MiniMax M2.5: Coding Benchmarks, Pricing, and Guide
MiniMax M2.5 scores 80.2% on SWE-Bench Verified in MiniMax’s runs and launched at $0.30/$2.40 per million tokens (Lightning). Benchmarks, pricing, API access.
SWE-Bench Verified
Task Cost vs Opus 4.6
Faster Than M2.1
Training Environments
Key Takeaways
Corrected October 21, 2026: we re-checked this guide against MiniMax’s announcement, model card, licence and launch-week price list, and against Anthropic’s, OpenAI’s and Google’s own figures. Standard M2.5’s input price ($0.30, not $0.15), Gemini 3 Pro’s SWE-bench Verified score (76.2%, not 78%), rival output prices, MiniMax’s home city (Shanghai) and M2.1’s release date are corrected. Rival scores MiniMax measured itself are now labelled, benchmark descriptions follow MiniMax’s methodology, and an invented “architect mode” and unsourced cost arithmetic are removed.
Update, October 2026: within a week of launch MiniMax renamed M2.5-Lightning to MiniMax-M2.5-highspeed and listed its input at $0.60 per million tokens (output unchanged at $2.40). In March 2026 the weights moved to the MiniMax Model License; the code stays under modified MIT. MiniMax has since released M2.7 (March 18), M3 (June 1) and M3.1 Flash (October 19, 2026), and its current price list shows only the M3 models.
MiniMax released M2.5 on February 12, 2026. MiniMax reports 80.2% on SWE-Bench Verified, 0.6 percentage points below Anthropic’s published 80.8% for Claude Opus 4.6, and by output price it costs roughly a tenth to a twentieth as much as Opus 4.6. It follows M2.1, released in December 2025, which scored 74% on SWE-Bench Verified. M2.5’s published configuration is unchanged from M2.1’s, a mixture-of-experts design MiniMax describes as 230 billion total and 10 billion active parameters.
M2.5 ships in two API variants: a standard model at about 50 tokens per second and a Lightning variant at 100. MiniMax says both were trained with its Forge reinforcement learning framework across hundreds of thousands of real-world environments, including more than 200,000 for coding. MiniMax’s API serves a 204,800-token context window, and the weights were released on Hugging Face under a modified MIT licence.
This guide covers M2.5's benchmark performance, pricing, technical architecture, and how it compares to Claude Opus 4.6, GPT-5.2, and Gemini 3 Pro across coding, search, tool calling, and office work tasks.
What Is MiniMax M2.5?
MiniMax M2.5 is a frontier AI model built for agentic workflows — tasks that require the model to use tools, write code, search the web, and complete multi-step processes autonomously. Unlike models optimized primarily for chat or single-turn generation, M2.5 is designed to work as what MiniMax calls a “professional employee” capable of sustained, independent work.
- • 100 tokens per second output
- • $0.30/M input, $2.40/M output at launch
- • ~$1/hour at 100 TPS (MiniMax’s figure)
- • Optimized for high-throughput workflows
- • About 50 tokens per second output
- • $0.30/M input, $1.20/M output
- • ~$0.30/hour at 50 TPS (MiniMax’s figure)
- • Same capability as Lightning
MiniMax says a spec-writing habit emerged during training: before writing code, M2.5 plans a project’s features, structure and UI the way a software architect would. This is model behaviour rather than a separate mode, and MiniMax pitches it at full-stack projects across web, Android, iOS and Windows, from system design through feature iteration to code review and testing.
Key context: MiniMax is a Shanghai-based AI company that previously released M2.1 in December 2025. M2.5 improves on it: 80.2% vs 74% on SWE-Bench Verified, 37% faster end-to-end task completion on that benchmark, and gains across search, tool calling, and office work.
Coding Performance
M2.5’s headline number is 80.2% on SWE-Bench Verified, a benchmark of human-validated real GitHub issues from Python repositories. MiniMax ran it in Claude Code with its own system prompt, averaged over four runs. That is 0.6 points below Anthropic’s 80.8% for Claude Opus 4.6 (averaged over 25 trials) and 0.2 points above OpenAI’s 80.0% for GPT-5.2 Thinking; Google reports 76.2% for Gemini 3 Pro.
SWE-Bench Verified
80.2%Real-world GitHub issue resolution in Python repositories. In the Droid and OpenCode harnesses MiniMax measured 79.7% and 76.1%.
Multi-SWE-Bench
51.3%A multilingual issue-resolving benchmark covering Java, TypeScript, JavaScript, Go, Rust, C and C++. MiniMax’s own runs put Opus 4.6 at 50.3% and Gemini 3 Pro at 42.7%; neither vendor publishes a score.
SWE-Bench Pro
55.4%Scale AI’s separate, harder benchmark of long-horizon tasks from 41 repositories, often needing multi-file patches. OpenAI reports 55.6% for GPT-5.2 Thinking on the public set.
Agentic Coding Harnesses
MiniMax ran SWE-Bench Verified, SWE-Bench Multilingual, SWE-Bench Pro and Multi-SWE-Bench on its own infrastructure with Claude Code as the agent scaffold, overriding the default system prompt and averaging four runs. To test how well the model generalises to unfamiliar harnesses, it also ran SWE-Bench Verified in Droid (79.7%, against 78.9% for Opus 4.6) and OpenCode (76.1%, against 75.9%) with their default prompts. This reflects real-world usage where AI models interact with codebases through tool calls rather than generating isolated code snippets.
MiniMax says M2.5 was trained on more than ten languages, including Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart and Ruby. SWE-Bench Verified is Python-only, so the multilingual evidence comes from SWE-Bench Multilingual (74.1%) and Multi-SWE-Bench (51.3%), both MiniMax’s runs.
Search and Tool Calling
Beyond coding, M2.5 shows competitive results in web search and tool calling benchmarks — capabilities essential for agentic workflows that require gathering information and interacting with external systems.
- BrowseComp (w/ context): 76.3% — finding hard-to-locate information across many web pages; MiniMax’s harness discards the history once it passes 30% of the context window. This is the figure in MiniMax’s announcement and overview chart; its search-section chart shows 75.1%
- Wide Search: 70.3% — broad information retrieval across diverse domains
- BFCL multi-turn: 76.8% — ahead of Opus 4.6 (63.3%) and Gemini 3 Pro (61%) in MiniMax’s runs
- 20% fewer rounds: About 20% fewer search and tool-calling rounds than M2.1 on BrowseComp, Wide Search and RISE
The BFCL (Berkeley Function Calling Leaderboard) multi-turn result is M2.5’s widest lead over Opus 4.6 among the benchmarks in this guide: 76.8% against 63.3%, a 13.5-point lead, with both figures from MiniMax’s own runs. MiniMax attributes its gains to RL training in real tool environments. Needing about 20% fewer rounds than M2.1 to finish search tasks should also cut latency and cost per task in production deployments.
Office Work Capabilities
M2.5 extends beyond coding and search into office productivity tasks. MiniMax says senior professionals in finance, law and the social sciences helped define requirements, give feedback and build training data for Word, PowerPoint and Excel financial-modeling work. The office scores below come from MiniMax’s internal benchmarks.
GDPval-MM
59%Average pairwise win rate against other mainstream models, judged by an LLM over full agent trajectories on tasks built from the open GDPval test set
MEWC
74.4%179 Microsoft Excel World Championship problems from 2021–2026, scored cell by cell against the answers
Supported Office Tasks
- Word documents: Formatting documents and producing structured outputs such as research reports that follow a set framework
- PowerPoint presentations: Building and editing slide decks
- Excel spreadsheets: Calculations and spreadsheet work through Excel tools
- Financial modeling: Building and validating financial models that follow an organisation’s own modeling and risk-control standards
AI-powered office automation is here. Models like M2.5 can handle document creation, data analysis, and reporting at scale. Explore our AI & Digital Transformation Services to integrate AI into your business workflows.
Efficiency and Speed
M2.5 finishes SWE-Bench Verified tasks 37% faster than M2.1: average end-to-end runtime fell from 31.3 to 22.8 minutes, level with Claude Opus 4.6’s 22.9 minutes in MiniMax’s measurement. The Lightning variant serves 100 tokens per second, which MiniMax describes as nearly twice the speed of other frontier models.
Speed
- • 37% faster than M2.1 on SWE-Bench Verified
- • Lightning: 100 TPS output
- • Standard: about 50 TPS output
- • 22.8 min per task vs 22.9 for Opus 4.6
Task Efficiency
- • 3.52M tokens average per SWE-Bench task (M2.1: 3.72M)
- • About 20% fewer search rounds than M2.1
- • Parallel tool calling
- • Spec-first planning before coding
The 3.52 million tokens per SWE-Bench Verified task is significant for cost planning, but it counts input and output tokens together, so it cannot be priced at the output rate alone. MiniMax’s own estimate is that M2.5’s total cost per task is about 10% of Claude Opus 4.6’s, for scores within a point of each other.
Pricing and Cost Analysis
MiniMax positions M2.5 as a cost-disruption play: by its reckoning the output price is a tenth to a twentieth of Opus, Gemini 3 Pro and GPT-5. Against list prices that holds for Claude Opus 4.6 ($25 per million output tokens); against GPT-5.2 ($14) and Gemini 3 Pro ($12) the gap is roughly 5x to 12x. Either way, sustained agentic use becomes viable for tasks that would be expensive on the alternatives.
| Model | Input ($/M) | Output ($/M) | Speed (TPS) | Hourly Cost |
|---|---|---|---|---|
| M2.5-Lightning | $0.30 | $2.40 | 100 | ~$1.00 |
| M2.5 Standard | $0.30 | $1.20 | 50 | ~$0.30 |
Source: MiniMax’s pay-as-you-go price list as archived on February 13, 2026; hourly costs are MiniMax’s own figures. Both models also supported prompt caching, at $0.03 per million tokens read and $0.375 per million written.
$1 per hour math: At 100 TPS, M2.5-Lightning generates 360,000 output tokens per hour. At $2.40/M, that’s $0.86 in output tokens plus input costs — roughly $1 per hour of continuous generation. MiniMax’s own illustration uses the 50 TPS model: four instances running for a year for about $10,000. This makes 24/7 agentic operation economically feasible. For teams looking to integrate cost-effective AI models into development workflows, see our Web Development Services.
How M2.5 Compares
The following comparison puts M2.5 against three rival frontier models from MiniMax’s launch charts, as of February 2026. M2.5 figures are MiniMax’s own runs. Rival figures are each vendor’s own published score where one exists; cells marked * are MiniMax’s runs of the rival model, taken from its launch charts.
| Benchmark | M2.5 | Opus 4.6 | GPT-5.2 | Gemini 3 Pro |
|---|---|---|---|---|
| SWE-Bench Verified | 80.2% | 80.8% | 80.0% | 76.2% |
| Multi-SWE-Bench | 51.3% | 50.3%* | — | 42.7%* |
| BrowseComp | 76.3% | 84.0% | 65.8% | 59.2% |
| BFCL Multi-Turn | 76.8% | 63.3%* | — | 61%* |
| Output Price ($/M) | $1.20–$2.40 | $25 | $14 | $12 |
Sources: MiniMax’s M2.5 announcement and its benchmark charts; Anthropic’s Claude Opus 4.6 announcement; OpenAI’s GPT-5.2 announcement (GPT-5.2 Thinking scores); Google’s Gemini 3 developer post and Google DeepMind’s Gemini Pro evaluation table. M2.5’s price range is standard to Lightning; Gemini 3 Pro’s is for prompts up to 200K tokens.
Comparison Date: February 2026. Harnesses differ: MiniMax ran its SWE benchmarks in Claude Code and its BrowseComp with context management, Anthropic averaged SWE-bench Verified over 25 trials, and each lab uses its own search setup, so gaps of a point or two are not like-for-like. M2.5’s BrowseComp is the 76.3% in MiniMax’s announcement; its search-section chart shows 75.1%. Anthropic later revised Opus 4.6’s BrowseComp to 83.7%. Dashes mark scores that neither the vendor nor MiniMax published. Verify current specifications before making production decisions.
Technical Deep Dive: RL Scaling
MiniMax credits much of M2.5’s performance to scaling reinforcement learning with its Forge framework, a purpose-built system for training AI agents across diverse real-world environments. MiniMax says most of the tasks and workspaces inside the company have been turned into RL training environments, hundreds of thousands so far.
The Forge Framework
Forge is MiniMax’s in-house, agent-native RL training infrastructure. It adds an intermediary layer that fully decouples the training-inference engine from the agent, so arbitrary agent scaffolds and tools can be plugged in and the model can be optimised to generalise across them.
- 200,000+ coding environments: Real-world codebases across more than ten programming languages, part of the hundreds of thousands of environments MiniMax has built
- About 40x training speedup: From asynchronous scheduling that balances throughput against off-policy samples, and tree-structured merging of training samples
- Process rewards: A process reward mechanism monitors generation quality end to end, and task completion time is scored to trade intelligence against response speed
CISPO Algorithm
CISPO (Clipped IS-weight Policy Optimization) is the RL algorithm MiniMax introduced with its MiniMax-M1 model in June 2025. MiniMax kept using it for M2.5 to keep its mixture-of-experts model stable during large-scale training:
- It clips importance-sampling weights rather than token updates, which the M1 paper found outperformed other competitive RL variants
- For M2.5, MiniMax pairs it with a process reward to handle credit assignment across long agent rollouts
Why this matters: Forge trains M2.5 to complete tasks inside real environments and many different agent scaffolds, and MiniMax reports that this carries over to harnesses it did not train for: its SWE-Bench Verified scores hold up in Droid and OpenCode as well as Claude Code. That is the behaviour agentic benchmarks like BFCL and SWE-Bench reward, where models must use tools across multi-step workflows.
MiniMax Agent Platform
M2.5 powers MiniMax’s Agent platform, a product that packages the model’s capabilities into ready-to-use workflows. The platform extends beyond API access to provide integrated tools for office work, coding, and information retrieval. For context on the platform's evolution from the original M2 Agent release, see our earlier coverage.
Standardized skills that MiniMax Agent loads automatically by file type in MAX mode, for Word formatting, PowerPoint editing and Excel calculations. Users interact through natural language instructions.
Users have built more than 10,000 reusable Experts that combine Office Skills with domain know-how, such as a research framework or in-house financial-modeling standards. MiniMax also ships Expert suites for office work, finance and programming.
MiniMax reports that M2.5 autonomously completes 30% of overall tasks across its research and development, product, sales, HR and finance functions, and that M2.5-generated code makes up 80% of newly committed code at the company.
The Agent platform positions MiniMax beyond the API-only model provider category. By combining M2.5 with integrated tools, MiniMax offers a product in the same space as enterprise AI assistants from Microsoft (Copilot), Google (Gemini), and startups building agentic coding tools.
Getting Started with MiniMax M2.5
M2.5 is accessible through the MiniMax API and the Agent platform, and its weights are on Hugging Face for self-hosting with SGLang, vLLM or Transformers. The API offers Anthropic-compatible and OpenAI-compatible endpoints, making integration straightforward for teams already using standard LLM APIs.
1. API Access
Create an account on MiniMax’s API platform to get credentials. At launch MiniMax recommended its Anthropic-compatible API and also offered an OpenAI-compatible one, both with streaming and tool calling. Standard and Lightning are separate model names (MiniMax-M2.5 and MiniMax-M2.5-lightning at launch), each with a 204,800-token context window.
2. Choose Your Variant
Select based on your primary use case:
- M2.5-Lightning: Best for real-time applications, interactive coding assistants, and latency-sensitive workflows
- M2.5 Standard: Best for batch processing, background agentic tasks, and cost-sensitive high-volume operations
3. Agent Platform
For non-developer users or teams wanting ready-to-use AI workflows, the MiniMax Agent platform provides pre-built office skills, domain experts, and a conversational interface, available as MiniMax Agent on the web.
Conclusion
MiniMax M2.5 delivers frontier-level coding scores at a fraction of competitor pricing. MiniMax’s 80.2% on SWE-Bench Verified places it alongside Claude Opus 4.6 and GPT-5.2, while $2.40 per million output tokens (Lightning) against $14 to $25 for those two rivals makes sustained agentic operation far cheaper.
MiniMax credits the Forge RL framework and the CISPO algorithm, training in hundreds of thousands of real-world environments, for its strongest results: multi-turn tool calling (76.8% vs 63.3% for Opus 4.6 on BFCL, in MiniMax’s runs) and fewer rounds per task than M2.1.
For teams evaluating AI coding models, M2.5 warrants serious consideration. Benchmark scores close to models with roughly 5x to 20x higher output prices make it a strong option for cost-conscious deployments, while the office work and search capabilities position it as a general-purpose agentic model rather than a coding-only tool.
Integrate Frontier AI Into Your Workflow
We help businesses evaluate, integrate, and scale AI-powered development and automation workflows using the latest frontier models.
Sources
- MiniMax: M2.5 announcement and benchmark charts (February 12, 2026); MiniMax-M2.5 model card and configuration; launch licence (modified MIT) and the MiniMax Model License that replaced it for the weights in March 2026 (the Hugging Face card still labels the model Modified-MIT); M2.1 announcement.
- MiniMax API documentation: price list, February 13, 2026; price list, February 17, 2026 (highspeed rename); text-generation guide (context window, compatible APIs); models overview (M2.1 parameters); model release notes and the current price list (read October 22, 2026).
- Rival models: Anthropic, Claude Opus 4.6 (February 5, 2026); OpenAI, GPT-5.2; Google, Gemini 3 and Gemini 3 for developers; Google DeepMind, Gemini Pro evaluations.
- Papers: MiniMax-M1 (CISPO); Multi-SWE-bench; SWE-Bench Pro.
- Company: Caixin Global’s report on MiniMax’s Hong Kong listing (January 1, 2026), for its Shanghai base.
Frequently Asked Questions
Deep dives on AI, marketing and development.
Practical guides and fresh insights by email. No recycled takes.
Related Guides
Continue exploring AI coding models and developer tools