MiniMax-M3 raises the pressure on costly frontier APIs

MiniMax-M3 is another sign that the open-weight and low-cost model race is not slowing down. VentureBeat reported that the Chinese AI startup released M3 with a 1-million-token context window, native multimodality, and pricing of $0.30 per million input tokens and $1.20 per million output tokens during an initial discount window. The company says M3 uses MiniMax Sparse Attention to reduce long-context compute, with per-token compute at maximum context falling to 1/20th of the previous generation, plus 9x faster prefilling and 15x faster decoding. Vendor-run benchmarks put M3 at 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, 74.2% on MCP Atlas, and 83.5 on BrowseComp. Treat the claims carefully until open weights and independent evaluations arrive, but the pricing and context-window targets are exactly the pressure enterprise buyers are putting on closed frontier APIs.

Key details: June 1, 2026, MiniMax-M3, 1M-token context window, Native multimodality, $0.30 per 1M input tokens during discount, $1.20 per 1M output tokens during discount, 59.0% SWE-Bench Pro claimed, Open weights promised within 10 days.

Continue swiping for more AI Brief stories.