v0.5.97

Added

• 📡 Full 25-provider audit, 2 researchers at a time: 265 → 271 live models. Every provider verified against its live API or official docs on 2026-09-29.
• onomeo free pool exploded to 33 models: Gemini 3.8/3.7/3.5 Flash, GLM 4.5/4.6V/4.7 Flash and Ministral 3 14B all joined the free pool this week (maintainer-confirmed catalog).
• Novita free tier is back: Qwen3.6 Plus and Qwen3.5 Plus (1M ctx, S+), Ling 3.1 Flash and dev/glm46 are zero-price again after the August dry spell.
• NVIDIA NIM: DeepSeek V4.1 Flash (1M ctx, 552B MoE agentic coder), Kimi K2.6 (256k), Nemotron Nano 3 30B and Codestral 22B re-verified live.
• LLM7: DeepSeek V4 Flash 0731 (400k ctx, reasoning + tools) verified answering with no key at all.
• OpenRouter: Space Bunny Alpha (anonymous stealth, 1M ctx) and the openrouter/free meta-router.
• OVHcloud: Qwen3-Coder-30B-A3B and Mistral Small 3.2 live on the router even though the marketing page hides them.
• ⚙️ Env variable placeholders in the config file (issue #196): any apiKeys value can now be "${MY_VAR}" and is resolved from the environment at read time. The file on disk keeps the placeholder so secrets never land on disk - perfect for Docker env_file setups. The shell env-file generator skips placeholder entries instead of exporting a literal that would clobber your real variable.

Changed

• 🪦 6 DashScope models flagged deprecated, shutdown 2026-10-10 per official Alibaba notices: qwen3-coder-plus, qwen3-coder-next, qwen3-coder-480b, qwen3-coder-30b, qwen3.6-max-preview and qwen3-vl-flash. They still serve until then; qwen3.7-plus/3.7-max are the replacements.
• OrcaRouter quota strings now reflect reality: 5 $0 models + the router, 10 req/min, 50 req/day.
• Docs provider table regenerated against the shipped catalog.

Fixed

• 🐛 CORS in Docker finally tamed (issue #198): FCM_ALLOWED_ORIGINS="*" now works. The list used to be exact-match only, so the * entry was silently ignored and cross-machine browser tabs failed CORS checks. Host-header DNS-rebinding protection stays on.
• Removed 13 dead or quietly-paid models: mistral-nemotron (EOL 09-28), kimi-k2-thinking (EOL 05-12), Mistral Large 3 (403 paid-tier-only on a free key), Z.ai GLM-5.3 + FlashX (paid-only), Zen DeepSeek-V4-Flash-Free (upstream dead) and Muse Spark 1.2 (de-promoted), the ling-3.0-flash-fin :free variants on OpenRouter/Kilo, 2 paid Ling models on Novita, and the OrcaRouter fusion trio, which bills the sum of its paid frontier panel legs despite reporting $0.
• Context windows corrected against live data: GLM-5.3 on NVIDIA 200k → 1M, muse-glimmer and gpt-oss-20b 128k → 131k, Zen space-bunny and longcat 200k → 1M, Pollinations nova-2-lite → nova-2-lite-v1 rename, Nex N2.5 pair ctx filled at 256k.
• Kimi K3 on Ollama Cloud now shows its 76.8% SWE-bench Verified score.
• onomeo provider text synced with the site's current model (PR #201, thanks @Owaid0723666).