Flow AI is managed inference for AI agents. One OpenAI- and Anthropic-compatible endpoint routes each request to the cheapest model that completes the task across many providers, billed at the provider's true cost plus a flat 2.5%.

How much does Flow AI cost?

Start with a 7-day free trial (no subscription needed). After that, a membership of $4.99/month or $45/year, then you pay each request's pass-through model cost plus a 2.5% buyer fee. Prompt-cache hits are billed at the provider's cache rate.

Which clients work with Flow AI?

Any OpenAI, Anthropic Messages, or OpenAI Responses (Codex) client — set the base URL to https://api.flowaiapi.com/v1 and use your fa- key with model 'auto'.

Harness guide

Flow AI in DeepSeek Harness

dsh ships a complete tool loop — web fetch, shell, files, sandbox, context compaction — and needs exactly one thing from outside: a model endpoint. Point it at Flow and its per-step router chooses among three routes on one key, while Flow's router picks the cheapest lane that completes each step and verifies the result.

Setup — five lines

$DSH_HOME/settings.yaml

llm-pi-ai:
  providers:
    flow:
      apiKeyEnv: FLOWAI_API_KEY          # your Flow key
      api: openai-completions            # streams tokens as they arrive
      baseURL: https://api.flowaiapi.com/v1
      models:
        - id: flow-1                     # balanced — the default
        - id: flow-1-savings             # max savings — cheapest capable lane
        - id: flow-1-quality             # quality first — best lane (add ensemble: true for two proposers + verifier pick)
      compat:
        supportsDeveloperRole: false
        maxTokensField: max_tokens

agent-default-model:
  provider: flow
  model: flow-1

The three ids are the three policies — the policy is the model name — so a harness can choose per step (savings for exploration, quality for the write-up) without a second key or a dashboard setting. Every response carries _flowaiapi: the lane used, the price paid, the routing candidates, the derived harness state and the verification verdict.

A real session, unedited

The question: “Look at buywhere.ai and its MCP server listings. Is it a good MCP server? Recommend how to test MCP servers.” What the harness did before any model wrote a word: read the site, the developer docs and the MCP endpoint; pulled the server's package.json, README, src/index.tsand its smoke-test script from GitHub; then ran the protocol itself — registered an account, initialised the MCP session, listed the tools, called one. The answer was a synthesis of ~120K tokens of evidence it had gathered. That is why it was good; the model was the cheapest lane that could hold the context.

session · 16 steps · 14 tool calls · 412 s of harness time

turn·step	tool	target	result	step time	lane (Flow)
1·1	web_fetch	buywhere.ai	13 KB	9.6 s	GLM Coding Max
1·2	web_fetch	buywhere.ai/developers	4.6 KB	13.2 s	GLM Coding Max
1·3	web_fetch	api.buywhere.ai/mcp	0.7 KB	9.6 s	GLM Coding Max
1·4	web_fetch	api.buywhere.ai/docs/guides/mcp	9.6 KB	12.9 s	GLM Coding Max
1·5	web_fetch	github.com/BuyWhere/buywhere-mcp	4 KB	13.3 s	GLM Coding Max
1·6	web_fetch	…/buywhere-mcp/main/package.json	2.8 KB	12.4 s	GLM Coding Max
1·7	web_fetch	…/buywhere-mcp/main/README.md	19.1 KB	8.6 s	GLM Coding Max
1·8	web_fetch	…/buywhere-mcp/main/src/index.ts	5.3 KB	11.2 s	GLM Coding Max
1·9	web_fetch	…/buywhere-mcp/main/scripts/smoke-tools.mjs	3.7 KB	35.1 s	GLM Coding Max
1·10	bash	curl POST api.buywhere.ai/v1/auth/register (registers an account, gets a key)	0.2 KB	15.7 s	GLM Coding Max
1·11	bash	curl POST api.buywhere.ai/mcp · initialize + tools/list	27 KB	16.4 s	GLM Coding Max
1·12	bash	curl POST api.buywhere.ai/mcp · tools/call	19.7 KB	78.0 s	GLM Coding Max
1·13	synthesis	turn failed here — a Flow bug, since fixed (see below)	—	111.1 s	GLM Coding Max
2·1	bash	curl POST api.buywhere.ai/mcp · re-run with the key	1.3 KB	18.4 s	MiniMax-M3
2·2	bash	curl POST api.buywhere.ai/mcp · tools/call	15.9 KB	13.1 s	MiniMax-M3
2·3	synthesis	the answer — 43K tokens in, 2,476 out	—	33.8 s	MiniMax-M3

Flow requests

input tokens

124K + 213K cached

output tokens

4,832

total cost

$0.0042

Honesty note: turn 1 died at its final step. That was a Flow bug — the Responses stream could not serialise a value from one lane's pricing and closed the connection, which dsh reported as server_error: [object Object]. Fixed the same day; the recommended openai-completionssetting above also streams tokens progressively, which the Responses path does not.

What Flow adds to a harness

Routing on harness state

Each step is routed on the state of the loop — step index, tool failures, recovery loops, context pressure, the verifier's last verdict — derived by Flow when the harness does not send it, or sent explicitly as harness_state.

Verification on every step

A reward model we trained on execution truth scores each response (pass/fail with confidence, ~50 ms). Add ensemble: true on the quality route for two proposers and a verifier pick.

Price, not markup

Subscription, free-tier and marketplace supply at the lane's live price. The session above cost less than half a cent.

All harness guides · Docs