Step 5 Preview: a 1M-context, $1/M agentic coder, live on OpenRouter
A 600B MoE doing agentic coding at a fraction of frontier price. Plus: OpenAI watermarks EU ChatGPT/Codex text, and its fired safety trio escalates.
Copy markdown
A 1M-context agentic coder for $1 per million
StepFun's Step 5 Preview is live on OpenRouter: a 600B-param sparse MoE (27B active) with a 1M-token window and 64k max output, at $1/M input, $2.70/M output, and just $0.05/M cache reads—frontier-class agentic coding at a fraction of GPT-6 or Claude pricing. It's tuned for long, multi-step work over large codebases, supports tool calling and structured outputs, and is already topping usage in Cline and Kilo Code.
OpenAI starts watermarking EU ChatGPT and Codex text
Over the coming weeks, OpenAI will statistically watermark eligible ChatGPT and Codex text for EU users via 'textGrain,' which nudges word choices into a detectable pattern for EU AI Act compliance. It's not global and the API stays opt-in, off by default—but EU-facing output you ship is now fingerprinted. The out: rewriting erases it—swapping 10% of words cut detection from 92% to 66%, and 25% dropped it to 17%.
OpenAI's fired safety trio takes it to the board
The three researchers OpenAI fired for allegedly mishandling info—one was its technical contact for external evaluators METR and Redwood Research—are now urging the board to halt 'opaque-reasoning' models and mandate independent audits, and an internal OpenAI memo reportedly backs their Oct 7 letter. Why it matters: the people gating frontier releases through outside evals are walking out, days after OpenAI pulled GPT-6.1 Astra for failing safety checks.