Surogate 3.7 Flash

Every agent needs a model under it. Renting one means someone else's prices, someone else's region and someone else's rules about your prompts. Surogate 3.7 Flash is ours, built for the work agents do: reading long case files, calling tools and filling in structured records.

What it does

Surogate 3.7 Flash reads text and images and writes text. It holds up to 1M tokens of context and writes up to 32K tokens in one answer. It supports tools, prompt caching and structured outputs, so an agent can call your systems and hand back records your software can read without parsing prose.

It is listed on llm.buzz as surogate/surogate3.7-flash.

How it scores

WildClawBench runs a model through 60 real agent tasks in the OpenClaw harness. On the public leaderboard, read on 29 October 2026, Surogate 3.7 Flash scores 54.6.

The Surogate score is our own run. The others are as the leaderboard lists them.

Where it runs

Two providers in the EU serve it. Neither trains on prompts and both keep them for zero days.

• Surogate Spark EU: $0.15 per million input tokens, $1.00 per million output. 0.04 s to the first token.
• Simpleminds EU: $0.10 input, $0.75 output, in FP4. 0.07 s to the first token.

These are the prices llm.buzz showed on 27 October 2026. They move, and the live numbers are on the model page.

You can also run it yourself. Our engine serves it in NVFP4 on a single NVIDIA DGX Spark, and on the same box it beats vLLM:

The last row is the one that matters for agents: many of them at once, each with a long case open.

Call it

One key works for every model on llm.buzz, through an OpenAI compatible API:

curl https://api.llm.buzz/api/v1/chat/completions \
  -H "Authorization: Bearer $LLM_BUZZ_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "surogate/surogate3.7-flash",
       "messages": [{"role": "user", "content": "Hello"}]}'

Inside Surogate, the AI Gateway can route any agent to it.

Part of a bigger stack

Surogate 3.7 Flash sits next to the rest of our models: Surogate Rune for decisions on text and images, and Surogate Speech for agents that listen and talk. We train all of them with the same open source engine we give our clients, so the model you run can be fine tuned on your own sessions, inside your own network.