[opencode-agent]

Changes

• Use Bailing's API base URL (/api/llm/v1) so the OpenAI-compatible adapter appends /chat/completions once.
• Remove Mistral's unsupported bare magistral-small ID; the model lifecycle list identifies dated Magistral Small IDs (magistral-small-2506, -2507, -2509) and their retirement, not the bare ID. Lab metadata for dated models is untouched.
• Add published cached-input prices (USD/M tokens) for currently priced Mistral Large 3, Medium 3.5, Small 4 and Codestral aliases/dated IDs. Mistral pricing; no separately priced cache-write field is documented.
• Remove Together's Kimi K2.5, K2.6 and K2.7 Code serverless routes; Together's deprecation history records their removal. Dedicated-endpoint availability is a different surface. Keep K3.
• Remove NVIDIA's unavailable hosted openai/gpt-oss-120b and moonshotai/kimi-k2-instruct-0905; both are absent from its live model list. The 120B entry already recorded a 410 Gone/end-of-life probe. Keep NVIDIA's live 20B and newer Kimi models.

Checked but unchanged

• Cloudflare Workers AI's GLM served-context limits were corrected and sync-pinned already in fix(cloudflare): correct overstated context windows #8314.
• OpenRouter's Qwen3 Coder Plus cache prices already match the model API. That API doesn't expose a cache-read rate for Qwen3.6 Plus, so one cannot be supplied safely from a different host's rates.
• Perplexity Agent Claude pricing publishes the existing cache-read rates but no cache-write price; leave it unset rather than inventing a charge.

Validation: bun validate and git diff --check pass.

Requested by: @rekram1-node (Aiden via Slack)

[github-actions]

No actionable findings.