One API. Intelligence allocated to the work.
FUMO Lab builds Fusion, a compound intelligence system available through OpenAI-compatible Chat Completions and Responses APIs.
An agent's needs change as work unfolds. A tool result may settle a question; a failed action may reveal a missing constraint; a working implementation may need verification rather than another solution. Fusion allocates model capability to the uncertainty that matters at the current step, then produces one authoritative response or tool call.
Get started · View pricing · See how Fusion works
Why Fusion
Fusion selects capabilities for each stage of a task: which model can help, what evidence it should provide, and when that evidence is sufficient to act.
Fusion is designed around four properties:
- State-aware allocation — new observations can change the next model, role, context view, or compute budget.
- Complementary capability — a model can be useful as a solver, explorer, critic, or verifier without being the best standalone answerer.
- One authoritative result — contributions are evaluated together, and only the final response can issue a client-visible action.
- One integration surface — clients receive one result through the Chat Completions or Responses interface they already use.
Choose a model ID
The models below share the same API key and Base URL. Set model to a model ID authorized for your key.
Fusion models
| Product | Model ID | Product focus |
|---|---|---|
| Fusion Max 2.0 | fusion-max-2.0 | High-performance coding, agents and document analysis, built for cost efficiency. |
| Fusion Max 1.0 | fusion-max-1.0 | Complex reasoning, research, and demanding agentic workloads |
| Fusion Code 1.0 | fusion-code-1.0 | Coding, debugging, terminal work, and software-agent workflows |
How it works
Fusion organizes computation around decision-relevant uncertainty:
- Understand the current state — preserve the goal, constraints, observations, and acceptance conditions.
- Acquire useful evidence — choose a source, role, context view, and budget, or stop when another call would add little value.
- Commit one result — reconcile the evidence and return one answer or tool call under the original client contract.
The API is stateless. The client maintains conversation history, executes tools, and sends the full history on each turn.
Start building
- Get an API key. Online self-service customers can sign up for FUMO and create a key in the console. Enterprise customers can follow the Enterprise onboarding to set up an account and create a key.
- Configure your client. Enter your API key, Base URL, and model ID.
- Send a request. Run a quickstart example or configure your preferred Coding Agent.
If you use FUMO and have a sales representative, enter their Sales Code at signup or bind it later in Account. For enterprise purchasing or setup assistance, contact sales.
Send your first request · Configure Codex
Evaluation results
FUMO Lab's August 2026 whitepaper reports scores and full benchmark-run costs for GPQA Diamond, τ³-Banking, and Terminal-Bench v2.1. The results show the capability–cost trade-offs of the two Fusion profiles under the reported evaluation conditions. API pricing is quoted separately.
Read the results and limitations · Read the whitepaper↗
Fusion exposes Chat Completions and Responses. Anthropic Messages clients require a protocol adapter.
All three models are AVAILABLE with 128K maximum output. Max 2.0 / Max 1.0 have 1M context; Code 1.0 has 512K. Text/image input and text/tool-call output are supported. Agents read files; native file upload is not supported.