Model catalogue
41 models from 9 vendors and the open-weight fleet. Prices per 1,000,000 tokens. Rows sorted by weekly token volume.
codingreasoningagents & toolstranslationRP & personaslong contextmultimodalopen weightsself-hostEU residency
| Model | Context | Prompt, $/M | Completion, $/M | Quantization | Tok/s | Tokens this week |
|---|---|---|---|---|---|---|
| gemini-3-flashGooglemultimodal1M ctx | 1M | $0.10 | $0.40 | fp8 | 240 | 83.6B |
| qwen-flash-next-3.8Alibabaopen weightsruns locally | 256k | $0.06 | $0.24 | fp8 / int4 | 168 | 71.3B |
| gpt-5-miniOpenAIvisiontools | 256k | $0.25 | $2.00 | fp8 | 145 | 61.4B |
| deepseek-flash-4.1DeepSeekopen weightsruns locallytools | 164k | $0.08 | $0.32 | fp8 / int4 | 132 | 58.9B |
| gpt-5OpenAIvisionreasoning | 256k | $1.25 | $10.00 | fp8 | 62 | 48.2B |
| gemini-3-proGooglemultimodal1M ctx | 1M | $1.25 | $10.00 | fp8 | 110 | 44.8B |
| qwen3.8-27bAlibabaopen weights24 GB GPUtools | 128k | $0.10 | $0.40 | q4 / int8 | 96 | 42.7B |
| claude-sonnet-4.5Anthropiclong taskstools | 200k | $3.00 | $15.00 | — | 78 | 39.5B |
| deepseek-r2DeepSeekreasoningMoE | 164k | $0.27 | $1.10 | fp8 | 45 | 35.2B |
| kimi-k2.5Moonshotopen weightsagentic | 256k | $0.60 | $2.50 | fp8 | 58 | 24.8B |
| gpt-5-codexOpenAIcodeagentic | 400k | $1.25 | $10.00 | fp8 | 54 | 18.7B |
| llama-4-70bMeta (EU PoP)open weightsEU | 1M | $0.60 | $1.80 | fp8 | 88 | 15.9B |
| claude-opus-4.5Anthropictop qualitytools | 200k | $15.00 | $75.00 | — | 41 | 14.3B |
| glm-4.6Zhipuopen weightsagentic | 200k | $0.30 | $1.20 | fp8 | 71 | 12.6B |
| grok-4.1xAIrealtime search | 512k | $2.00 | $8.00 | fp8 | 66 | 11.2B |
| devstral-3-coderMistralopen weightscodeлокально | 256k | $0.80 | $2.40 | fp8 / int4 | 92 | 9.8B |
| mistral-large-3MistralEUtools | 128k | $2.00 | $6.00 | fp8 | 63 | 8.4B |
| llama-4-405bMetaopen weights1M ctx | 1M | $2.50 | $8.00 | fp8 | 34 | 6.7B |
| mistral-small-4Mistralopen weightsfast | 128k | $0.20 | $0.60 | fp8 | 187 | 5.9B |
| qwen3.8-vl-32bAlibabaopen weightsvisionлокально | 256k | $0.15 | $0.60 | fp8 / int4 | 84 | 5.2B |
Quantization shows the weight format for open-weight models; vendor models do not disclose the exact format (—). Tok/s is the 24-hour median generation speed. Hidden models (21) are available on request for Scale tier.
Where the traffic goes
| Coding & code agents | 34% |
| Support & business agents | 21% |
| Translation & localization | 14% |
| Document summarization | 12% |
| RP & personas | 9% |
| Search & RAG | 6% |
| Other | 4% |
How to read the catalogue
- Open weights — published weights: same API through our fleet, same API on your GPU. qwen-flash-next-3.8, qwen3.8-27b, deepseek-flash-4.1, devstral-3-coder and qwen3.8-vl-32b are mirrored with stable checksums.
- Runs locally — the model fits 24 GB VRAM at 4-bit quantization; run instructions in the docs.
- Route modifiers — the
:fast,:cheap,:euand:exactsuffixes pick a strategy without changing keys:deepseek-flash-4.1:fast. - BYOK — bring your own vendor key on Starter and up: the gateway fee is waived.