GILD verdict
Adopt Fireworks AI to serve open models through an OpenAI-compatible API, billed per token on shared Serverless or per GPU second on dedicated deployments.
It fits teams ready to keep Auto Reload on, retry 429s and 503s, and reserve capacity before a new model carries production traffic, since GitLab redirected Codestral traffic when an on-demand deployment would not scale.
Fireworks AI
Inference Hosting by Fireworks AI
−17 MixedFit and limits
Who it fits
- You run open models such as GLM, Kimi, DeepSeek or gpt-oss and want one OpenAI-compatible API for them, with the OpenAI Python client and frameworks such as LangChain, LlamaIndex, CrewAI and AutoGen.
- Prompts must stay unstored: Fireworks has Zero Data Retention by default and does not log or store prompt or generation data for open models without explicit opt-in.
- Traffic is spiky enough that per-token Serverless beats renting GPUs, and offline jobs on models that support the Batch API can run at 50% off Serverless per-token prices.
- You want to fine-tune an open model and serve it at base-model prices with the same vendor.
Where it falls short
- Serverless is a shared pool. Fireworks says users see minor latency and performance variability during high-volume periods, and Serverless can return 429 Too Many Requests or 503 Service Overloaded.
- A new model may not scale on demand. GitLab's incident record of March 24, 2026 put a 12.03% code-completion error rate on Fireworks being unable to scale a new Codestral 2508 deployment.
- One developer reports a price rise of about 30% (September 22, 2026), another found cache reads on GLM 5.3 Flash cost more than on DeepSeek (September 24, 2026), and a third pays 3x to 5x DeepSeek's own price for zero-retention hosting (July 19, 2026).
- The compatible APIs have gaps: the Anthropic-compatible endpoint serves only POST /v1/messages without server-side tools, and developers report 400 errors on OpenAI and Anthropic request fields.
Implementation reality
Pick capacity for each workload
Put variable traffic on Serverless and steady or latency-critical traffic on on-demand deployments, which reserve dedicated GPUs billed per GPU second. An H100 80 GB lists at US$8.00 per hour, and region-restricted deployments cost a 1.5x premium.
Plan for limits and errors
Serverless limits are tokens per minute per model. Stay below the adaptive rate limits to avoid 429s, and move to the Priority tier to reduce the likelihood of 503s. All API usage on an account shares one envelope of up to 6,000 requests per minute with a payment method and active credits.
Keep billing from stopping traffic
Usage pauses when the balance reaches zero and Auto Reload is off, and all API requests pause at 100% of the monthly spend limit, which does not apply to Enterprise accounts. Turn on Auto Reload and size the spend limit above expected peaks.
Reserve capacity before a model upgrade
GitLab ran a new Codestral model on on-demand deployments with legacy direct routing, redirected traffic to the older deployment when it would not scale, and moved to reserved deployments behind Fireworks' standard API gateway the same day.
What to verify before you buy
Limits and errors
- Variability: Fireworks says serverless users are subject to minor latency and performance variability during high volume periods.
- 429 and 503: When using Serverless, you may experience 429 Too Many Requests or 503 Service Overloaded. To avoid 429s, stay below the adaptive rate limits. To reduce the likelihood of 503s, upgrade to Priority tier.
- Account request limit: All API usage on an account shares one request-throughput envelope: 10 RPM with no payment method or no credits, and 6,000 RPM (maximum) with a payment method and active credits. The 6,000 RPM cap applies account-wide and is a fixed, non-adaptive ceiling.
- Dedicated deployments: The inference error codes page says there are no account-level rate limits on dedicated and on-demand deployments, and that a 429 there indicates the deployment's processing capacity is saturated. The account quotas page says on-demand and dedicated deployments are not limited by adaptive serverless TPM upper bounds, and that their requests still count toward account-wide request limits.
Billing
- Zero balance: If your balance reaches zero and Auto Reload is not enabled, usage pauses until you add credits.
- Spend limit: When usage reaches 100% of the monthly spend limit, all API requests pause automatically across serverless inference, deployments, and training. This suspension does not apply to Enterprise accounts.
- Post-paid: Contracted customers may have the option to move to post-paid billing.
Identity, data and audit
Custom SSO, SCIM and audit logs need an Enterprise account.
- SSO and SCIM: By default, Fireworks supports Google SSO. If you have an enterprise account, Fireworks supports bringing your own identity provider using an OpenID Connect (OIDC) or SAML 2.0 provider. Coordinate with your Fireworks AI representative to enable the integration. SCIM provisioning is available for enterprise accounts and works with supported directory providers, including Okta, Microsoft Entra ID, and Google Workspace. Custom SSO must be configured before you can enable SCIM provisioning. To request SCIM enablement, contact your Fireworks AI representative, and Fireworks will enable Directory Sync for your account. API keys and service accounts are not blocked by SSO enforcement.
- Retention and training: Fireworks has Zero Data Retention by default. Fireworks does not log or store prompt or generation data for any open models, without explicit user opt-in. For the Response API, with
store=True(the default), stored conversation data is kept for up to 30 days. Settingstore=Falseprevents any conversation data from being retained. Across every training surface, your training data is never used to train Fireworks-owned or shared models. - Certifications: Fireworks has achieved all three ISO certifications covering security, privacy, and AI management: ISO 27001, ISO 27701 and ISO 42001. Fireworks lists SOC 2 Type II as certified. Fireworks is HIPAA compliant.
- Audit logs: Audit logs are available for Enterprise accounts. Audit logs include data access logs. All read, write, and delete operations on storage are logged, normalized, and enriched with account context. You can view audit logs with the Fireworks CLI or retrieve them with the List Audit Logs API.
Regions and compatibility
- Regions: By default, Fireworks deployments are multi-region, and Fireworks can move and spread them across regions as needed. Supported multi-regions are GLOBAL, US, CANADA, EUROPE and APAC. New accounts receive GPU quota for the GLOBAL multi-region only. The US, CANADA, EUROPE and APAC multi-regions and all single regions start at zero quota and must be granted by Fireworks. The
--regionflag pins a deployment to a single region, and a deployment with a--regionyou have no quota for is rejected at creation. - Data residency: On Enterprise accounts, data residency restricts inference on the account to a single region. Contact your Fireworks representative if you need this enabled. The listed restriction is US, on
us.api.fireworks.ai. For a region that is not listed, contact Fireworks sales. - Anthropic compatibility: Fireworks Anthropic compatibility serves only POST /v1/messages and does not support Anthropic server side tools.
What to measure in your pilot
- Load-test each model at your expected concurrency on Serverless and on a dedicated deployment, and record 429s, 503s and latency before you pick one for production.
- Track your own spend per model weekly, including cached input tokens, since practitioners report a price rise and cache rates that differ by model.
- Before a new open model becomes a default, confirm reserved capacity with Fireworks and keep the previous deployment ready as a redirect.
Sources
- Official documentation
- 33
- Public source
- 9
- Fireworks AI: Pricing
- Fireworks AI: Billing and credits FAQ
- Fireworks AI: Quickstart
- Fireworks AI: OpenAI compatibility
- Fireworks AI: Agent frameworks
- Fireworks AI: LLM gateways
- Fireworks AI: Microsoft Foundry
- Fireworks AI: Cost FAQ
- Fireworks AI: Nexus Auto
- Fireworks AI: Nexus open models
- Fireworks AI: Serverless pricing
- Fireworks AI: Serverless pricing (text)
- Fireworks AI: Batch inference
- Fireworks AI: Introduction
- Fireworks AI: Serverless latency FAQ
- Fireworks AI: US-only Serverless
- Fireworks AI: On-demand deployments
- Fireworks AI: Regions
- Fireworks AI: Data residency
- Fireworks AI: Serverless overview
- Fireworks AI: Embeddings
- Fireworks AI: Vision models
- Fireworks AI: Tool calling
- Fireworks AI: Serverless rate limits
- Fireworks AI: Anthropic compatibility
- Fireworks AI: Serverless rate limits (text)
- Fireworks AI: Account quotas
- Fireworks AI: Inference error codes
- Fireworks AI: SSO and SCIM
- Fireworks AI: Data handling
- Fireworks AI: Secure training
- Fireworks AI: Data security
- Fireworks AI: Audit logs
- Cursor blog: Editing Files at 1000 Tokens per Second (May 14, 2024)
- GitLab merge request !4997: redirect Codestral 2508 to 2501 (Mar 24, 2026)
- GitLab production incident 21621, INC-8707 (Mar 24, 2026)
- GitLab merge request !4844: move Fireworks from direct routing to the standard API gateway (Mar 24, 2026)
- Hacker News, pkulak: Fireworks raising a price by about 30% (Sep 22, 2026)
- Hacker News, vardalab: GLM 5.3 Flash cache rates on Fireworks (Sep 24, 2026)
- Hacker News, bwfan123: zero-data-retention hosts at 3x to 5x DeepSeek's price (Jul 19, 2026)
- GitHub, earendil-works/pi issue #3854: 400 errors on tool requests through anthropic-messages (Apr 27, 2026)
- GitHub, mozilla-ai/any-agent issue #778: invalid_request_error with Fireworks models (Aug 26, 2025)