Step 3.5 Flash
Language Available Full comparison ↗ stepfun/step-3.5-flash · by StepFun
· mixture-of-experts
Pricing — 1 offering(s)
Input tokens
- $0.10 / 1M tokens (input) Current 2026-01-29 → present
Output tokens
- $0.30 / 1M tokens (output) Current 2026-01-29 → present
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Step 3.5 Flash fits into a cost-aware routing setup
See how →Capability profile
How Step 3.5 Flash rates across core capability dimensions, with the task-level evidence behind each rating.
StepFun's technical report: BBH 88.2 (within 0.5 of the best listed), MMLU 85.8, IMO-AnswerBench 85.4. Strong for an 11B-active model — "retains the memory of a 196B model at the speed of an 11B one".
LiveCodeBench-v6 (2024.08-2025.05) 86.4, Terminal-Bench 2.0 51.0 (technical report).
"Fast, Sharp & Reliable Agentic Intelligence" is StepFun's own tagline; tau2-Bench 88.2, BrowseComp 69.0 (with context management) in the report. Function calling supported.
~256-262K tokens (StepFun docs / OpenRouter / technical report).
Chinese and English focus; broader coverage not specifically claimed.
11B active parameters — 100-300 tok/s typical, peaks ~350 tok/s on coding tasks per the technical report. The point of the 'Flash' tier.
$0.10 / 1M input, $0.30 / 1M output on StepFun's own developer platform (a $0.02/1M cache-hit input rate exists but isn't modelled). Among the cheapest models in this registry for this benchmark level.
Operator guidance
A strong value pick: frontier-adjacent benchmark numbers (from StepFun's own technical report) at one of the lowest price points in this registry, backed by a real published report rather than marketing claims. Prefer it over pricier flagships when the task tolerates an 11B-active model and cost dominates. Note the report's benchmarks are StepFun-run; no independent third-party reproduction found yet.
Use cases
- Cost- and latency-sensitive agentic and coding workloads that still need frontier-adjacent reasoning quality
- High-throughput production serving where the 11B active footprint keeps per-token cost very low
- Long-context tasks (up to ~256K) at a fast, cheap tier
- Self-hosted deployment (open weights on GitHub)
Limitations
- Benchmarks are from StepFun's own technical report — credible and detailed, but not independently reproduced
- A later model, Step 3.7 Flash (198B MoE, vision-language), exists — this entry is specifically Step 3.5 Flash
- The pricing-page row `step-3.5-flash` is a substring of an adjacent dated `step-3.5-flash-2603` row (see registry note) — harmless while prices match
- Multilingual/instruction-following not separately benchmarked in the report