AI tools · September 21, 2026

StepFun’s Step 5 Preview: a 600B agent model that undercuts rivals on price

A new frontier AI model launched on September 20 — and it’s built for agents, not chat. StepFun, the Shanghai lab founded in 2023 by former Microsoft veteran Jiang Daxin, released Step 5 Preview, a 600-billion-parameter sparse mixture-of-experts model with a 1-million-token context window. It accepts text, images (up to 60 per request), and video as input, and it ships the full agent-workload toolkit: tool calling, streaming, JSON mode and JSON Schema, prompt caching, and low, medium, and high reasoning-effort levels.

The pitch is price. At $1.00 per million input tokens and $2.70 per million output tokens, Step 5 Preview costs roughly one-seventh of the $10.00 median output price Artificial Analysis reports for comparable models ($1.88 input median). StepFun says the model coordinated 950 web fetches in a single agent action, tuned an H100 kernel to 508 TFLOPS versus Claude Opus 5’s 493 in a 24-hour experiment, and lifted Qwen3-30B-A3B’s AIME24 score from 53.3% to 60% — company-reported figures, not independently reproduced. Artificial Analysis measured 99.8 tokens per second on StepFun’s API and scores the model 44 on its Intelligence Index, against a median of 24 for the price tier.

The API is live today, and the date that matters is October 15, 2026 — when StepFun has promised full open weights. A 600-billion-parameter release (about 1.2TB in BF16 before KV cache) would be one of the largest open-weights drops ever, and it would give self-hosters a near-frontier agentic model to run for the cost of electricity. Price-per-task is becoming the new leaderboard, and this model just reset it.

Why it matters

If you pay per agentic task — long coding runs, research sweeps, document pipelines — a near-frontier model at roughly one-seventh the output price changes the unit economics of everything you build. And if the October 15 open-weights drop lands as promised, self-hosters get their biggest agentic model yet, with no U.S. API bill attached.

Key facts

Launched September 20, 2026 by Shanghai-based StepFun (founded 2023 by ex-Microsoft veteran Jiang Daxin)
600B-parameter sparse MoE with ~27B active per token (~4.5%); 1M-token context; text, image (up to 60 per request), and video inputs; up to 64k text output tokens
API pricing: $1.00 per million input tokens, $2.70 per million output tokens — about one-seventh of the $10.00 median output price for comparable models (Artificial Analysis)
Company-reported agent feats: 950 web fetches in a single action; H100 kernel at 508 TFLOPS vs Claude Opus 5’s 493; AIME24 score of Qwen3-30B-A3B lifted from 53.3% to 60% (not independently verified)
Artificial Analysis: Intelligence Index score 44 (price-tier median 24); measured 99.8 tokens/sec output on StepFun’s API
Open weights promised October 15, 2026 — roughly 1.2TB in BF16, one of the largest open-weights releases on record

Sources

Spot an error? We correct quickly and note it. Contact us via the contact page.

← More AI news