Pricing
Pay for tokens. Nothing else.
No subscriptions, no minimums, no per-seat anything. Prices are per million tokens, input and output priced separately — the way every rate card you'd compare us to reads. Batch is half price.
| Class | Input | Output | Batch input | Batch output | Operator share | What lives here |
|---|---|---|---|---|---|---|
| Nano · ≤3.5B | $0.014 | $0.03 | $0.007 | $0.015 | 40% | classification, extraction, routing |
| Small · 3.5–9B | $0.018 | $0.036 | $0.009 | $0.018 | 45% | the volume workhorse |
| Mid · 9–40B | $0.075 | $0.29 | $0.038 | $0.145 | 65% | 24GB+ GPUs and 64GB+ Macs earn here |
| Large · 40–150B | $0.105 | $0.34 | $0.053 | $0.17 | 72% | 96GB+ Apple Silicon; scarce supply |
| XL · >150B | $0.3 | $0.97 | $0.15 | $0.485 | 75% | 256–512GB unified; the rarest supply |
| Embeddings | $0.004 | — | $0.002 | — | 50% | bulk vectorization; input tokens only |
Classes are by total parameters, mixture-of-experts included: you buy the big model's quality, and the operator has to provision the big model's memory.
Decision models (/v1/systemone) have no
row of their own. They are billed on input tokens at their class's input rate — Julia-1 and
Laya at nano ($0.014), Kev-4B and Clef-flash at small
($0.018), Clef at mid ($0.075) — and they
generate no tokens, so the output column never applies. Batch is half, as everywhere.
What they are →
The multipliers, in the open
Latency class
Batch 0.5× · interactive 1.0×. Batch jobs queue for idle capacity — that's why they're half price, not a gimmick — and batch is never surged.
Privacy tier
Open is the listed price; verified and private carry premiums for scarcer, accountable capacity.
Surge
When demand on a model outruns the supply serving it, interactive requests pay up to 1.5×: nothing below 80% utilisation, ramping to 1.5× at 100%. Never silent — the multiplier you paid is on every API response and shown in the playground before you send. Operators earn the surge too, which recruits supply.
Where the money goes: operators keep 40–75% of what customers pay, rising with how scarce their hardware is — 40% on nano, 45% small, 65% mid, 72% large, 75% xl — on input and output tokens alike. The rest funds verification hardware and margin. The split is part of the product, not a secret.
How we compare
Every published open-weight host we could find, the cheapest included, plus the frontier APIs — as their rate cards read on 30 Aug 2026. Prices move; the date is part of the claim. Ours are the table above: interactive output, and batch output, per million tokens.
| Provider | Class | Model | Their input | Their output | Ours, output | Ours, batch | Source |
|---|---|---|---|---|---|---|---|
| DeepInfra | nano | Llama 3.2 3B class | $0.02 | $0.04 | $0.03 | $0.015 | card |
| DeepInfra | small | Llama 3.1 8B | $0.02 | $0.04 | $0.036 | $0.018 | card |
| DeepInfra | mid | Qwen 3.2 32B | $0.08 | $0.28 | $0.29 | $0.145 | card |
| DeepInfra | large | Llama 3.3 70B | $0.1 | $0.32 | $0.34 | $0.17 | card |
| DeepInfra | embedding | bge-base-en-v1.5 | $0.005 | $0.005 | $0.004 | $0.002 | card |
| Novita | small | Llama 3.1 8B Instruct | $0.02 | $0.05 | $0.036 | $0.018 | card |
| Novita | mid | Gemma 4 26B A4B | $0.13 | $0.4 | $0.29 | $0.145 | card |
| Novita | large | Llama 3.3 70B Instruct | $0.135 | $0.4 | $0.34 | $0.17 | card |
| Novita | embedding | BAAI BGE-M3 | $0.01 | $0.01 | $0.004 | $0.002 | card |
| Groq | small | Llama 3.1 8B Instant | $0.05 | $0.08 | $0.036 | $0.018 | aipricing.guru (third-party) |
| Groq | large | GPT-OSS-120B | $0.15 | $0.6 | $0.34 | $0.17 | card |
| Fireworks | small | 4–16B tier | $0.2 | $0.2 | $0.036 | $0.018 | card |
| Fireworks | mid | >16B tier | $0.9 | $0.9 | $0.29 | $0.145 | card |
| Fireworks | large | GPT-OSS-120B | $0.15 | $0.6 | $0.34 | $0.17 | card |
| Fireworks | embedding | ≤150M params | $0.008 | $0.008 | $0.004 | $0.002 | card |
| Together | small | Llama 3 8B Instruct Lite | $0.14 | $0.14 | $0.036 | $0.018 | card |
| Together | mid | Gemma 4 31B | $0.39 | $0.97 | $0.29 | $0.145 | card |
| Together | large | Llama 3.3 70B | $1.04 | $1.04 | $0.34 | $0.17 | card |
| Together | embedding | multilingual-e5-large | $0.02 | $0.02 | $0.004 | $0.002 | card |
| Cerebras | large | GPT-OSS-120B | $0.35 | $0.75 | $0.34 | $0.17 | pricepertoken.com (third-party) |
| Chutes | mid | Qwen3-32B | $0.104 | $0.416 | $0.29 | $0.145 | card |
| Hyperbolic | small | Llama 3.1 8B (BF16) | $0.1 | $0.1 | $0.036 | $0.018 | card |
| OpenAI | small | gpt-5.6-luna, short context | $0.1 | $0.6 | $0.036 | $0.018 | card |
| small | Gemini 2.5 Flash-Lite | $0.1 | $0.4 | $0.036 | $0.018 | card | |
| Anthropic | small | Claude Haiku 4.5 | $1 | $5 | $0.036 | $0.018 | card |
| OpenAI | embedding | text-embedding-3-small | $0.02 | $0.02 | $0.004 | $0.002 | card |
Read it straight: on the 8B class we're 10% under the cheapest host and 4× under Together, 11–17× under the OpenAI and Google entry tiers; on mid and large we sit within a few percent of the cheapest host at the interactive rate and about half of it on batch. Your embedding job runs at $0.004 per million tokens, $0.002 in batch — 5× under Together and under the cheapest host we found — and the money goes to the people whose machines run it.