Home › Blog › Novita is exactly 20% cheaper than Together on every model we track

Novita Is Exactly 20% Cheaper Than Together on Every Model We Track

16 November 2026 · Site Methodology & Corrections · 5 min read

Open our LLM price comparison table, pick any open-weight model that's hosted on more than one reseller, and sort by price. Novita comes out cheapest. Every time. I wanted to know if that was just a few popular models lining up by chance, so I pulled our pricing database and ran the actual ratios instead of eyeballing a sorted table.

41 models, 6 hosts, zero exceptions

Our database prices 41 open-weight models — Llama, Qwen, Mistral, Gemma, DeepSeek, Mixtral and a dozen smaller ones — each listed across the same six resale platforms: Together, Fireworks, Groq, DeepInfra, Novita and OpenRouter. I took Together's rate as the baseline for every model and computed every other platform's rate as a ratio against it, on input and output tokens separately.

The ratio doesn't drift model to model. It's identical to four decimal places on all 41:

PiattaformaPrice vs TogetherModels matching exactly
Fuochi d'artificio+5.0%41 / 41
OpenRouter0.0% (same as Together)41 / 41
Groq−5.0%41 / 41
DeepInfra−15.0%41 / 41
Novita−20.0%41 / 41

(Kimi K2 is excluded from this set on purpose — we already flagged Novita's listed Kimi K2 rate as stale in last week's audit, and a known-bad data point has no business sitting inside a "zero exceptions" claim.)

What that looks like in real dollars

Ratios are abstract until there's a workload behind them. Say you're running 100M input tokens and 20M output tokens a month — a mid-size production app, not a toy project.

OspiteDeepSeek V4Llama 4 MaverickQwen 3 235B
Fireworks (priciest)$110.88$60.90$41.58
Together / OpenRouter$105.60$58.00$39.60
Groq$100.32$55.10$37.62
DeepInfra$89.76$49.30$33.66
Novita (cheapest)$84.48$46.40$31.68

Novita vs Fireworks on that DeepSeek V4 workload: $84.48 against $110.88. Same model, same weights, same context window, same output — a $26.40/month gap that's purely which vendor you clicked "deploy" on. Run a workload ten times that size and it's $264/month for doing nothing except picking a different host.

Why would a markup be this exact?

I don't have Novita's or Fireworks' internal margin sheet, so I won't pretend to know their reasoning. But a flat percentage applied identically across an entire catalog looks like a deliberate pricing policy, not something that fell out of per-model GPU-cost math. If DeepInfra were actually 15% cheaper to run specifically because of how GLM-4.6 maps to their hardware, there's no reason that same exact 15% would also hold for Falcon 180B, Mixtral 8x7B and Qwen 2.5 7B — three models with very different parameter counts and compute profiles. A flat rake on top of (or under) Together's list price explains the pattern better than a cost-driven one does.

That also makes the pattern fragile in a way real spot pricing isn't. If any one of these six platforms reprices its margin policy, the lockstep breaks immediately and this whole table needs rechecking — which is exactly the kind of change our ongoing audit pass exists to catch, the same process that caught the Kimi K2 price and the Adyen and Binance fee drifts in recent weeks.

What I'd actually do with this

If you're locked into a specific open-weight model and the only decision left is which host serves it, check Novita first — on this dataset it wins outright, with no model where it loses. If you're on Fireworks today for latency, support or integration reasons that have nothing to do with price, that's still a legitimate call — just go in knowing you're paying a flat 25 percentage points more than Novita for the identical weights, not a different product.

Run your own workload through the LLM price comparison tool to see current numbers host by host, or price a specific model directly on Insieme, Fuochi d'artificio, Groq O OpenRouter's own pages. New to comparing inference hosts? Start with the guide gratuite sui costi API.

Ratios computed directly from our pricing database, 16 November 2026, across 41 open-weight models listed on all 6 resale platforms we track. Kimi K2 excluded as a known-stale data point. Reseller margins can change — confirm current pricing on the comparison page before budgeting a production workload.