Put production traffic on Fireworks AI only after you have chosen capacity per workload and planned for its limits. Decide Serverless or dedicated GPUs, retry 429s and 503s, keep the balance from pausing usage, reserve capacity before a model upgrade, and clear identity, retention and audit checks.
When you need this
- An app or agent that called an open model on Fireworks Serverless in development is moving to production traffic.
- A team is about to make a newly launched open model its production default.
- Security asks where prompts go, who can sign in, and what is logged.
Evaluation criteria
Choose capacity per workload
Which traffic goes to Serverless, and which to dedicated GPUs?
Look forServerless shares a pool of GPUs with models pre-provisioned and bills per token. On-demand deployments reserve GPUs billed per GPU second, with an H100 80 GB at US$8.00 per hour on November 1, 2026. Region-restricted deployments cost a 1.5x premium.
Plan for limits and errors
What does a client do with a 429 or a 503?
Look forServerless limits are tokens per minute per model. Stay below the adaptive limits to avoid 429s, and move to the Priority tier to reduce the likelihood of 503s. An account shares one envelope of up to 6,000 requests per minute, and on a dedicated deployment a 429 means its capacity is saturated.
Keep billing from stopping traffic
What happens when the balance or the spend limit runs out?
Look forUsage pauses at a zero balance unless Auto Reload is on, and all requests pause at 100% of the monthly spend limit except on Enterprise accounts. Contracted customers can ask about post-paid billing.
Reserve capacity before a model upgrade
Can a new model carry your traffic on day one?
Look forGitLab's incident record of March 24, 2026 shows a 12.03% completion error rate when Fireworks could not scale a new on-demand Codestral deployment. GitLab redirected traffic to the previous deployment and moved to reserved deployments. Keep the previous deployment ready.
Check API compatibility
Will your client's requests pass validation?
Look forThe Anthropic-compatible endpoint serves only POST /v1/messages and runs no server-side tools. Developers report 400 errors for OpenAI message fields after several agent steps (August 26, 2025) and for tool fields on the Anthropic-compatible API (April 27, 2026). Test your client's full request shape.
Clear identity, retention and audit
What does security need on the account?
Look forGoogle SSO is the default. OIDC or SAML SSO, SCIM, audit logs and data residency need an Enterprise account, and a Fireworks representative enables SSO and SCIM. Zero Data Retention is the default for open models, and the Response API keeps stored conversations up to 30 days unless store=False.
Questions to ask vendors
- Which workloads stay on Serverless, and which move to on-demand deployments?
- How do our clients retry 429s and 503s, and do we need the Priority tier?
- Is Auto Reload on, and is the monthly spend limit above our expected peak?
- Who confirms reserved capacity with Fireworks before a new model becomes a default?
- Do we need an Enterprise account for custom SSO, SCIM, audit logs or data residency?
Red flags
- A new model made the production default on on-demand capacity with no redirect to the previous deployment.
- Production traffic on prepaid credits with Auto Reload off.
- A monthly spend limit set below expected peak on a self-serve account.
- Clients that treat a 429 or 503 as a final failure.
- Anthropic-format agents that rely on server-side tools.
Products in this piece
Sources
- Official documentation
- 14
- Public source
- 5
- Fireworks AI: Serverless latency FAQ
- Fireworks AI: Pricing
- Fireworks AI: On-demand deployments
- Fireworks AI: Serverless rate limits
- Fireworks AI: Serverless rate limits (text)
- Fireworks AI: Account quotas
- Fireworks AI: Inference error codes
- Fireworks AI: Billing and credits FAQ
- Fireworks AI: Cost FAQ
- Fireworks AI: Anthropic compatibility
- Fireworks AI: SSO and SCIM
- Fireworks AI: Data handling
- Fireworks AI: Audit logs
- Fireworks AI: Data residency
- 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)
- 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)