Step 5 Preview
by StepFun
Today's price · per 1M tokens
Input
$1.00
per 1M tokens
Output
$2.70
per 1M tokens
Blended
$1.43
blended $/1M — 3:1 weighted input:output
Cached input
$0.050
5% of input — prompt caching
Price receipt
We read StepFun's own pricing page on and found step-5-preview listed with a price on the same row — that read is where the number above comes from, recorded as a new model. We keep the page text we read; its content hash is ca8a14ccf4.
Overview
StepFun's new flagship, released 2026-10-08 as a preview on its Open Platform and built for agentic work, coding and long-document tasks. It takes text, image and video input with a 1M-token context and up to 64K output tokens, and output tokens include its reasoning. StepFun has not released weights, so it is API-only. $1.00/$2.70 per 1M tokens.
Capabilities
struck through = not supportedBenchmark performance
accuracy % · higher is betterWe hold no per-benchmark accuracy scores for this model yet, so it has no accuracy average. That is a gap in our coverage, not a sign the model is untested — independent evaluators often publish a composite index for a new model long before releasing its per-benchmark numbers. The percentile above is its standing across the independent composites below.
- AA Intelligence Index: 28.5 (Artificial Analysis composite across reasoning, knowledge and coding evals)
- AA-Omniscience Index: 16.4 (−100–100)
- GDPval-AA v2: 1590.7 ELO (Real-world work tasks, human baseline = 1000)
- Arena (arena.ai): 1448.8 ELO (Human preference)
- Arena Vision: 1246.5 ELO (Human preference on image-input prompts)
- Arena Agent: -0.051 (Score from real agent sessions (IPS); higher is better)
- Arena Agent cost per task: $0.58 (Median measured cost of one real Agent Arena task)
- HLE-Diamond: 24.3% (measured 2026-11-22) — A revised 1,000-question set of Humanity's Last Exam, half reasoning, half knowledge
- CritPt v2, hard subset: 32.5% (measured 2026-11-22) — The hardest research-level physics problems in CritPt v2, solved with a code sandbox
- FrontierCode-AA 2.0: 12% (measured 2026-11-22) — Coding tasks in real open-source repositories, graded on whether a maintainer would merge the change
- Terminal-Bench-Science: 2.4% (measured 2026-11-22) — Expert-written science research workflows completed in a terminal
- GDP.pdf: 14.8% (measured 2026-11-22) — Tasks over long professional documents; share of tasks that pass every criterion
- AutomationBench-AA: 51% (measured 2026-11-22) — Multi-step work across simulated business apps, without breaking the rules set
- MLCR-AA: 18.3% (measured 2026-11-22) — Reasoning over long, fragmented medical records; pass rate
- AA-LCR (long-context reasoning): 88.3% (measured 2026-11-22)
- Terminal-Bench 4.0: 33.3% (measured 2026-11-22)
- SciCode: 58.9% (measured 2026-11-22)
- MMMU-Pro: 76.4% (measured 2026-11-22)
- Cost to run the AA Intelligence Index: $2.35 per task, 17 min per task
- Measured by AA: 111 tokens/sec output, 14.81s to first token, 40.22s to first answer token
- Ranks #58 of 207 comparably-measured models by percentile score, across 4 independent measurements
- Ranks #60 of 207 comparably-tested models by normalized performance per dollar
111 tokens/sec output 14.81s latency to first token (TTFT)
Throughput: Artificial Analysis, measured 2026-11-22 · Source: Artificial Analysis (artificialanalysis.ai)
Source: · updated · See full rankings →
Specifications
- Provider
- StepFun
- Context window
- 1M tokens
- Modality
- text, image, video
- Parameters
- Proprietary
- Open source
- No — proprietary
- Released
- Status
- Preview
- Last updated
- Tags
Availability verified: — listed on StepFun's own page