[sundar-flex]

Adds FlexAI (https://flex.ai) as an AI Gateway provider, following INTEGRATE_PROVIDER_TO_GATEWAY.md.

FlexAI is an OpenAI-compatible inference provider serving open-weight models. This PR registers the provider and adds endpoints for the six models FlexAI serves that already have a model definition in this registry — no new authors or model families, so nothing in the catalogue had to be invented.

Provider

Base URL	https://api.flex.ai/ → v1/chat/completions
Auth	Authorization: Bearer <key> (auth: "api-key", no override needed)
Pricing page	https://flex.ai/pricing
Models page	https://flex.ai/models
Priority	4 (default third-party)
PTB	ptbEnabled: false on every endpoint — BYOK only

Endpoints added

Every value below was measured against the live API on 2026-10-02, not copied from a listing.

Endpoint key	providerModelId	input $/M	output $/M	ctx / max out	quant	tools
gpt-oss-120b:flexai	gpt-oss-120b	0.03	0.17	131072	fp4	yes
gpt-oss-20b:flexai	gpt-oss-20b	0.02	0.10	131072	fp4	yes
llama-3.1-8b-instruct:flexai	Meta-Llama-3.1-8B-Instruct-FP8	0.02	0.05	131072	fp8	yes
llama-3.3-70b-instruct:flexai	Llama-3.3-70B-Instruct-FP8	0.135	0.40	131072	fp8	yes
mistral-nemo:flexai	Mistral-Nemo-Instruct-2407-FP8	0.019	0.03	131072	fp8	yes
qwen3-coder-30b-a3b-instruct:flexai	Qwen3-Coder-30B-A3B-Instruct-FP8	0.07	0.26	262144	fp8	yes

How the values were verified

Prices come from FlexAI's own GET /v1/models (pricing.input_per_mtok / output_per_mtok), and were then reconciled against billed cost on a live completion:

• gpt-oss-120b: prompt_tokens: 74, of which cached_tokens: 64; reported upstream_inference_prompt_cost: 5.9e-07. Hand-computed with cached tokens inside prompt_tokens and billed at the published cached rate: (74−64)×3e-8 + 64×4.5e-9 = 5.88e-07. ✅
• Llama-3.3-70B-Instruct-FP8: prompt_tokens: 42, cached_tokens: 16, reported 3.83e-06. (42−16)×1.35e-7 + 16×2.03e-8 = 3.835e-06. ✅

That reconciliation establishes two things the price table alone doesn't: cached tokens are counted inside prompt_tokens (so the registry's cache handling is correct), and the published cached rate is 0.15 × input on every model — hence cacheMultipliers: { cachedInput: 0.15 } on each endpoint.

Context / max output. /v1/models reports max_output_length equal to context_length on every row, which on its own would be an unverified claim, so the ceiling was probed directly. The API rejects max_tokens strictly greater than the context window ("max_tokens=131073 cannot be greater than the model's maximum context length of 131072 tokens"), so maxCompletionTokens is set equal to contextLength, which is the measured limit.

supportedParameters is the intersection of FlexAI's per-model supported_parameters with this repo's StandardParameter union. The one exception: Llama-3.3-70B-Instruct-FP8 omits min_p and logit_bias. FlexAI serves it with speculative decoding, which rejects both with a 400 even though /v1/models lists them (verified live 2026-10-06).

One correction worth flagging: FlexAI's /v1/models metadata does not list tool_use for Mistral-Nemo-Instruct-2407-FP8, but the deployment emits a well-formed get_weather tool call on two independent live requests (finish_reason: "tool_calls"). The measurement is what's declared here, and tools / tool_choice are included for it.

Two things were deliberately left out rather than guessed:

• rateLimits — FlexAI publishes a per-account request-rate limit, not the per-model tpm/tpd shape this field wants, so declaring numbers here would be fiction.
• Models with no existing definition in the registry (FlexAI also serves GLM, DeepSeek V4, Qwen 3.5/3.6/3.8, Gemma 4, Nemotron, MiniMax, Step, embeddings, image, ASR and TTS models). Adding those means adding model families, which belongs in a separate PR rather than being smuggled into a provider registration.

Qwen3-30B-A3B-Thinking-2507 is served by FlexAI but is not mapped onto this registry's qwen3-30b-a3b (41k context, non-thinking Instruct weights) — a different model, so it is omitted rather than mis-attached.

Registration points touched

Found by grepping an existing comparable provider (canopywave, deepinfra) repo-wide and diffing the commit that added Canopy Wave (#5292):

1. packages/cost/models/providers/flexai.ts — provider class
2. packages/cost/models/providers/index.ts — import, registry entry, ResponsesAPIEnabledProviders (chat-completions-only group; BaseProvider.buildRequestBody already maps RESPONSES → chat completions)
3. packages/cost/models/providers/priorities.ts
4. packages/cost/providers/mappings.ts — URL pattern, providersNames, provider row
5. packages/cost/models/provider-helpers.ts — heliconeProviderToModelProviderName, dbProviderToProvider
6. packages/cost/usage/getUsageProcessor.ts — OpenAI usage processor
7. web/data/providers.ts + web/public/assets/home/providers/flexai.webp (512×512, from FlexAI's brand mark)
8. endpoints in four existing author families
9. worker/test/setup.ts provider-key fixtures (both BYOK and Helicone-key blocks)
10. worker/test/ai-gateway/registry-{openai,meta-llama,mistralai,alibaba}.spec.ts
11. packages/__tests__/cost/__snapshots__/registrySnapshots.test.ts.snap — regenerated with npx jest --updateSnapshot (step 9 of the guide)

Checks run

Baseline measured on untouched main in the same install, then re-run on this branch.

Command	main	this branch
npx vitest run test/ai-gateway/registry-{openai,meta-llama,mistralai,alibaba}.spec.ts (worker)	4 files, 628 passed, 0 failed	4 files, 644 passed, 0 failed (+16 new, and registry-mistralai alone re-run after the tool-call fix: 101 passed)
npx jest __tests__/ (packages)	41 passed / 1 skipped, 578 passed	41 passed / 1 skipped, 578 passed, 8 snapshots passed
npx tsc --noEmit (worker)	exit 0	exit 0

Snapshot delta is exactly what you'd expect: totalProviders 21 → 22, totalEndpoints 329 → 335, { provider: "flexai", modelCount: 6 }.

Not verified

The gateway was not exercised end-to-end against the live FlexAI API — the registry and worker test suites use mocked upstreams, and I have no Helicone deployment to route a real request through. Everything above the transport (URL construction, auth header shape, prices, context, parameters, tool calls) was measured directly against api.flex.ai.

Happy to adjust naming, priority, or the endpoint set.

🤖 Generated with Claude Code

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[vercel]

@sundar-flex is attempting to deploy a commit to the Helicone Team on Vercel.

A member of the Team first needs to authorize it.

[cursor]

Bugbot is paused — on-demand spend limit reached

Bugbot uses usage-based billing for this team and has hit its on-demand spend limit.

A team admin can raise the spend limit in the Cursor dashboard, or wait for the next billing cycle to continue.

[cursor]

Bugbot is paused — your team has an unpaid invoice

Bugbot is paused because your team has an unpaid invoice.

Pay it from the billing page in the Cursor dashboard and Bugbot resumes automatically.

(requestId: 8bedc825-6029-4d14-a820-3c8ea8481f9c)

[cursor]

Bugbot is paused — your team has an unpaid invoice

Bugbot is paused because your team has an unpaid invoice.

Pay it from the billing page in the Cursor dashboard and Bugbot resumes automatically.

(requestId: 54009ad1-3542-42e6-b5ad-6997b0685aee)