Step 3.7 Flash: Vision, Effort Levels and a Higher Output Price

Step 3.7 Flash is StepFun's successor to Step 3.5 Flash, and the short version of what changed is this: it can see, it lets you choose how hard it thinks, and its output tokens cost nearly four times as much as listed. Whether that is a good trade depends almost entirely on whether your workload has images in it.

OpenRouter lists it from late May 2026. Like its predecessor, the weights are published under Apache 2.0. This guide covers what is new, how to read a benchmark table that shares almost nothing with the previous one, the price shape, and when the older model is still the better pick.

What changed: a vision encoder and three effort levels

StepFun's model card describes a 198B-parameter sparse mixture-of-experts vision-language model with about 11B active per token. It pairs a 196B language backbone, the same size as Step 3.5 Flash, with a 1.8B vision encoder. Context stays at 256K tokens.

So the language side is the same scale and the same active count. The two additions are what you are paying for:

• Native image input. The vendor card describes native image understanding. OpenRouter's listing also shows video input; the card does not mention video, so test before relying on it.
• Three reasoning levels, low, medium and high, exposed so you can trade depth for latency and cost per request rather than per deployment.

StepFun quotes throughput of up to 400 tokens per second, up from the 350 peak it gave for 3.5 Flash. Treat both as best-case figures from the vendor's own serving stack.

A benchmark table that does not line up

The 3.7 card reports a different set of evaluations from the 3.5 card, which makes a direct before-and-after harder than it looks:

• SWE-Bench Pro: 56.3
• Terminal-Bench 2.1: 59.5
• Toolathlon: 49.5
• HLE with tools: 48.1
• GDPval-AA: 45.8
• SimpleVQA (search): 79.2, V* (Python): 95.3

Step 3.5 Flash reported SWE-bench Verified and Terminal-Bench 2.0. Step 3.7 reports SWE-Bench Pro and Terminal-Bench 2.1. These are different tests. A 59.5 on Terminal-Bench 2.1 is not an eight-point gain over 51.0 on 2.0, because 2.1 revised 28 of the benchmark's 89 tasks; the real difference may be more or less than that. The same goes for SWE-bench Verified against SWE-Bench Pro, where Pro is the harder benchmark and scores run lower across the field.

The shift itself is informative. StepFun has moved its headline numbers towards long-horizon agent work and professional tasks, and added vision evaluations. That is the job it wants 3.7 Flash judged on. To compare the two models on your own work, run both through the same harness on the same tasks; building a 20-task eval covers how.

Price: output is where the gap is

StepFun lists 3.7 Flash at $0.20 per million input tokens on a cache miss, $0.04 on a cache hit, and $1.15 per million output tokens, and OpenRouter showed the same rates at the time of writing. Step 3.5 Flash was listed at $0.10 input and $0.30 output.

Input doubled. Output went up almost four times. For agent loops, where input usually dominates token counts, the real increase sits somewhere between those two, and the cache rate pulls it back down: a loop that resends a stable system prompt and tool definitions pays a fifth of the input rate on the cached part. Input vs output token pricing covers how to work out which side of the bill your workload lands on.

The reasoning levels interact with this directly. High effort produces more output tokens, which are now the expensive ones. Default to low or medium and move up only where you have measured a gain.

Same footprint, plus a projector

Local deployment looks much like 3.5 Flash. The card lists GGUF language-model files from about 102.5 GB at Q3_K_L to 111.5 GB at Q4_K_S, plus a 3.97 GB multimodal projector, with about 7 GB of runtime overhead. Minimum memory is 120 GB, with 128 GB recommended, on machines such as Mac Studio, a MacBook Pro with 128 GB, AMD Ryzen AI Max+ 395 systems or a DGX Station.

For servers, StepFun ships BF16, FP8 and NVFP4 builds; its vLLM examples use eight-way tensor parallelism for BF16 and FP8, and four-way for NVFP4. Transformers support needs version 5.0 or later.

When 3.5 Flash is still the better choice

The upgrade is not automatic. A decision rule:

• Pick 3.7 Flash when the workload includes screenshots, diagrams or documents as images; when you want per-request control over reasoning depth; or when your own tests show it finishing agent tasks in fewer turns, which can offset the higher output rate.
• Stay on 3.5 Flash for text-only, high-volume work where output tokens dominate the bill and 3.5 already passes your tests. Paying for a vision encoder you never use is waste.
• Compare against other vision models before committing, since multimodal is now the selling point. Vision-language models compared lists what to test.

The practical next step is a fifty-request sample of your real traffic, run through both models at matched settings, scored on task completion and total cost. That answer will be more reliable than either benchmark table.