Open-Weight Hosting Economics: Groq vs. Together AI vs. Amazon Bedrock vs. Mistral

Case Study 1: openai/gpt-oss-120b (Groq vs. Together AI vs. Amazon Bedrock)

All three providers list the exact same headline synchronous price of $0.15/1M input and $0.60/1M output (in US standard regions). Yet real-world monthly spend diverges as soon as you enable caching, Batch, or residency:

• Groq (~500 tok/s LPU, 131K ctx): Offers 50% automatic prompt caching ($0.075 read, 2h inactivity TTL) AND 50% Batch ($0.075/$0.30)—but Batch and caching do not stack. Sibling gpt-oss-20b drops to $0.075/$0.30 ($0.0375 cached/Batch).

• Amazon Bedrock (US Std): Offers 50% Batch AND 50% Flex tier ($0.075/$0.30), plus Priority tier (+75% = $0.2625/$1.05), but zero prompt caching on gpt-oss models (Bedrock OpenAI caching is GPT-5.6+ only). Sydney region edges up slightly to $0.1545/$0.6180.

• Together AI (Serverless MXFP4): Offers zero prompt caching ('-') and 0% Batch discount ($0.15/$0.60 even in Batch). Pinning US-only execution via us-api.together.ai adds a +25% (1.25x) uplift to $0.1875/$0.75.

Case Study 2: DeepSeek-V4.1-Flash (DeepSeek 1P vs. Mistral vs. Together AI)

DeepSeek's own API bills $0.30/$1.20 during Peak UTC hours and drops 50% to $0.15/$0.60 Off-Peak (outside 01:00–04:00 and 06:00–10:00 UTC Mon–Fri, plus all day on weekends/CN holidays), with automatic $0.003/$0.006 disk KV cache reads (98% off), 384K max output, and no Batch API.

If your enterprise requires non-China hosting or true async Batch queues, Mistral hosts deepseek-v4-1-flash-1017 at $0.30/$1.20 flat 24/7 with $0.007 cached reads (64-token blocks) AND 50% Batch ($0.15/$0.60)—matching DeepSeek's Off-Peak rate at any hour of the day. Together AI hosts DeepSeek-V4.1-Flash (FP8) at $0.30/$1.20 with $0.006 best-effort caching (~1h TTL), 0% Batch discount, and a 1.25x US-only uplift ($0.375/$1.50).