StepFun
Step 5 Preview
StepFun pitches Step 5 Preview as frontier-level software engineering at $1.00/$2.70 per million tokens, but only three of its eight tracked scores are independent: Humanity's Last Exam (46.48) and AA-AnalystAgent (35.0), both Artificial Analysis, plus vals.ai's Terminal-Bench 4.0 (31.82). The coding and agent numbers behind the pitch remain StepFun's own.
Step 5 Preview benchmarks and pricing, every number sourced: 8 tracked Step 5 Preview benchmark scores (3 independently run, 5 still resting on a vendor’s own claim), priced at $1.00 per million input tokens and $2.70 per million output.
Step 5 Preview architecture: Mixture-of-Experts; 600B total parameters (27B activated per token); 1M-token context window.
- Released
- 2026-09-20
- License
- open-weights
- Context window
- 1M tokens
- Knowledge cutoff
- Not disclosed
- Verified
- sources checked 2026-09-22–2026-10-08
- Parameters
- 600B (27B active)
- Architecture
- Mixture-of-Experts
Step 5 Preview’s verified record
Step 5 Preview’s most-compared rival is GLM-5.3-Flash: 1 lead, 1 tie, and 6 not callable across their 8 shared comparisons. Step 5 Preview is priced at $1.00/$2.70 per 1M tokens (in/out) vs GLM-5.3-Flash’s $0.15/$0.50.
Against the 215 head-to-head comparisons Step 5 Preview shares with other tracked models:
A gap counts for Step 5 Preview only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Step 5 Preview trails on 28 of them.
HLE · no tools Reasoning
±2 is noiseAnalystAgent Spreadsheet & document analysis
±11.2 is noiseTerminal-Bench 4.0 Terminal ops
±12.4 is noiseNo verdict for Step 5 Preview anywhere on DeepSWE, HLE · with tools, Terminal-Bench 2.1, Toolathlon-Verified ( nothing independently confirmed on both sides); GPQA Diamond ( saturated).
- None of Step 5 Preview’s coding comparisons are independently confirmed on both sides yet.
Step 5 Preview API pricing
$1.00 in / $2.70 out per 1M tokens — official pricing source
What Step 5 Preview costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.127 |
| A codebase review | 1,000K / 100K | $1.27 |
| A day of agent work | 10,000K / 1,000K | $12.70 |
Computed from Step 5 Preview’s list rates above — cache discounts and batch tiers are not applied.
Step 5 Preview benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| HLE(with tools)[2] Reasoning · ±2 is noise | |
| GPQA Diamondsaturated[3] Expert science Q&A — not ranked at any gap size | |
| Terminal-Bench 2.1[4] Terminal ops · ±10.6 is noise | |
| DeepSWE[5] Long-horizon coding · ±9.5 is noise | |
| Toolathlon-Verified[6] Multi-tool chores · ±9.7 is noise | |
| AnalystAgent[7] Spreadsheet & document analysis · ±11.2 is noise | |
| Terminal-Bench 4.0[8] Terminal ops · ±12.4 is noise |
Who ran these numbers: 3 of 8 independent — artificialanalysis.ai (2), vals.ai (1); vendor self-reported (5).
- HLE: Artificial Analysis's own run (text-only 2,158-question subset, pass@1; the board rounds this to 46.5). AA lists the model as "Step 5 Preview" with no effort level, so which reasoning tier produced the score is not stated. StepFun's launch table reports the same 46.5, at its High setting.
- HLE: StepFun's own launch-table figure (High effort). The vendor's footnote says Step 5 Preview and GLM-5.3 were evaluated on the text-only subset while every other model in its table used the full dataset, and that the settings are not directly comparable. This site has found no independent with-tools run.
- GPQA Diamond: StepFun's own launch-table figure (High effort; harness not stated). GPQA Diamond is graded saturated on this site, so the number carries little weight either way; vals.ai no longer runs new models on its GPQA page, and Artificial Analysis had no Step 5 Preview entry there (checked 2026-09-21).
- Terminal-Bench 2.1: StepFun's own launch-table figure, labeled Terminal-Bench v2.1 (High effort); the agent harness is not stated. tbench.ai's 2.1 board (last updated 2026-09-03) does not list the model. Artificial Analysis now scores Terminal-Bench 4.0 instead, where it reports 33.3 (the same figure StepFun shows on its own Terminal-Bench v4 row) — a different benchmark this site does not track.
- DeepSWE: StepFun's own launch-table figure (High effort), run on the SWE-agent harness at temperature 1.0 and top_p 0.95. The independent deepswe.datacurve.ai board runs every model on mini-swe-agent, so this is not a like-for-like number, and that board (v1.1, updated 2026-09-03) does not list the model.
- Toolathlon-Verified: StepFun's own launch-table figure (High effort; harness not stated). toolathlon.xyz's board (latest entry 2026-08-30) does not list the model.
- AnalystAgent: AA's own run, board row 'Step 5 Preview': analystAgent 0.35 = 35.0 pass@5. Second independent score for this model alongside AA's HLE 46.48.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 31.818 at no effort tier stated. Artificial Analysis same-conditions run (no tier suffix): 33.33.
Notes on the record
StepFun announced Step 5 Preview on 2026-09-20 as its flagship for agentic work. Pricing on StepFun's international USD page (checked 2026-09-21): $1.00 per million input tokens on a cache miss, $0.05 on a cache hit and $2.70 for output, reasoning tokens included. It is one flat rate, with no length, time-of-day or promotional condition anywhere in StepFun's docs. Context window is 1M tokens; maximum output is 64K tokens — the docs listed 1M at launch and were edited within two days.
Per StepFun, it is a sparse mixture-of-experts with 600B total and 27B active parameters, taking text, images and video and returning text. Its docs offer three reasoning-effort levels; every StepFun figure here is at High. StepFun published the weights on 2026-10-15 (Hugging Face stepfun-ai/Step-5-Preview-BF16, plus FP8 and NVFP4) under the StepFun Model License 1.0, so this site lists it as open-weights. Terms are in the FAQ. Unofficial copies had circulated since 2026-09-20; those were not this release.
Three of the eight benchmark scores tracked here are independent: Humanity's Last Exam without tools (46.48, Artificial Analysis, no effort tier stated), AA-AnalystAgent (35.0 pass^5, Artificial Analysis, added 2026-10-03) and vals.ai's Terminal-Bench 4.0 (31.82, added 2026-10-08). The other five come from StepFun's launch table. Its with-tools HLE figure uses the text-only subset while most rivals' figures use the full dataset. StepFun's Agents' Last Exam figure is the harder ALE-CLI split, not the split this site tracks, so it is not recorded.
Two gateways listed it on 2026-10-08: OpenRouter at StepFun's own rates ($1/$2.70, $0.05 cache read, output capped at 64,000) and OpenCode Zen as step-5-preview-free at $0 with zero retention, output capped at 65,536. OpenCode calls that route limited time and publishes no end date (checked 2026-10-09), so it is a gateway subsidy rather than a price cut.
Compare with
FAQ
Has Step 5 Preview been independently benchmarked?
Partly. Independent boards have run it on three of this site's fourteen tracked benchmarks: Humanity's Last Exam (46.48, Artificial Analysis, text-only, no effort level shown), AA-AnalystAgent (35.0 pass^5, Artificial Analysis, added 2026-10-03), and — added 2026-10-08 — Terminal-Bench 4.0, where vals.ai's run reads 31.82 (default effort) against Artificial Analysis's 33.3. StepFun's own launch table supplies GPQA Diamond (93.5), HLE with tools (59.4), Terminal-Bench 2.1 (85.0), DeepSWE (67.7) and Toolathlon-Verified (74.1), all self-reported. As of 2026-09-21 the boards for DeepSWE, Terminal-Bench 2.1 (tbench.ai), Toolathlon, Agents' Last Exam, ARC Prize, LiveBench and MathArena did not list the model; Artificial Analysis's own Terminal-Bench 4.0 reading (33.3) sits close to vals.ai's tracked 31.82.
Is Step 5 Preview open source?
Yes, as of 2026-10-15. StepFun posted BF16, FP8 and NVFP4 checkpoints on Hugging Face (stepfun-ai/Step-5-Preview-BF16 and the sibling repos) under the StepFun Model License 1.0. The license allows use, modification and redistribution, and it adds two conditions: a "Powered by StepFun AI" notice once a commercial product passes 100 million monthly users or $20 million in revenue, and a separate commercial agreement before a device maker ships the model as a feature or a large model-as-a-service operator serves it. The copies labeled Step-5-Preview-BF16 that appeared on 2026-09-20 were not this release. This page's license field says open-weights because the weights are public; the terms are not MIT.
What does Step 5 Preview cost?
StepFun's international USD price list (checked 2026-09-21) charges $1.00 per million input tokens on a cache miss, $0.05 on a cache hit and $2.70 per million output tokens, with reasoning tokens billed as output. It is a single flat rate: no length threshold, time-of-day pricing or promotional price appears in StepFun's docs. StepFun's China-region site lists the same model in yuan (¥7, ¥0.35 and ¥20 per million tokens).
How does Step 5 Preview compare with the frontier on coding?
There is no independent evidence on the coding benchmarks this site tracks yet. StepFun's own launch table puts its DeepSWE score (67.7) within that benchmark's 9.5-point noise band of Kimi K3 (67.5), GLM-5.3 (66.9), GPT-6 Astra (74.1) and Claude Opus 5 (74.0) — none of those distances would count as a gap here even if independently run — and the table is StepFun's, and its DeepSWE run used the SWE-agent harness, not the mini-swe-agent the independent DeepSWE board applies to every model. That board does not list the model, so this site calls none of those gaps. Artificial Analysis reports 33.3 on the tracked Terminal-Bench 4.0, where the 31.82 recorded here is vals.ai's independent run of this model.
What can Step 5 Preview take as input, and how much can it write?
Text, images and video in; text out. The context window is 1M tokens and the maximum output is 64K tokens — StepFun's docs listed 1M for maximum output at launch and changed it to 64K within two days. Reasoning effort can be set to low, medium or high; StepFun's benchmark figures use high.
Did Step 5 Preview leak before StepFun announced it?
Not the weights, on the evidence this site could check. Artificial Analysis had a page for the model, with results, a price and a 2026-09-18 release label, in an archive capture from 05:44 UTC on 2026-09-19 — at least 21 hours before StepFun's announcement post. AA's page does not say when it was given access, and this site does not read that gap as a leak. The first of several unofficial copies labeled as its weights appeared on Hugging Face about two hours after the post, at 05:07 UTC on 2026-09-20. StepFun's site, docs and X account say nothing about either, and this site found no anonymous OpenRouter, OpenCode or Arena listing tied to the model.
Can I use Step 5 Preview for free?
Yes, on OpenCode Zen, which has listed step-5-preview-free at $0 input, $0 output and $0 cache read since 2026-10-08, with a zero-retention policy and a 65,536-token output limit. The start date is the only firm figure: OpenCode calls the route free for a limited time and publishes no end date (checked 2026-10-09). It is a gateway subsidy, not a StepFun price cut — StepFun still charges $1.00/$2.70 per million tokens, and OpenRouter charges the same.
Further reading
- Models with 10M token context windows 2026 — Step 5 Preview is one of the 49 models it compares.
Benchmark guides
- Humanity's Last Exam — what Step 5 Preview’s reasoning score on it does and doesn’t prove.
- Terminal-Bench 2.1 — what Step 5 Preview’s agentic score on it does and doesn’t prove.