StepFun previews a 600B open model at $1 per million tokens, with weights due October 15
StepFun launched Step 5 Preview on September 20, a 600 billion parameter MoE priced at $1 per million input tokens with a 95 percent cache discount. Open weights are scheduled for October 15.
The decision it puts on your desk
Test Step 5 Preview on one cache-heavy workflow by October 15, when open weights land. Run an A-B against your current API before you commit, and if you build agent products that live on context reads, price workloads against the $1 per million tokens and the 95 percent cache discount, since your cost floor just dropped below the DeepSeek benchmark.
The mid-frontier got its own floor on Sunday, and it sits in China.
StepFun launched the Step 5 Preview API on Sunday.
The launch specs, as trade reporting recorded them: "a 600 billion parameter sparse mixture-of-experts model with 27 billion active parameters per token and a 1 million token context, priced at $1 per million input tokens and $2.70 per million output with a 95 percent cache discount."
Artificial Analysis scores Step 5 Preview at 44 on its Intelligence Index, a mid-frontier result that measures against GPT-5.6 Terra and sits below Opus 5.
Full open weights arrive October 15, StepFun said.
The cache discount is the products math
Follow the money on one line of that release. Dropping repeated-context input to $0.05 per million tokens undercuts what DeepSeek charges off-peak, and context-heavy agent work runs hundreds of repeated reads inside a session.
The economics favor one builder class: teams running long agent sessions with big shared contexts, where repeated reads dominate the bill.
The parameter design cuts inference cost again. With 27 billion active parameters on a 600 billion base, the compute per token is intentionally lean.
Under the hood: the release follows an open cadence the market already noticed.
Shanghai AI Lab's free 744B Atria Dawn and DeepSeek's V4.1 Flash landed inside the same window, and the Chinese open-weight release pace is roughly one flagship a week.
The dated commit matters more than the score. Weights on October 15 turn the preview into a countdown, and three anchored open-weights releases land inside one quarter.
Ecosystem check: the pricing cuts across two channels at once.
US closed APIs price capability at a premium, and self-hosted open weights price at zero. Step 5 Preview rents frontier-adjacent results at a rate the market rarely sees.
No rival has repriced yet. Anthropic and OpenAI's public rates stayed put this weekend, so I get the sense the response will arrive in benchmarks months rather than press releases.
The unresolved edge is independent testing. Artificial Analysis put it at 44, and the launch scores stay vendor-conditioned until the downloaded weights get rerun by outsiders after October 15.
Source
← Previous
Trump calls AI safety a hoax and promises an AI Force. The split with Sacramento is now public
Next →
Google disclosed that Gemini reached live systems at three real companies during a cyber test. The timing is part of the story too
The Morning Recap
We sift the day. You make the calls.
The decision on every story, in your inbox before your first call.