DeepInfra pricing page (per-model per-1M-token table), re-read 2026-10-17

Fetched 2026-10-17. DeepSeek rows unchanged from the 2026-10-14 capture: DeepSeek-V3.1 $0.25 in / $0.13 cached / $0.95 out (160k); DeepSeek-V3.2 $0.26/$0.13/$0.38 (160k); DeepSeek-V3 $0.32/$0.89 (160k); DeepSeek-R1-0528 $0.50/$0.35/$2.15 (160k); DeepSeek-V4-Flash $0.09/$0.018/$0.18 (1024k); DeepSeek-V4-Flash-0731 $0.06/$0.015/$0.18 (1024k); DeepSeek-V4-Pro $1.30/$0.10/$2.60 (1024k). Qwen rows also held: Qwen3-32B $0.08/$0.28 (40k); Qwen3-30B-A3B $0.12/$0.50 (40k); Qwen3-235B-A22B-Instruct-2507 $0.09/$0.55 (256k); Qwen3-Next-80B-A3B-Instruct $0.09/$1.10 (256k); Qwen3.5-35B-A3B $0.14/$0.05 cached/$1.00 (256k). Kimi-K2.6 $0.75/$0.15/$3.50 (256k); Kimi-K3 $2.85/$0.285/$14.25 (1024k). Llama 3 section: Llama-3.3-70B-Instruct-Turbo $0.10/$0.32 (128k printed); Meta-Llama-3.1-70B-Instruct-Turbo $0.40/$0.40 (128k); Meta-Llama-3.1-8B-Instruct-Turbo $0.02/$0.04 (128k). Page still states all models run on H100 or A100 and the account cap is 200 concurrent requests. The 128k context printed on the Llama 3.3 pricing row and the 131072 context_length on the models API are the same window rounded differently; the API integer is the one stored on the service record.