StepFun Step 3.7 Flash Review: We Tested the 1/9-Cost Claim

StepFun's Step 3.7 Flash arrived with an eye-catching pitch: roughly 97% of Claude Opus's coding ability at about one-ninth the cost, from an open Apache-2.0 model you can run on a single 128GB machine. That's exactly the kind of claim worth executing rather than repeating, so we ran it through our coding benchmark. The short version: the per-token price really is low - but the model is slow and verbose enough that the per-task economics look nothing like 1/9, and cheaper, better coders already exist.

What Step 3.7 Flash is

Step 3.7 Flash is an open-weights model from StepFun (阶跃星辰), released around May 28-29, 2026 under Apache 2.0. It's a ~198B-parameter Mixture-of-Experts — a 196B language backbone plus a 1.8B vision encoder, ~11B active per token — so it's both a coding model and a vision-language model (text + image in, text out). It has a 256K context, selectable reasoning effort, and is small enough to run on a single 128GB unified-memory device.

The claim we tested

StepFun's headline pitch is specific and falsifiable: that Step 3.7 Flash (with its "Advisor" mode) reaches ~97% of Claude Opus 4.6's coding ability at roughly 1/9 the per-task cost, alongside vendor benchmarks of SWE-bench Verified 76.5%, SWE-bench Pro 56.3%, and Terminal-Bench 2.1 59.6%. Those are StepFun's own numbers. We don't repeat them as fact — we ran the model ourselves on the same nine-task coding harness as our coding-cost benchmark and measured real cost, latency, and pass rate.

Results: 8/9, and mid-pack cost

Step 3.7 Flash solved eight of nine — it missed only the valid-parentheses check — and, crucially, it was neither fast nor cheap in practice:

So on pure coding, Step 3.7 Flash is mid-table: a notch behind the 9/9 models on correctness (it missed one), and one of the slower and pricier options per task. The headline "1/9 the cost of Opus" did not hold up the way the sticker price implies.

Cheap per token vs cheap per task

Here's the trap the marketing leans on. Step 3.7 Flash's per-token price is genuinely low ($0.20/$1.15 vs Opus's $5/$25 — roughly 1/15 to 1/20 per token). But per-task cost = price × tokens generated, and Step 3.7 Flash generated long, slow outputs (~19 seconds per task). Multiply a low rate by a high token count and you get a mid-pack bill — about $2.66 per 1,000 tasks, only ~1.5× cheaper than Opus, not 9×. Meanwhile Claude Opus answered tersely and quickly, shrinking its apparent cost disadvantage. Sticker price is not spend — the only way to know real cost is to run the tasks, which is exactly why we do.

The real draw: vision + Apache 2.0

If you only need to write code, cheaper text coders beat Step 3.7 Flash on every axis we measured. Its genuine differentiators are elsewhere: it's natively multimodal (it reads images, not just text) and it's Apache-2.0 open-weights that fits on a single 128GB machine. For a self-hosted assistant that needs to reason over screenshots, diagrams, or mixed media — with no per-token lock-in — that combination is rare at this size. Pick it for that, not for cheap coding throughput.

Pricing & access

On OpenRouter (stepfun/step-3.7-flash): $0.20 input / $1.15 output per 1M tokens, no free tier. The weights are Apache-2.0 on Hugging Face (with GGUF builds) for self-hosting. Budget by real output volume, not the sticker rate — for verbose models the two diverge sharply.

Who should use it

FAQ