7 September 2026 · AI & LLMs · 5 min read
I was pricing out a fast-inference provider for an agent workload running a 70B-class open model, and the three names that keep coming up — Groq, Together AI, Fireworks — all quote per-token pricing that looks similar at a glance. It isn't. On Llama-70B-class models, Groq works out to $0.59 input / $0.79 output per 1M tokens, Together AI is a flat $0.88 / $0.88, and Fireworks bills a blended ~$0.90 per 1M for its 16–80B tier. At a realistic 70/30 input/output split, that's a $250-a-month gap between Groq and Fireworks once you're pushing a billion tokens — and the cheapest one isn't automatically the right pick.
All three run open-weight models — Llama, Qwen, Mixtral, DeepSeek — behind OpenAI-compatible APIs, so switching providers is usually a base-URL change. Here's what Llama-70B-class inference actually costs on each, pulled from their own pricing pages:
| Provider | Model | Input $/1M | Output $/1M |
|---|---|---|---|
| Groq | Llama 3.3 70B | $0.59 | $0.79 |
| Together AI | Llama 3.1 70B | $0.88 | $0.88 |
| Fireworks | Mid tier (16–80B, e.g. Llama 70B) | ~$0.90 blended | |
Groq is the cheapest on paper for a typical chat-shaped workload — more input tokens than output — because its input price is 33% below Together's flat rate. Together AI prices input and output identically, which is simpler to forecast but costs more on the input-heavy prompts most RAG and agent workloads actually send. Fireworks doesn't split input/output at all; its ~$0.90 is a single blended rate across the whole 16–80B model class, which makes it easy to estimate but means you can't optimize by shrinking output length the way you can on Groq or Together.
Groq's lower price comes from the same place its speed does: custom LPU (Language Processing Unit) hardware built specifically for token generation, instead of general-purpose GPUs. That's also Groq's real selling point over the other two — hundreds of tokens per second, well past what H100-based serving typically delivers. For a voice agent or a chat UI where the user is watching the cursor, that latency difference matters more than the $0.29/1M gap on input tokens. The catch: Groq's hosted model catalogue is narrower than Together's or Fireworks's, and it doesn't offer dedicated capacity or fine-tuning — if the model you need isn't on Groq's list, price stops being the deciding factor.
Together AI carries the broadest open-model catalogue of the three (Llama, Mixtral, Qwen, DeepSeek and more) plus fine-tuning and dedicated endpoints, so if the job is "run my own fine-tuned variant" rather than "call a stock model," Together's flat, easy-to-forecast pricing and tooling win over Groq's narrower, speed-first lineup. Fireworks sits in a similar spot — serverless per-token pricing for quick starts, but also per-GPU-hour dedicated deployment for teams that want reserved throughput and predictable latency at scale, which neither Groq's fixed catalogue nor Together's serverless-first setup covers the same way. Both are the better default once the workload needs more than "cheapest tokens for a supported model."
Assuming a realistic 70% input / 30% output split, typical of chat and RAG-style prompts:
| Tokens / month | Groq | Together AI | Fireworks |
|---|---|---|---|
| 10 million (small app) | $6.50 | $8.80 | $9.00 |
| 100 million (steady product) | $65 | $88 | $90 |
| 1 billion (high-throughput agent) | $650 | $880 | $900 |
The gap scales linearly with volume since all three are pure per-token pricing with no plan floors or tiers on the model itself — at 1 billion tokens a month, Groq's $650 vs Fireworks's $900 is a real $250 difference, not a rounding error. But none of these numbers account for what happens when the model you want isn't on the cheaper provider, or when Groq's speed cuts your effective cost elsewhere — fewer retries, shorter user sessions, less need to cache aggressively just to hide latency.
Latency-sensitive product (voice, live chat, anything a user is staring at while it streams): start with Groq if your model is on its list — the speed is the actual product decision, and it happens to be cheaper too. Need a fine-tuned or less common open model, or want dedicated capacity with predictable throughput: Together AI or Fireworks, and pick based on whether you value Together's broader catalogue or Fireworks's GPU-hour dedicated option more. Running all three side by side isn't unreasonable either — the OpenAI-compatible API means the code doesn't change, only the base URL and the bill.
New to usage-based AI pricing? Start with the free API-cost guides. For the closed-model side of the same decision, see GPT vs Claude vs Gemini: what the same workload actually costs.
Rates verified against Groq, Together AI and Fireworks's official pricing pages, last checked 2026-08-15. Reference estimates — model availability, volume discounts and negotiated enterprise rates vary; confirm current pricing before budgeting.