All three run open-weight models behind an [OpenAI](/entry/notebooklm-vs-chatgpt-projects-vs-claude-projects-which-for-research-2026)-compatible API so you can host Llama, [DeepSeek](/entry/deepseek-vs-chatgpt-vs-claude-is-the-cheap-option-good-enough-2026), Qwen and friends without buying GPUs. On price they're close; the real differences are speed and what else you can do on the platform.
Rough rates for a mid-size model (Llama 3.1 70B class), per million tokens:
| Input | Output | Angle | |
|---|---|---|---|
| Groq | ~$0.59 | ~$0.79 | fastest tokens/sec by a wide margin (custom LPU hardware) |
| Together AI | ~$0.88 | ~$0.88 | dedicated GPUs and raw HGX clusters, cheapest for reserved capacity |
| Fireworks AI | ~$0.90 | ~$0.90 | strong on newest models, good fine-tuning and function-calling |
Groq is the speed play. Its custom LPU chips push tokens dramatically faster than GPU-based providers — if your product's UX depends on near-instant responses (voice agents, live assistants), Groq is worth the integration on that alone. Its per-token price on legacy models is also the lowest of the three.
Together AI is the one to pick if you'll want dedicated capacity: reserved H100s run about $6.49/hour and it offers full HGX clusters, which Fireworks doesn't match at comparable rates. Good for teams that outgrow serverless and want predictable reserved spend.
Fireworks tends to have the newest models available first and its function-calling and structured-output support is strong — the better pick if you're building agents that need reliable tool use, or if you'll be fine-tuning.
Latency-critical product: Groq. Planning to reserve GPU capacity: Together. Agentic workloads or you want the freshest models: Fireworks. For casual serverless use, benchmark all three against your actual prompts — the per-model pricing varies enough that the winner changes by workload.