feat: add Pareto Inference (Pass)

[dylanduyvu]

Summary

Add Pareto Inference (Pass) with three models:

• z-ai/glm-5.3
• z-ai/glm-5.3-flash
• deepseek/deepseek-v4-flash

I work on Pareto Inference. The provider uses the OpenAI-compatible Chat Completions endpoint at https://api.paretoinference.com/v1, with PARETO_API_KEY for authentication. The model files inherit existing base models. This change adds only one provider file, one logo, and three model files.

Subscription pricing

This is paid Pass access, not free access. Pareto Pass costs $3/week and includes up to $20 of model usage per day. There are no additional per-token charges within that allowance. The zero cost fields follow the existing zai-coding-plan convention. The provider name, documentation link, and source comments identify the paid plan.

This entry does not represent legacy keys or a separate pay-as-you-go product. If a different catalog convention is preferred for subscription providers, I can adjust it.

Sources

• Pass price, daily allowance, and key setup
• Public model IDs
• API reference
• OpenCode setup
• OpenRouter catalog, checked September 2, 2026, for provider context limits and Flash input modalities. Pareto routes these IDs through OpenRouter.

Pareto accepts reasoning_effort. This entry does not advertise OpenRouter's separate reasoning.enabled toggle. GLM exposes low/high/max; DeepSeek exposes high/xhigh. Pareto maps max to OpenRouter's xhigh.

Verification

• bun validate passed.
• Web catalog build passed.
• SDK generation, type checking, and all 23 SDK tests passed.
• Two local catalog/logo checks passed with 71 assertions. The generated-provider check failed as expected before this addition.
• Real OpenCode 1.18.27 loaded the generated catalog through OPENCODE_MODELS_PATH, with no manual provider URL or adapter override. Each of the three models completed one permitted local tool call and its response continuation, using its high-effort variant.
• All eight advertised reasoning-effort values returned HTTP 200 on short synthetic requests.
• Test keys had a $0.50 budget, a 15-minute expiry, and one concurrent request. Both keys used during verification were revoked and subsequent requests were rejected.

The OpenCode tests used a 4,096-output-token cap. The effort probes used a 128-output-token cap. Maximum context/output lengths and image/video inputs were not live-tested; those metadata values follow the cited upstream catalog and existing base models.

[github-actions]

No actionable findings.