MiniMax M2.5 Review: Frontier AI at 1/10th the Cost
MiniMax M2.5 delivers frontier-class benchmarks, 80.2% SWE-Bench, 84.8% GPQA, at $0.30/$1.20 per million tokens. Full review with real data.
Chinese lab MiniMax released M2.5 on February 12, 2026, and within seven days it had become the single most-used model on OpenRouter, racking up 2.45 trillion tokens in one week, a 197% week-over-week increase (Dataconomy). The reason is straightforward: it posts benchmark numbers that compete with Claude Sonnet 4.6 and GPT-5, but costs roughly 6–10x less.
This review covers what M2.5 is, what the benchmarks actually mean, where it falls short, and when it makes sense to use it.
What is MiniMax M2.5?
MiniMax M2.5 is a Mixture-of-Experts (MoE) language model with 230 billion total parameters and 10 billion active per inference step. Because only a fraction of the network activates for any given token, MoE models can reach large-model quality while keeping compute, and therefore cost, closer to a 10B dense model (MiniMax official announcement).
Key specs:
- Parameters: 230B total, 10B active
- Context window: 205,000 tokens
- Release date: February 12, 2026
- License: Modified MIT (open weights, commercial use with attribution required)
- Weights: Available on Hugging Face
The 205K context window puts it ahead of many competitors in practical long-document workflows, legal review, full codebase ingestion, book-length translation, without hitting the per-token cost wall that makes GPT-5 Turbo impractical at scale.
Benchmark performance
MiniMax published results across several standard evaluation suites. All figures below are from the official release and cross-referenced against the Artificial Analysis intelligence leaderboard (Artificial Analysis).
Coding and software engineering
SWE-Bench Verified tests a model's ability to resolve real GitHub issues from open-source repositories. M2.5 scores 80.2%, placing it above Claude Sonnet 4.6 (which scores in the mid-70s range on the same benchmark) and making it one of the highest-scoring open-weights models available. Multi-SWE-Bench, a multilingual variant, comes in at 51.3%.
Science reasoning
GPQA Diamond is a set of graduate-level biology, chemistry, and physics questions designed to be hard enough that human domain experts score around 65%. M2.5 scores 84.8%, firmly in the frontier tier alongside leading proprietary models.
General knowledge
MMLU Pro, a harder variant of the standard MMLU benchmark covering 57 academic subjects, comes in at 82%. Competitive but not exceptional, several models score in the 85–88% range.
Web browsing and research
BrowseComp (with context management) reaches 76.3%, a strong result for agentic web-search tasks. This matters for developer use cases involving automated research pipelines.
AI capability index
The Artificial Analysis Intelligence Index assigns M2.5 a score of 42, ranking it approximately #5 overall across all models, open and proprietary, at the time of release.
The overall picture: M2.5 is a genuine frontier-class model for coding and science reasoning. It is not a compromise pick for developers who cannot afford Claude Opus.
Speed and efficiency
Two variants exist:
- M2.5 Standard: approximately 56.6 tokens per second
- M2.5 Lightning: approximately 100 tokens per second
Lightning is a distilled, faster variant. At 100 t/s, running M2.5 Lightning continuously for a full hour costs roughly $2.40 in output tokens at standard pricing, a figure MiniMax uses to argue "intelligence too cheap to meter" (The Decoder).
For real-time agentic loops, where a model must call tools, evaluate results, and decide next steps in tight cycles, Lightning's latency advantage over slower frontier models is meaningful. The difference between 30 t/s and 100 t/s compounds across multi-step agent pipelines.
Pricing compared
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| MiniMax M2.5 Standard | $0.30 | $1.20 |
| MiniMax M2.5 Lightning | $0.30 | $2.40 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
| GPT-5 (standard) | ~$2.50+ | ~$10.00+ |
| DeepSeek V3.2 | $0.28 | $0.42 |
For output-heavy workloads, code generation, long-form drafting, agentic tasks where the model generates thousands of tokens per run, the gap between M2.5 and Claude Sonnet is a 10–12x cost reduction. For input-heavy workloads like document analysis with long context, DeepSeek V3.2 is cheaper still, but M2.5 leads on SWE-Bench by a significant margin.
The modified MIT license allows commercial deployment of the weights with attribution, making it viable for teams that want to self-host on their own infrastructure and avoid API costs entirely.
What MiniMax M2.5 is good for
Software engineering and code generation
An 80.2% SWE-Bench score is the clearest signal here. M2.5 is well-calibrated for resolving real issues, not just writing syntactically correct code. It handles multi-file context well within its 205K window, making it practical for codebase-level tasks rather than just snippet generation.
Agentic and tool-use pipelines
The BrowseComp score (76.3%) and the model's strong GPQA performance suggest it handles multi-step reasoning chains reliably. For developers building RAG pipelines, research agents, or automated QA systems, M2.5 at $0.30/M input is dramatically more economical than proprietary alternatives.
High-volume API workloads
Any application that fires thousands of API calls per day, customer support bots, classification pipelines, document processing, sees direct cost impact from the per-token price. At 1/10th the cost of Claude Sonnet, teams can iterate faster, run larger evaluation sets, and serve more requests within a fixed budget.
Long-context document processing
205K tokens fits roughly 150,000 words, the equivalent of a full-length novel or a substantial legal contract bundle. Teams doing contract review, financial report analysis, or regulatory compliance checking can pass entire documents in a single request rather than chunking.
Science and research Q&A
The 84.8% GPQA Diamond score makes M2.5 a strong option for scientific literature synthesis, academic research assistance, and graduate-level tutoring tools.
Limitations
Writing quality is not a standout
The lechmazur creative writing benchmark ranked M2.5's predecessor (M2.1) at position 19 among evaluated models. M2.5 does not appear to represent a significant leap in prose quality. For creative writing, fiction, or copy that needs genuine stylistic flair, models like Kimi K2.5 or the Claude family perform better.
Chinese lab governance and data considerations
MiniMax is a Chinese company operating under Chinese data law. For enterprises with strict data residency requirements or sensitive IP concerns, self-hosting the open weights is the practical path. Using the MiniMax API means data travels to servers subject to Chinese jurisdiction, a compliance consideration for regulated industries.
Smaller active parameter count
At 10B active parameters per forward pass, there are edge cases, particularly in nuanced reasoning chains that require holding many variables simultaneously, where the sparse activation pattern may underperform a dense 70B model. The benchmarks are strong in aggregate, but individual complex tasks can still expose this.
Not multimodal
M2.5 is a text-only model. It does not process images or video. For multimodal pipelines, Kimi K2.5 is the more relevant open-weights option in the same pricing tier.
Modified MIT, not pure open source
The "modified MIT" license requires attribution. This is permissive by commercial standards, but developers building products that require completely unrestricted licensing should review the terms before production deployment.
Verdict
MiniMax M2.5 is the most compelling cost-efficiency story in AI as of February 2026. It earns an 80.2% SWE-Bench score and 84.8% GPQA while costing $0.30/$1.20 per million tokens, numbers that would have been considered frontier-model performance six months ago at proprietary prices.
For teams building coding assistants, agentic pipelines, scientific research tools, or any high-volume API product, M2.5 should be the first model you evaluate. The cost difference versus Claude Sonnet or GPT-5 is large enough to materially change what is economically viable to build.
It is not the right choice if you need the best possible creative writing, strict open-source licensing without attribution requirements, or native multimodal capabilities. But as a coding and reasoning workhorse available at open-weights prices, it sets the bar for what developers should expect from any model in its tier.
Available on AfricanAI at the same competitive pricing, try it directly in your browser without API setup.
Sources: