Same model, different bill: what open-weight LLMs cost across 6 inference hosts

Published 2026-08-06 · reference numbers, verify before budgeting

DeepSeek V4 is DeepSeek V4. The weights are identical whether you call it through Groq, Together, Fireworks, DeepInfra, Novita or OpenRouter — same architecture, same output quality, no fine-tuning difference between hosts. What's not identical is the bill. We pulled our per-1M-token pricing across all six for the exact same model, twice, and the gap between cheapest and priciest host lands at a consistent 31% — every single time.

The workload we're pricing

A mid-size RAG or coding-assistant backend running DeepSeek V4: 50 million input tokens and 10 million output tokens a month. That's a real production range — a support bot or internal coding tool with a few hundred active users, not a toy project.

Step 1: DeepSeek V4, six hosts

Prices are reference estimates, August 2026. Report outdated price →

$19.60 to $25.73 on identical weights for identical usage — a $6.13/month, 31% gap with zero quality difference to explain it. At 10× the volume (500M in / 100M out, a genuinely busy product) that's $196 vs $257, a $735/year difference from the host dropdown alone.

Step 2: Llama 4 Maverick, same six hosts

To check this wasn't a DeepSeek quirk, we ran the same six hosts against Llama 4 Maverick on the same 50M/10M workload:

Same ranking, same order, same 31% spread from Novita at the bottom to Fireworks at the top. That's not a coincidence — it's a fixed multiplier every host applies to the same reference price, and it holds across every open model we checked in our pricing set, from 12B parameter Gemma to 400B+ parameter Llama 4 Behemoth.

Where the 31% band comes from

• Together and OpenRouter price at exactly 100% of list. Both are aggregator-style platforms — OpenRouter especially is a marketplace, not an infra operator — so neither has a cost structure that lets them undercut the reference rate. You pay the sticker price for the convenience of one API key across dozens of models.
• Novita and DeepInfra run at 80% and 85% of list. Both compete primarily on price for open-weight serving and carry a leaner feature set — fewer regions, thinner dashboards, less enterprise support — than the bigger names. The discount is real, but so is the smaller safety net if something breaks at 2am.
• Groq sits at 95%, a small discount off list rather than a big one — Groq's pitch is inference speed on its own LPU chips, not being the cheapest line item, and the pricing reflects that positioning.
• Fireworks runs 5% over list, the only host of the six pricier than the reference rate, consistent with its enterprise-SLA and dedicated-deployment focus rather than lowest-cost self-serve.

What this doesn't tell you

Price parity across models doesn't mean the hosts are interchangeable. We're comparing per-token rate only — not latency, uptime history, region coverage, rate limits or how each host behaves under a traffic spike. A 20% discount on Novita is worth nothing if a support bot times out during your busiest hour. If you're pre-revenue and cost-sensitive, start with the cheapest host and a fallback. If you're already at meaningful volume, benchmark latency and uptime on your own traffic before switching purely on the price column — the $735/year gap from our example is real money, but it's not the only number that matters.

The takeaway

Open-weight doesn't mean open pricing. The model being free to download doesn't stop six different companies from charging six different rates to run it for you, and the spread — 31% top to bottom, in the same order, on every model we checked — is large enough to matter once you're past a few million tokens a month. Before committing to one host, run your own token volume through the cheapest LLM API tool, and compare the providers directly on their pages: Groq, Together, Fireworks, OpenRouter. It's the same lesson as the frontier tax across model tiers: the model name on the invoice tells you less than the vendor logo next to it.

Figures are reference estimates (August 2026) from our own pricing dataset. Inference host pricing changes frequently and volume discounts, committed-use rates and regional pricing can shift the ranking — always confirm current rates on each provider's official pricing page before budgeting.