MiniMax-M3 debuts, eclipsing GPT-5.5 and Gemini 3.1 Pro on key benchmark performance for just 5-10% of the cost
MiniMax says its new M3 model matches or beats GPT-5.5 and Gemini 3.1 Pro on some benchmarks while costing far less.
Intelligence analysis by GPT-5.4 Mini
MiniMax launched M3 as a frontier-style model with a 1-million-token context window, native multimodality, and aggressive pricing. The company says it pairs strong agentic and coding results with open-weights plans and a much lower cost than major U.S. rivals.
MiniMax made a new AI model called M3. It is meant to read very long documents, look at images, and help with coding and computer tasks, while costing much less than some big-name AI models.
Think of it like a very fast librarian who can search a giant room full of books without checking every single shelf each time. The company says this makes the model cheaper and faster when the text gets really long.
MiniMax also says M3 did very well on some tests and may be released in a way that lets companies download and change it. That could matter because smaller companies might get powerful AI without paying as much.
Analysis
What MiniMax launched
MiniMax introduced M3 as a new large language model aimed at enterprise AI and agentic workflows. The company says it combines a 1-million-token context window, native multimodality, and strong coding performance while keeping prices far below major proprietary rivals.
Pricing and access
For a limited period, MiniMax is offering API pricing of $0.30 per million input tokens and $1.20 per million output tokens on fresh cache. The article says the full price is still only a fraction of leading U.S. models, and that the model starts at $20 per month under new subscription token plans. MiniMax also says it plans to release the model with an open source license and open weights within about 10 days, which would allow enterprise downloading and customization.
Why the architecture matters
The piece credits the model’s efficiency to MiniMax Sparse Attention, a sparse attention design meant to avoid the quadratic cost growth of standard Transformer attention on long inputs. The article says this approach improves hardware use, reduces compute per token at very long context lengths, and outperforms some open-source sparse attention alternatives in internal tests.
Benchmark claims
MiniMax says M3 scores 59.0% on SWE-Bench Pro, 66.0% on Terminal Bench 2.1, 74.2% on MCP Atlas, and 83.5 on BrowseComp. VentureBeat says those results place it ahead of GPT-5.5 and Gemini 3.1 Pro on selected benchmarks, especially for autonomous agent and browsing tasks. The article also notes a tradeoff: Claude Opus 4.8 still leads M3 on some code-focused and terminal benchmarks.
Bottom line
The story frames M3 as an attempt to collapse the usual split between expensive closed models and cheaper open ones. If the pricing and open-weights plan hold, the model could change how startups think about cost, context length, and deployment control.
Key points
- MiniMax launched M3 with a 1-million-token context window and native multimodality.
- The company says the model is much cheaper than leading proprietary U.S. AI models.
- MiniMax claims M3 beats GPT-5.5 and Gemini 3.1 Pro on selected benchmarks.
- The company plans to release open weights and an open source license soon.
- The article says M3 still trails Claude Opus 4.8 on some code and terminal benchmarks.
If MiniMax’s pricing holds, startups could run long-context and agentic AI workflows at a much lower cost than with top-tier proprietary models. The planned open-weights release could also make it easier for companies to customize the model and keep more control over deployment.
The article also shows M3 is not the clear winner across every benchmark, with Claude Opus 4.8 still ahead in some code and terminal tasks. The big claims on price and openness will matter less if real-world performance, reliability, or release timing fall short of what MiniMax is promising.
