Key findings
- Cursor reports about 1000 tokens per second on its fine-tuned 70b fast-apply model with Fireworks (May 14, 2024), and credits its own speculative edits method, which Fireworks supported in its API.
- GitLab's incident record of March 24, 2026 reports a 12.03% code-completion error rate when Fireworks could not scale a new on-demand Codestral 2508 deployment, and GitLab redirected the traffic and moved to reserved deployments.
- Developers report bills of US$1 to US$2 a day on GLM 5.3 Flash (October 9, 2026) and about US$50 a day for 14 million tokens on Kimi K3 (July 29, 2026), and others report spend above their subscription or a 3x to 5x premium over DeepSeek's own API.
- One developer reports Fireworks raising a price by about 30% (September 22, 2026), and another found cache reads on GLM 5.3 Flash cost more than on DeepSeek (September 24, 2026).
What named teams and developers report about running open models on Fireworks AI, from three named deployments and 47 public practitioner accounts that a different model checked against their stored pages. Accounts split on cost and speed, and GitLab's record shows a new deployment that would not scale.
Named deployments
Cursor deployed its fine-tuned 70b fast-apply model with Fireworks in May 2024 and reports about 1000 tokens per second. Sourcegraph ran StarCoder for most Cody community-edition completions on Fireworks in 2023, and its co-founder said that spared the team from building an inference stack.
GitLab redirected Codestral 2508 traffic to its Codestral 2501 deployment on March 24, 2026, after its incident record gave the cause of a 12.03% completion error rate as Fireworks being unable to scale the new deployment. GitLab had run on-demand deployments with legacy direct routing and moved to reserved deployments the same day.
What developers pay
Developers who left a frontier subscription report lower bills: one spends US$1 to US$2 a day on GLM 5.3 Flash, and another went from well over US$100 a month to less than US$50 on Kimi K2.5. Others report the opposite: about US$7 of tokens in one day on DeepSeek V4.1 Flash, and zero-retention hosting at 3x to 5x DeepSeek's own price.
Fireworks lists Serverless prices per 1M tokens for input, cached input and output, and bills batch inference at 50% of Serverless prices. One developer was billed US$192 for an 8-minute supervised fine-tune of gpt-oss-20b on 10k records (September 19, 2025).
Limits and compatibility
Fireworks documents 429 and 503 responses on Serverless, an account-wide limit of up to 6,000 requests per minute, and latency variability during high-volume periods. One developer never hit rate limits on GLM-5.2 (July 27, 2026), while GitLab's new deployment did not scale.
Developers report 400 errors when OpenAI message fields reach Fireworks after several agent steps (August 26, 2025), and when tool fields reach its Anthropic-compatible endpoint (April 27, 2026), which Fireworks documents as serving only POST /v1/messages without server-side tools.
Methodology
GILD built this piece only from rows a different model accepted against their stored pages on November 1, 2026: three deployment accounts, 47 practitioner records from Hacker News, GitHub, Reddit, PeerSpot and GitLab, and Fireworks' own documentation. Each figure is the author's own report, with its date. Cursor's and Sourcegraph's accounts predate 2025, so they appear here and in Operator Insights and not in the Signals count.
Products in this piece
Sources
- Official documentation
- 7
- Public source
- 17
- Fireworks AI: Pricing
- Fireworks AI: Serverless pricing
- Fireworks AI: Batch inference
- Fireworks AI: Serverless rate limits (text)
- Fireworks AI: Account quotas
- Fireworks AI: Serverless latency FAQ
- Fireworks AI: Anthropic compatibility
- Cursor blog: Editing Files at 1000 Tokens per Second (May 14, 2024)
- Latent Space: Beyang Liu of Sourcegraph on running StarCoder on Fireworks (Dec 14, 2023)
- Sourcegraph blog: Cody October 2023 feature release (Oct 2023)
- 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, kelnos: US$1 to US$2 a day on GLM 5.3 Flash (Oct 9, 2026)
- Hacker News, drewnick: a daily workflow on Kimi K3 (Jul 29, 2026)
- Hacker News, vuldin: from Claude Code to Kimi K2.5 on Fireworks (Feb 18, 2026)
- Hacker News, bwfan123: zero-data-retention hosts at 3x to 5x DeepSeek's price (Jul 19, 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, mountainriver: a US$192 bill for an 8-minute fine-tune (Sep 19, 2025)
- Hacker News, switchbak: Fireworks speed against Baseten (Oct 26, 2026)
- Hacker News, verdverm: new models arriving late on Fireworks (Oct 19, 2026)
- GitHub, mozilla-ai/any-agent issue #778: invalid_request_error with Fireworks models (Aug 26, 2025)
- GitHub, earendil-works/pi issue #3854: 400 errors on tool requests through anthropic-messages (Apr 27, 2026)