Fireworks AI and Together AI are the two hosts most teams shortlist when they want open-model inference that feels production-grade, and in 2026 the honest answer is: prices are close, platforms differ. On the models where we can compare them directly, Fireworks is slightly cheaper on serverless per-token rates (90 cents versus 1.04 dollars per 1M output tokens on Llama 3.3 70B in our latest check), while Together carries a much wider model catalog and a bigger, better-defined startup credit program worth up to 50,000 dollars. Neither is the cheapest host we track for most models, and both know it; you pay a modest premium over the discount tier for reliability, tooling, and fine-tuning support.
All prices below come from our live rankings, which re-check 88 model price tables across every known provider weekly. Prices move monthly, so treat the live tables as the source of truth.
Head-to-head prices, per 1M output tokens
Where both platforms serve the same model in our latest verified check:
| Model | Fireworks | Together AI | Cheapest host we track |
|---|---|---|---|
| Llama 3.3 70B | $0.90 | $1.04 | OpenRouter, $0.32 |
| gpt-oss-120b | $0.60 | $0.60 | DeepInfra, $0.17 |
Two things jump out. First, the gap between Fireworks and Together is small: 13 percent on the 70B row, zero on gpt-oss-120b. Second, the gap between either of them and the discount tier is large: DeepInfra serves gpt-oss-120b at roughly a quarter of what both charge. If raw per-token price is your only criterion, this comparison is moot and you should read the cheapest AI inference APIs instead. This post is for teams who have decided the discount tier is not enough and are choosing between the two premium open-model hosts.
A note on how each prices. Fireworks uses size-tiered serverless pricing: models are grouped into parameter bands and every model in a band costs the same, which makes bills predictable when you swap models. Together prices per model. That is why Together shows up in far more of our ranking tables: it hosts a broader catalog, and each entry gets its own rate.
Together's catalog depth, in numbers
Together appears in our verified tables for far more models than Fireworks. A sample of Together rates from the latest check, per 1M output tokens:
- Llama 3.1 8B: $0.18
- Llama 4 Scout: $0.59
- Qwen3 235B A22B: $0.60
- Llama 4 Maverick: $0.85
- Qwen2.5 72B: $1.20
- DeepSeek-V3: $1.25
- Llama 3.1 405B: $3.50
- Kimi K2: $4.50
- DeepSeek-R1: $7.00
The Llama 3.1 405B row is worth flagging: at 3.50 dollars, Together is the cheapest verified host we track for that model, beating everyone including the discount providers. On most other rows Together sits mid-table: solid, but routinely 30 to 70 percent above the cheapest endpoint. DeepSeek-R1 is the extreme case, where Together's 7.00 dollars is more than triple DeepInfra's 2.15. If a Together model is core to your stack, check its specific row in the rankings before assuming the premium is small.
Where Fireworks quietly wins: audio
Fireworks serves whisper-v3-large transcription at 0.0015 dollars per minute in our latest check, which is the cheapest hosted Whisper rate we track anywhere: a quarter of OpenAI's gpt-4o-transcribe and below Groq. If your product transcribes at any volume, this single row can decide the comparison by itself. Together does not compete seriously in speech-to-text. Full table on the Fireworks Whisper ranking page, with the broader market in cheapest speech-to-text APIs.
Fireworks also leans into speed as a differentiator: its own FireAttention serving stack, aggressive speculative decoding, and image and embedding endpoints alongside LLMs. Together counters with its inference engine plus a GPU cluster business (dedicated instances and reserved capacity) that Fireworks matches only partially.
Fine-tuning: both are real, with different flavors
This is where both platforms genuinely earn their premium over discount hosts, most of which offer weak or no tuning.
- Fireworks supports LoRA fine-tuning with training billed per token (its listed LoRA SFT rates run from 0.50 to 10.00 dollars per 1M training tokens depending on model size), and the key economic feature: serving a fine-tuned LoRA costs the same per token as the base model. You can also hot-swap many LoRA adapters on one deployment.
- Together supports both LoRA and full fine-tuning across a wide slice of its catalog, also billed per training token, with downloadable checkpoints so you own the weights and can leave. That exit path matters: you can tune on Together and later serve the merged model anywhere, including a self-hosted setup.
If you plan to tune many small adapters and serve them cheaply, Fireworks' LoRA economics are hard to beat. If you want full fine-tunes and weight portability, Together is the safer bet.
Startup programs: Together's is bigger and clearer
Together AI runs a startup accelerator with credit tiers set by total funding raised: up to 15,000 dollars if you have raised under 5M, up to 30,000 dollars between 5M and 10M, and up to 50,000 dollars above 10M. Credits apply to inference, fine-tuning, and dedicated endpoints, but not reserved GPU clusters. It is selection-based, so treat it as an application, not a coupon. Full verified details, eligibility, and the application link are in the Together AI startup credits listing in the Perkstack catalog.
Fireworks runs a Fireworks for Startups program offering build credits and higher rate limits, but it does not publish fixed credit tiers the way Together does, so we cannot promise you a number. New Fireworks accounts also get a small starter credit (about a dollar, enough to smoke-test an endpoint, not to build on).
Practical read: if startup credits are part of your runway math, Together's program is the more bankable of the two, and at the top tier it can cover a serious inference bill for months. Stack it with the rest of the ecosystem using our startup credits checklist. Perkstack members get the full picture: the catalog tracks 200 plus hand-verified credit programs worth over 5.5M dollars combined, so you can browse the catalog and claim in priority order instead of discovering programs one blog post at a time.
How to choose
- Choose Fireworks if: you want the lowest premium-tier price on 70B-class serverless, you transcribe audio at volume, you plan to serve many LoRA adapters, or latency is a headline requirement.
- Choose Together if: you need a specific model Fireworks does not host, you want full fine-tuning with portable weights, you need 405B-class serving at the best verified rate, or you can land the accelerator credits.
- Choose neither if: your workload is price-sensitive batch processing on common open models. DeepInfra and Novita undercut both by 2x to 4x on most rows; see DeepInfra vs Together for that matchup.
Switching between them later is close to free. Both expose OpenAI-compatible endpoints, so moving is a base URL and key change. Prototype on whichever grants you credits first, and let a week of real traffic decide.
Bottom line
Fireworks and Together are the two credible premium hosts for open models in 2026, separated less by price than by shape: Fireworks is slightly cheaper where they overlap ($0.90 versus $1.04 on Llama 3.3 70B), owns the cheapest hosted Whisper we track, and has the better LoRA serving economics; Together has the deeper catalog, full fine-tuning with weight portability, the best verified Llama 3.1 405B rate at $3.50, and a startup program worth up to $50,000. Check the current per-model numbers in our live rankings, which are re-verified weekly, and if you qualify for credits, apply before you spend cash. Create an account to track both programs alongside everything else in the catalog.
Related reading: cheapest AI inference APIs, DeepInfra vs Together, LLM API pricing comparison.