We ran the same 100-prompt benchmark — code review, summarization, classification — against Together, Fireworks and Groq. Same model (Llama 3.3 70B Instruct, same quantization), same prompts, same evaluation criteria.
Tokens per second: Groq ~280, Fireworks ~95, Together ~120. Groq's hardware advantage is real and consistent.
Time to first token: Groq 180ms, Together 340ms, Fireworks 410ms. Groq again.
Quality: identical — same weights, same sampling. No surprises.
Cost per 1M tokens: Together $0.88, Fireworks $0.90, Groq $0.65. Groq wins on price too.
Uptime & reliability: Together has the longest track record — four nines. Fireworks good. Groq newer, still proving itself.
Verdict: for interactive UX, Groq wins. For cost at scale with batch workloads, Fireworks wins. For reliability and the deepest model catalog, Together wins.
The real answer: they're all good. Pick by your bottleneck — speed, cost, or breadth.