MiniMax M3: The Open-Source Model at $0.30 Per Million Tokens That Just Beat GPT-5.5
MiniMax M3 dropped on June 1, 2026, and the timing could not have been better positioned. Claude Fable 5 was shut down on June 12. Claude Code Max is $200 per month. Copilot Credits are accumulating on enterprise cards. Meanwhile M3 launched with a 1 million token context window, 59 percent on SWE-Bench Pro — beating GPT-5.5 on multiple benchmarks — at $0.30 per million tokens. Open weights releasing within ten days of launch. This is the cost escape hatch that cost-sensitive teams need to know about.
What MiniMax M3 Is
MiniMax M3 is a Chinese open-source large language model developed by MiniMax, a Shanghai-based AI research company. M3 is their third-generation model, built on a novel architecture they call MiniMax Sparse Attention (MSA) — designed specifically to handle very long contexts efficiently without the quadratic compute scaling that cripples most transformer models at 1 million tokens.
Key specifications:
- 1 million token context window — genuine, not theoretical. MSA makes this economically viable.
- 59 percent on SWE-Bench Pro Verified — the most demanding version of the coding benchmark
- $0.30 per million tokens input pricing through the MiniMax API
- Open weights — self-hostable, no API dependency required
- OpenAI-compatible API — drop-in compatible with any application built on the OpenAI SDK
The Benchmark Numbers in Context
The 59 percent SWE-Bench Pro score requires context to interpret correctly:
- GPT-5.5 scores approximately 54-56 percent on SWE-Bench Pro — M3 beats it by a meaningful margin on this benchmark
- Claude Opus 4.8 and MAI-Thinking-1 score higher — M3 is not the top-tier frontier model, but it outperforms GPT-5.5 at a fraction of the price
- Gemini 3.1 Pro and Claude Sonnet are close to M3's range — the competitive tier is accurate
The honest caveat: MiniMax published these benchmarks themselves. Independent third-party verification is limited at time of writing. The numbers are plausible given the architecture and training approach, but treat them as directional until independent replication is available.
The Pricing Math: 18x Cheaper Than Claude Opus
Claude Opus 4.8 costs approximately $15 per million input tokens. MiniMax M3 costs $0.30 per million input tokens through the MiniMax API. That is a factor of 50 difference at list price.
The 18x comparison referenced in the video is based on Claude Opus at its most common enterprise contract pricing, not list price. Either way: the order-of-magnitude cost difference is real and significant for high-volume workloads. For applications processing millions of tokens daily, the difference between $0.30 and $15 per million tokens is the difference between economically viable and prohibitively expensive.
MSA Architecture: Why 1 Million Tokens Is Feasible
Standard transformer attention scales quadratically with context length — doubling the context quadruples the compute. This is why most models with large context windows are slow or expensive at long contexts even when they technically support them.
MiniMax Sparse Attention addresses this by computing attention sparsely — not every token attends to every other token, but only to a structured subset based on the content relationships. The result is near-linear scaling with context length, making 1 million token contexts economically viable at $0.30 per million tokens rather than prohibitively expensive.
Open Weights: Why This Matters Post-Fable-5
The Fable 5 shutdown demonstrated concretely what "model sovereignty" means in practice: when a model is API-only and government-controlled distribution is imposed, your access disappears without warning. Open weights models are immune to this. You download the weights, you run them on your own infrastructure, and no government directive against the provider affects your access.
This is not a theoretical concern for teams with international operations or government-sensitive workloads. M3's open weights release is a sovereignty option, not just a cost option.
Three Honest Catches
- Self-reported benchmarks: MiniMax published the benchmark numbers. Independent verification at this level of scrutiny is limited. Validate on your actual use case, not the benchmark table.
- Chinese model with data handling implications: For organisations with data residency requirements, legal restrictions on Chinese-origin technology, or US government contract obligations, M3's origin requires compliance review before deployment. The open weights let you self-host, which resolves many (not all) of these concerns.
- No Claude Code integration yet: Claude Code currently cannot use M3 as its underlying model. If Claude Code is central to your workflow, M3 is an API alternative, not a direct replacement.
Key Takeaways
- MiniMax M3 scores 59 percent on SWE-Bench Pro — beating GPT-5.5 — at $0.30 per million tokens with open weights
- The MSA architecture enables genuine 1 million token context without quadratic cost scaling
- Open weights provide model sovereignty that API-only models cannot — relevant after the Fable 5 shutdown demonstrated API dependency risk
- Validate on your actual workload before switching — benchmarks are directional, not deterministic for production performance
- Run compliance review before deploying Chinese-origin models for workloads with data residency, government contracts, or export-sensitive content
