MiniMax M2.5 for Coding: The $1/Hour Frontier Model That Matches Opus

Summary

MiniMax shipped M2, M2.1, and M2.5 in three and a half months, from late October to mid-February 2026. Each version closed the gap to proprietary frontier models. M2.5 closed it almost entirely for coding tasks, while remaining 20x cheaper. The model weights are open, the API is cheap, and GGUF quantizations are available for local deployment.

Benchmark Breakdown

Two numbers stand out. First, M2.5 leads Multi-SWE-Bench at 51.3% vs Opus's 50.3%. Multi-SWE-Bench tests multi-file, cross-language tasks, the kind of work that most closely resembles real software engineering. Second, M2.5 scores 76.8% on BFCL multi-turn function calling vs Opus's 63.3%. For agentic workflows that rely on tool use, that 13.5-point gap is significant.

Architecture: Why MoE Makes This Possible

M2.5 has 229 billion total parameters but only activates 10 billion per token. It routes each token through 8 of 256 available experts. This is how MiniMax achieves frontier accuracy at a fraction of the compute cost: most of the model sits idle on any given inference step.

The MoE architecture explains why M2.5 can be both cheap and accurate. Dense models like Opus activate every parameter on every token. MoE models select a small subset of specialists per token. The trade-off is complexity in routing and training stability. MiniMax solved the training stability problem with CISPO. They solved the routing problem with 256 fine-grained experts instead of the 8-16 experts used by earlier MoE models like Mixtral.

Pricing: The 20x Advantage

At $0.30/M input and $1.20/M output, M2.5 is 50x cheaper than Opus on input tokens and 62x cheaper on output tokens. The M2.5-Lightning variant doubles throughput to 100 tok/sec for double the output cost, still 31x cheaper than Opus. For a team running 100 agentic coding tasks per day, M2.5 costs ~$15/day. The same workload on Opus costs ~$300/day. Over a month, that is $450 vs $9,000.

Stat Comparison: M2.5 vs Opus 4.6

Beyond raw benchmarks, these models differ in how they handle different types of coding work.

Where M2.5 Wins

The cost advantage compounds when you consider subagent architectures. Running 4 parallel subagents on Opus costs $12/task. On M2.5, the same 4-agent setup costs $0.60. Multi-agent workflows, where dedicated context windows prevent context pollution, are the direction coding agents are moving. M2.5 makes that architecture economically viable for teams that cannot justify Opus-level spend.

Where Opus 4.6 Still Wins

The honest assessment: Opus 4.6 is a better model. M2.5 is a better value. For 95% of coding tasks, you will not notice the 0.6-point accuracy difference. For the remaining 5%, complex terminal workflows, massive tool orchestration, and deep architectural reasoning, Opus is worth the premium. The smart approach is using M2.5 for volume and Opus for the hard problems.

Using M2.5 with WarpGrep

Model quality is one variable. Search quality is the other. Coding agents spend the majority of their tokens navigating codebases, not writing code. WarpGrep is an MCP server that gives any coding agent parallel, sub-6-second codebase search. It works with M2.5 the same way it works with Opus or GPT-5.2.

# WarpGrep works as an MCP server with any model
# Aider, Continue, OpenHands, Claude Code all support MCP

# In your MCP config:
{
  "mcpServers": {
    "warpgrep": {
      "command": "npx",
      "args": ["warpgrep-mcp"]
    }
  }
}

# M2.5 can then call WarpGrep tools:
# - search_codebase: parallel semantic + keyword search
# - find_references: cross-file dependency tracing
# - explore_architecture: structural codebase understanding

# The search happens in <6 seconds across 100K+ file repos
# M2.5's strong tool calling (76.8% BFCL) means
# reliable WarpGrep integration out of the box

On SWE-bench Pro, Opus 4.6 + WarpGrep v2 scores 57.5%, up from 55.4% stock. That 2.1-point improvement comes entirely from better search, not a better model. The same principle applies to M2.5: pair a strong model with strong search, and performance on real-world tasks improves more than switching to a more expensive model with worse search.

Frequently Asked Questions

How does MiniMax M2.5 compare to Claude Opus 4.6 for coding?

M2.5 scores 80.2% on SWE-Bench Verified vs Opus's 80.8%, a 0.6-point gap. M2.5 leads on Multi-SWE-Bench (51.3% vs 50.3%) and multi-turn function calling (BFCL: 76.8% vs 63.3%). Opus leads on Terminal-Bench 2.0 and MCP Atlas. The defining difference is cost: $0.15/task on M2.5 vs $3.00 on Opus. Both models complete SWE-Bench tasks in ~23 minutes on average.

Is MiniMax M2.5 open source?

Yes. Full weights are on HuggingFace under MiniMaxAI/MiniMax-M2.5. Community GGUF quantizations are available from Unsloth and others. You can self-host, fine-tune, or run quantized versions locally. The API is also available at $0.30/M input and $1.20/M output tokens.

What does "agent-native" mean?

MiniMax built a training framework called Forge that decouples the RL training engine from agent scaffolding. Instead of training the model to work well in one specific agent system, Forge introduces an intermediary layer that supports arbitrary agent integrations during training. The result is a model that generalizes across Claude Code, Aider, Continue, OpenHands, and other agent frameworks without scaffold-specific tuning.

Should I use M2.5 or M2.5-Lightning?

Same model, different speed/cost trade-off. M2.5 runs at 50 tok/sec for $1.20/M output tokens. M2.5-Lightning runs at 100 tok/sec for $2.40/M output. Use Lightning for interactive development where latency matters. Use standard M2.5 for batch/background agentic tasks where cost matters more than speed.

Can I fine-tune M2.5 on my codebase?

Yes. Open weights mean full fine-tuning, LoRA, or QLoRA on your proprietary code. For teams with domain-specific patterns (internal frameworks, custom APIs, proprietary DSLs), fine-tuning M2.5 can close the remaining gap to Opus on your specific workload. Unsloth provides optimized fine-tuning scripts for M2.5.