Step 3.5 Flash
Step 3.5 Flash is a StepFun open-weights reasoning model, released 2 Apr 2026. Its best configuration ranks #269 of 775 on the BenchLeader Index at 53.8 ±3.8, in the upper half. The ± is the point: 279 other configurations score within that range, so they and this one cannot be told apart on quality alone — price and speed are what separate them. It scores highest in instruction following (66) and lowest in knowledge (42). At $0.150 per million tokens blended it is among the cheapest fifth of ranked models. Output speed of 145 tokens per second puts it in the fastest quarter, with a first answer in 18.8 s. Last measured 15 Oct 2026.
- Blended price
- $0.150/M
- $0.100 in · $0.300 out
- Output speed
- 145 tok/s
- measured by Artificial Analysis
- First answer
- 19 s
- first token 1.18 s
- Context
- 256k
- Full answer
- 23 s
- median, reasoning included
- Cost per run
- –
- Released
- 2 Apr 2026
- Overall index54
- Reasoning49
- Coding58
- Agents & tools62
- Maths54
- Knowledge42
- Instruction following66
- Human preference57
- Long context56
- Composite49
Versions
StepFun has shipped 2 models under this name. Each is ranked on its own results; a newer version often has fewer results so far, which holds its index nearer the average until more arrive.
| Model | Released | Index | Rank |
|---|---|---|---|
| Step 3.7 Flash | 29 May 2026 | 51.5 | #341 |
| Step 3.5 Flashthis page | 2 Apr 2026 | 53.8 | #269 |
How its index has moved
9 Sept 2026 to 19 Oct 2026The index is recomputed from scratch every day, so a line moves when a new benchmark result lands, when a publisher revises a score, or when the models it is normalised against change. Early movement usually means the score is still settling.
Benchmark results
See all LLM benchmarks. One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.
Reasoning
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| Arena Hard Prompts | 1412#163 | Arena (LMArena) | |
| GPQA Diamond (AA)not in index | 82.6%#167 | Artificial Analysis | |
| Humanity's Last Exam (AA)not in index | 24.5%#185 | Artificial Analysis | |
| CritPt | 2.3%#194 | Artificial Analysis |
Coding
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| Arena Coding | 1449#147 | Arena (LMArena) |
Agents & tools
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| Terminal-Bench Hard | 32.6%#97 | Artificial Analysis | |
| τ²-Bench Telecom (AA)not in index | 87.4%#77 | Artificial Analysis |
Maths
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| AIME 2026 | 96.7%#6 | MathArena | |
| HMMT February 2026 | 86.4%#20 | MathArena | |
| USAMO 2026not in index | 44.6%#7 | MathArena | |
| MathArena Apexpossible contamination | 13.5%#18 | MathArena | |
| Arena Maths | 1407#139 | Arena (LMArena) |
Knowledge
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| AA-Omniscience | -44.2#342 | Artificial Analysis | |
| AA-Omniscience: accuracynot in index | 24.6%#236 | Artificial Analysis | |
| AA-Omniscience: non-hallucinationnot in index | 8.7%#449 | Artificial Analysis |
Instruction following
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| IFBench | 66.5%#94 | Artificial Analysis | |
| Arena Instruction Followingnot in index | 1384#158 | Arena (LMArena) |
Human preference
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| Arena Text | 1393#159 | Arena (LMArena) | |
| Arena Creative Writingnot in index | 1345#172 | Arena (LMArena) | |
| Arena Multi-turnnot in index | 1396#158 | Arena (LMArena) | |
| Arena Longer Queriesnot in index | 1403#158 | Arena (LMArena) |
Long context
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| AA-LCR | 63.0%#265 | Artificial Analysis |
Composite
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| AA Intelligence Index v4.3.2 | 17.0#261 | Artificial Analysis |
Where to run it
Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.
| Provider | Speed | First token | Input $/M | Output $/M | Context | Quantisation |
|---|---|---|---|---|---|---|
| SiliconFlow | 72 tok/s | 1.18 s | $0.100 | $0.300 | 262k | fp8 |
What a task costs
Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.020 per 1M. Reasoning tokens are not modelled.
| Workload | Tokens in / out | Cost | With caching | Time |
|---|---|---|---|---|
| Chat reply | 400 / 300 | $0.0001 | $0.0001 | 20.9 s |
| Summarize a 30-page report | 12,000 / 600 | $0.0014 | $0.0007 | 23.0 s |
| Code edit | 6,000 / 1,500 | $0.0011 | $0.0007 | 29.2 s |
| Agentic coding session | 60,000 / 4,000 | $0.0072 | $0.0036 | 46.4 s |
| Structured extraction | 2,000 / 200 | $0.0003 | $0.0001 | 20.2 s |
See also
Data as of 21 Oct 2026. Compare with another model.
Cite as: BenchLeader, “Step 3.5 Flash: benchmarks, pricing, speed and rank”, https://www.benchleader.com/models/step-3-5-flash, data as of 21 Oct 2026.