StepFun's Step 5 Preview undercuts Opus 5 on price

StepFun released Step 5 Preview, a 600 billion parameter sparse mixture-of-experts model with 27 billion active parameters and a 1 million token context window. It's live through StepFun's products and API, with full weights due October 15.

Step 5 scores 44 on the Artificial Analysis Intelligence Index, at about one-eighth Claude Opus 5's per-task cost, according to StepFun's release page. For teams running high volume inference, that ratio changes which model gets the default slot.

StepFun published its own gaps alongside the price. Step 5 scores 67.7% on DeepSWE v1.1 against GPT-6 Astra's 74.1%. On StepCodeBench it scores 49.0% against Opus 5's 63.9%. Those are real deficits, but ones a lot of workloads can absorb.

Terminal-Bench v4 shows the widest gap: 33.3% for Step 5 against Astra's 57.9%. A 24-point deficit on terminal work, running commands, reading output, recovering from a failed one, sits right where a coding agent burns through budget on retries. An eighth of the price stops looking cheap once the agent needs three or four extra turns to finish a task Astra completes in one.

Step 5 fits work that doesn't loop through a shell: chat, summarization, single-shot generation where the 1 million token context matters more than agentic reliability. For terminal-heavy coding, the benchmark gap is StepFun's own number to explain.