GILD verdict
Pick Together AI when dedicated GPU hours dominate the bill and an admin will turn prompt storage off, since its on-demand H100 rate is lower.
Pick Fireworks AI when prompts must not be stored by default or dedicated traffic is spiky enough to favor per-second over per-minute billing. On gpt-oss-120b the per-token list prices match, so compare throughput, limits and cached-input pricing on your own workload.
Together AI and Fireworks AI both host open models on serverless per-token APIs and on dedicated GPUs, and both list gpt-oss-120b at US$0.15 per 1M input tokens and US$0.60 per 1M output tokens. They differ on what happens to prompts by default and on how dedicated GPUs are billed. Together stores prompts and responses until an organization admin turns storage off, while Fireworks states Zero Data Retention by default. Together lists an on-demand H100 at US$5.49 per GPU per hour for dedicated inference, and Fireworks bills on-demand deployments per GPU second, with an H100 80 GB at US$8.00 per hour.
Choosing between Together AI and Fireworks AI
Choose Together AI if
- You will run a model on its own GPUs for many hours, where Together's listed US$5.49 per H100 hour on demand is below Fireworks' US$8.00.
- An admin can turn prompt storage off before production traffic, or stored prompts are acceptable for your data.
- You batch work on the models Together's batch page lists at 50% off, and accept standard rates on the rest.
Choose Fireworks AI if
- Prompts must not be logged unless you opt in, since Fireworks states Zero Data Retention by default for open models.
- Dedicated traffic is spiky, where Fireworks bills on-demand deployments per GPU second with no start-up charge, against Together's per-minute billing per replica.
- Prompt caching matters, since Fireworks lists a cached-input price of US$0.015 per 1M tokens for gpt-oss-120b.
Side by side
| Dimension | Together AI | Fireworks AI |
|---|---|---|
| Prompt and response storage by defaultFireworks' data-handling page also says its Response API keeps conversation data for up to 30 days when store=True, which is the default. | Stores prompts and responses by default and may use them for product improvements. Organization admins can turn storage off, which enables zero data retention1 | Zero Data Retention by default for open models: prompt and generation data are not logged without explicit opt-in6 |
| Batch discount | Up to 50% lower cost. Selected serverless models run at 50% off, and the batch page lists one, openai/whisper-large-v3. Models not listed run at standard rates2 | 50% off serverless per-token prices for batch jobs. Not all models support the Batch API7 |
| Dedicated GPU price (H100, on demand)List prices in US dollars on each vendor's pricing page, Nov 1, 2026. | Dedicated inference: US$5.49 per H100 GPU per hour on demand3 | On-demand deployments billed per GPU second with no extra charge for start-up time: H100 80 GB at US$8.00 per hour, or US$0.134 per minute8 |
| gpt-oss-120b serverless list priceList prices in US dollars on Together's model page and Fireworks' serverless pricing page, Nov 1, 2026. | US$0.15 per 1M input tokens and US$0.60 per 1M output tokens4 | Standard tier US$0.15 input, US$0.015 cached input and US$0.60 output per 1M tokens9 |
What practitioners report, side by side
Each product's figures come from its own practitioners, for the group you pick.
Overall signal
All practitioners rate Fireworks AI −17, 32 points above Together AI at −49. Most of the gap is in complaints, with 26 of Together AI's 40 sources negative against 20 of Fireworks AI's 47.
The score counts the authors from all practitioners that take a side, one vote each with their latest view, and is how far positive votes lead negative ones, per 100 votes, so Together AI's −49 rests on votes split 5 positive, 8 mixed and 22 negative and Fireworks AI's −17 on 11, 13 and 18.
A lead of a fifth of the votes reads positive and a lead of half strong positive, with the same lines below zero. The dots show each source, positive, mixed, negative. Run a pilot on your own prompts at expected concurrency to measure time to first token, output speed, error rate and cost per million tokens.
Footprint on the AI Application Platform stack
They differ on two of 9 layers. On Model gateway, Together AI touches it and Fireworks AI supports it. On Retrieval, Together AI touches it and Fireworks AI supports it. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.
| Layer | Together AI | Fireworks AI |
|---|---|---|
| Model | ||
| Model gateway | Touches | Supports |
| Data | ||
| Retrieval | Touches | Supports |
The 7 layers where they match
| Layer | Together AI | Fireworks AI |
|---|---|---|
| Model | ||
| Model APIs | Supports | Supports |
| Build | ||
| Orchestration | Touches | Touches |
| Tool use | Touches | Touches |
| Evaluate | ||
| Evaluation | Touches | Touches |
| Operate | ||
| Observability | Touches | Touches |
| Cloud deployment | Core | Core |
| Guardrails | Touches | Touches |
What practitioners praise
Among all practitioners, Fireworks AI alone has three praise themes that reach 5 sources from 3 authors. Low Per-Token Bills and Serving Speed lead, at 8 of 10 and 8 of 12.
| Praised for | Together AI | Fireworks AI |
|---|---|---|
| Low Per-Token Bills | Not a theme in its sources | 8 of 10 |
| Serving Speed | Not a theme in its sources | 8 of 12 |
| Moving Work to Open Models | Not a theme in its sources | 5 of 7 |
Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.
A count ending in "so far" has under 5 sources from 3 authors among all practitioners. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.
What practitioners raise as concerns
Among all practitioners, Together AI alone has three concern themes that reach 5 sources from 3 authors. Tool Calling & API Compatibility and Python SDK & Integration Packages lead, at 9 of 9 and 5 of 5. Fireworks AI alone has three concern themes that reach 5 sources from 3 authors. Price Rises and Cached Tokens and API Compatibility Gaps lead, at 9 of 12 and 6 of 6.
| Raised as a concern | Together AI | Fireworks AI |
|---|---|---|
| Tool Calling & API Compatibility | 9 of 9 | Not a theme in its sources |
| Python SDK & Integration Packages | 5 of 5 | Not a theme in its sources |
| Errors, Rate Limits & Outages | 5 of 6 | Not a theme in its sources |
| Price Rises and Cached Tokens | Not a theme in its sources | 9 of 12 |
| API Compatibility Gaps | Not a theme in its sources | 6 of 6 |
| Varies by Model and Day | Not a theme in its sources | 4 of 5 |