Confirmed independently by both passes — listed free on OpenRouter (1M context, 524K output, multimodal) and live inside Command Code's harness. Worth testing on non-sensitive work only: provider identity is unconfirmed and OpenRouter's listing says prompts/outputs may be retained (not for training, but still retained).
OpenAI says training agents posted 53 user images (from accounts that had allowed training use) to external image hosts; full review will take months. If you've opted into training use anywhere with OpenAI, this is worth checking.
OpenAI Devs confirmed a Codex outage caused spurious 401s on Sep 25. If a job died with a 401 that day, re-queue it — rotating your key does nothing and just adds churn.
An Anthropic employee clarified low/medium effort is fine for normal feature work, high for when you need real test coverage, max for long jobs or security review. Cranking effort adds checks — it does not fix a wrong plan.
Plugins now package MCP servers + skills together. The 110x MCP-usage jump is Anthropic's own number — worth a look if you're not yet using Claude plugins/MCP for repeat workflows.
LFM2.5-VL-3B-DSpark, a speculative-decoding draft model, claims 3.13x faster decoding on Apple silicon / 2.66x on H100 (Safetensors + GGUF, works in SGLang/MLX-VLM/llama.cpp). The LFM Open License only gives free commercial use under $10M annual revenue — review before embedding it in a paid product.
"Code" builds apps/dashboards from natural language; "Autopilot" is a longer-running named agent with a role and objective. Early access only, end of September. Don't budget around it yet unless you're already in the preview.
One IDE workspace supporting Claude Agent, Codex, Gemini CLI/3 Flash, and Junie side by side. Worth inspecting if you want to avoid rebuilding your workflow every time you switch agent vendors. No pricing/availability confirmed yet.