v0.5.21

Changed

• cloudflare: 5 new models added — @cf/meta/llama-3.2-90b-instruct (A+,128k), @cf/mistralai/mistral-7b-instruct-v0.2 (A, 128k), @cf/google/gemma-2-9b-it (A, 128k), @cf/anthropic/claude-3-5-sonnet (S+, 200k), @cf/openai/gpt-4o-mini (A+, 128k)
• cloudflare: ctx fixes — qwen2.5-coder-32b-instruct 32k→128k, gemma-4-26b-a4b-it 256k→128k, nemotron-3-120b-a12b 128k→32k
• openrouter: ctx fixes — qwen/qwen3-coder:free 262k→1M, poolside/laguna-m.1:free 131k→262k

Fixed

• cloudflare: removed @cf/openai/gpt-oss-120b (OpenAI deprecated) and @cf/deepseek-ai/deepseek-r1-distill-qwen-32b (upstream removal)
• googleai + gemini: individual deprecation comments added to all 3 Gemini 2.5 models (gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite) marking shutdown Oct 16, 2026

Notes

• All 14 audited providers verified manually against live provider APIs and docs
• Clean providers (no changes): cerebras, github-models, mistral, sambanova, ovhcloud, codestral, zai, scaleway, qwen, opencode-zen
• Total model count updated: 154→157