A Shanghai lab just released a model that costs a tenth of Claude Opus, claims to outperform GPT-5.5 on coding benchmarks, holds a million tokens of context, reads images and video natively, and will - within days - let you download the weights and run the whole thing on your own hardware.
That is the pitch. Here is what your team actually needs to decide.
MiniMax launched M3 on June 1, 2026, positioning it as the first open-weight system to combine frontier coding-agent performance, a one-million-token context window, and native multimodal capabilities - including image, video, and desktop computer operation - in a single model. The architecture behind the context window is genuinely interesting: MiniMax Sparse Attention uses a lightweight index branch to scan incoming tokens and select which blocks of past tokens require attention, then runs attention only on those relevant blocks. At 1-million-token context length, this reduces per-token compute to one-twentieth of the prior generation, with more than 9x faster prefill and more than 15x faster decoding.
The cost gap is real. MiniMax M3 API input pricing starts at approximately $0.30 per million tokens during the launch promotion, while Claude Opus 4.7 charges $5.00 per million input tokens and $25.00 per million output tokens. For teams running agent loops where context gets resent repeatedly, that difference compounds fast.
The price is not a rounding error. It is an order-of-magnitude gap that changes what you can afford to build.
So why not just use it?
The benchmarks are vendor-run. Several benchmark results were obtained on MiniMax's own infrastructure using agent scaffolding such as Claude Code and Mini-SWE-Agent. Independent third-party verification is still pending, and M3 has not yet appeared on the DeepSWE board for long-horizon software tasks. That is not unique to MiniMax - most labs run their own evals at launch - but it matters more here because the numbers are the entire argument for switching.
The open weights had not shipped at launch. Developers cannot yet inspect the architecture details, verify the training setup, assess the safety behavior under edge cases, or confirm the licensing terms. Until the weights ship and independent engineers can reproduce the architecture claims, M3's open-weight designation is a company commitment - not a verifiable fact. MiniMax put the target at approximately June 11. If you are reading this after that date, check Hugging Face directly - the model card is the source of truth. MiniMax's previous M2.7 license restricted commercial use without prior written authorization. If M3 follows that precedent, expect downloadable weights with a non-commercial default and enterprise licensing through direct sales.
The API route carries a jurisdiction risk you need to decide on deliberately. For workloads involving proprietary source code, customer data, or confidential business information, any prompt processed through MiniMax's API falls under the jurisdiction of China's 2017 National Intelligence Law, which requires MiniMax to cooperate with Chinese government intelligence requests. That obligation is not altered by server location, Western subsidiaries, or MiniMax's stated privacy policy.
No confirmed backdoor in M3 specifically has been found. The National Intelligence Law obligation, however, is structural and legally confirmed - it does not require a demonstrated breach to create risk.
This is the same question teams face with DeepSeek, Qwen, and other Chinese-origin models. A constellation of Chinese AI labs - DeepSeek, Moonshot, and MiniMax chief among them - are increasingly popular fixtures of a global open-source marketplace. Though they are open-weight and free to use, Chinese-origin AI models are still developed by companies subject to the country's National Intelligence Law and liable to "support, assist, and cooperate" with the Chinese government's national security investigations.
The answer is not automatically no. It depends on what you are sending.
Here is the practical breakdown:
Use the API today if your workloads involve non-sensitive data, you want to test the long-context performance claims against your own tasks, and you can run the evaluation before putting anything real through it. If you are calling the hosted API, you need to make a deliberate decision, not a default one.
Wait for the weights if your plan is self-hosting. If you self-host the open weights on your own infrastructure, the concern changes substantially. If you use the hosted API, decide explicitly based on your data sensitivity and compliance requirements. Once the weights land and the license is known, running M3 on your own hardware means data never leaves your network - which is exactly the posture that makes the price advantage worth pursuing for sensitive use cases.
Skip it for now if your team handles health data, financial records, or anything under a compliance framework that requires you to control where data is processed. The price advantage does not override a regulatory obligation.
The model's internal demos are worth paying attention to. In one test, the model worked for nearly twelve hours without intervention, produced 18 commits and 23 figures, and confirmed the key findings of an ICLR 2025 paper it was asked to reproduce. M3 independently reproduced that paper, achieving a score of 0.650. In another, M3 was asked to optimize a compute kernel for matrix multiplications on Nvidia Hopper GPUs. Experienced teams typically need one to two weeks for this, according to MiniMax. M3 got only a task description, a benchmark script, and a non-functional code skeleton. After about 24 hours, the model had pushed Hopper hardware utilization from 7.6 to 71.3 percent. These are MiniMax's own numbers, and they warrant independent scrutiny - but they point at what a million-token context combined with sustained agentic iteration actually looks like in practice.
The question is not whether M3 is interesting. It clearly is. The question is whether the decision to use it was made deliberately or by default - which is true of any model, but matters more when the tradeoffs include legal jurisdiction over your data.
A teammate like Beagle, which lives inside Slack and Teams, is in a similar position: the data it touches is organizational, and the model routing underneath should reflect a conscious choice, not whatever was cheapest at signup.