz/en/capability/reasoning).
+# All reasoning models therefore advertise interleaved.field = "reasoning_content".
+name = "Neosantara"
+env = ["NEOSANTARA_API_KEY"]
+npm = "@ai-sdk/openai-compatible"
+api = "https://api.neosantara.xyz/v1"
+doc = "https://docs.neosantara.xyz"