MiniMax M2.5

MiniMax M2.5 is an open-weights coding model from MiniMax, suited to teams that want cheap agent runs and can accept that it has been superseded. OpenRouter listed it on 12 February 2026 with a 200,000 token context (204,800 when we first checked), text in and text out, at $0.27 per million input tokens and $0.95 per million output tokens on its cheapest route, as listed on 5 November 2026. MiniMax's own Hugging Face card gives a different price: it says the Lightning variant costs $0.3 per million input tokens and $2.4 per million output tokens, and the standard variant half that, so check which route you are paying for. The weights are on Hugging Face, 229B parameters in total by the file metadata, under a modified MIT licence. The card does not summarise that licence's terms, so read the licence file before you ship a product on it. MiniMax reports 80.2 percent on SWE-Bench Verified, though the card's own metadata says 75.8, so treat the score as company-reported. MiniMax's current pay-as-you-go page now lists M3 and M3.1 Flash, not M2.5. We have not run it ourselves. We would only use it where you already depend on it, and start new work on M3. Figures come from OpenRouter's model list and the Hugging Face card.