MiniMax M3: What the First Open-Weight Model to Beat GPT-5.5 on SWE-Bench Pro Means for Codex CLI Model Routing

Updated

codex-climinimax-m3open-weight-modelsmodel-routingswe-benchcost-optimisationnamed-profiles

MiniMax M3: What the First Open-Weight Model to Beat GPT-5.5 on SWE-Bench Pro Means for Codex CLI Model Routing

On 1 June 2026, Shanghai-based MiniMax released M3 — a 229.9-billion-parameter Mixture-of-Experts model with 9.8 billion parameters active per token, a one-million-token context window, and native multimodal support for text, image, and video input 1. Its headline claim: 59.0% on SWE-Bench Pro, edging past both GPT-5.5 (58.6%) and Gemini 3.1 Pro on the benchmark that has become the de facto measure of agentic coding capability 2. At $0.30 per million input tokens and $1.20 per million output tokens on OpenRouter — roughly 8-12x cheaper than GPT-5.5 — M3 forces a recalculation of every Codex CLI model routing decision 3.

This article examines the benchmark results in context, unpacks the architectural innovations that make M3’s long-context performance practical, and provides concrete Codex CLI configuration recipes for integrating M3 into a multi-model routing strategy.

The Benchmark Picture: Competitive but Complicated

M3’s SWE-Bench Pro score of 59.0% places it in genuinely frontier territory for an open-weight model 2. But the number deserves scrutiny.

Benchmark MiniMax M3 GPT-5.5 Claude Opus 4.7 Kimi K2.7 Code
SWE-Bench Pro 59.0% 58.6% — —
Terminal-Bench 2.1 66.0% 82.0% — —
MCP Atlas 74.2% — — —
KernelBench Hard 28.8% — — —
PostTrainBench 0.37 0.39 0.42 —

Three caveats matter for practitioners:

  1. Vendor-reported scores. M3’s SWE-Bench Pro result is self-reported by MiniMax and has not yet appeared on the independently verified leaderboard 4. GPT-5.5’s scores are third-party verified.

  2. The noise floor. The 0.4-point gap between M3 (59.0%) and GPT-5.5 (58.6%) on SWE-Bench Pro falls within the benchmark’s measurement uncertainty. An ICSE 2026 patch correctness study found SWE-bench systematically overestimates scores by 3.8-5.2 percentage points 5.

  3. Terminal-Bench tells a different story. GPT-5.5 scores 82% on Terminal-Bench 2.1 versus M3’s 66.0% — a 16-point gap that reflects GPT-5.5’s native advantage when operating inside the Codex CLI harness 2.

The practical reading: M3 is a genuine peer on repository-level bug-fixing tasks but remains behind on the interactive terminal workflows where Codex CLI spends most of its time.

Architecture: Why MSA Matters for Long Sessions

M3 introduces MiniMax Sparse Attention (MSA), a mechanism that makes one-million-token contexts economically viable 1. Traditional dense attention scales quadratically with sequence length. MSA uses a “KV outer gather Q” approach — each key-value block is read once, memory access is contiguous, and arithmetic intensity improves substantially over earlier sparse methods such as DSA and MoBA 1.

The numbers are striking: per-token compute at one million tokens drops to one-twentieth of M3’s predecessor, with more than 9x faster prefilling and more than 15x faster decoding 1. MiniMax reports MSA is 4x faster than open-source Flash-Sparse-Attention and flash-moba implementations 1.

graph LR
    A[Input Tokens] --> B[MoE Router<br/>256 experts]
    B --> C[9.8B Active<br/>Parameters]
    C --> D[MSA Layer<br/>KV outer gather Q]
    D --> E[1M Token<br/>Context Window]
    E --> F[Output]
    style D fill:#f9f,stroke:#333

For Codex CLI users, this matters in two scenarios:

  • Large-repository exploration. When tool_output_token_limit is set high and the agent reads multiple files in sequence, M3 can maintain coherence across a broader context without the compaction events that interrupt GPT-5.5 sessions.
  • Long-horizon agentic tasks. MiniMax demonstrated M3 autonomously reproducing an ICLR 2025 Outstanding Paper across 18 commits over twelve hours, and improving a CUDA kernel’s hardware utilisation from 7.6% to 71.3% over 24 hours with 1,959 tool calls 1.

The Cost Equation

The pricing gap between M3 and GPT-5.5 is the most immediately actionable finding for teams managing Codex CLI budgets.

  MiniMax M3 (OpenRouter) GPT-5.5 (OpenAI) Ratio
Input (per 1M tokens) $0.30 $5.00 16.7x cheaper
Output (per 1M tokens) $1.20 $30.00 25.0x cheaper

MiniMax also offers subscription tiers through its own platform: Plus ($20/month, ~1.7B tokens), Max ($50/month, ~5.1B tokens), and Ultra ($120/month, ~9.8B tokens) 1. For teams running batch operations via codex exec, the per-token API pricing through OpenRouter is typically more economical.

The cost arithmetic is simple. A typical Codex CLI session consuming 100,000 input tokens and 20,000 output tokens costs approximately $0.17 with GPT-5.5 versus $0.05 with M3. Over 50 sessions per day, that is $8.50 versus $2.50 — a saving of $180 per developer per month.

Configuring M3 as a Codex CLI Provider

Direct API Access

Add MiniMax as a custom provider in ~/.codex/config.toml:

[model_providers.minimax]
name = "MiniMax"
base_url = "https://api.minimax.io/v1"
env_key = "MINIMAX_KEY"

Set your API key:

export MINIMAX_KEY="<your-key-here>"

Via OpenRouter

If you already use OpenRouter for multi-provider routing:

[model_providers.openrouter]
name = "OpenRouter"
base_url = "https://openrouter.ai/api/v1"
env_key = "OPENROUTER_KEY"

Then reference the model as minimax/minimax-m3 when using the OpenRouter provider 3.

Named Profile for M3

Create a dedicated profile at ~/.codex/minimax.config.toml:

model = "minimax-m3"
model_provider = "minimax"

[model_providers.minimax]
name = "MiniMax"
base_url = "https://api.minimax.io/v1"
env_key = "MINIMAX_KEY"

Activate with:

codex --profile minimax "refactor the authentication module"

A Practical Routing Strategy

The benchmark data suggests a tiered routing approach where task complexity determines model selection.

flowchart TD
    A[Incoming Task] --> B{Task Type?}
    B -->|Interactive terminal<br/>multi-step debugging| C[GPT-5.5<br/>Terminal-Bench: 82%]
    B -->|Repository-level<br/>bug fix / feature| D{Budget<br/>Constraint?}
    B -->|Batch codex exec<br/>bulk operations| E[MiniMax M3<br/>8-12x cheaper]
    D -->|Cost-sensitive| E
    D -->|Quality-critical| C
    E --> F[Review Output<br/>PostToolUse hook]
    C --> G[Standard Flow]
    F --> H{Passes<br/>quality gate?}
    H -->|Yes| I[Accept]
    H -->|No| C

Profile-Based Routing in Practice

Define three profiles that encode the routing decision:

# ~/.codex/default.config.toml — GPT-5.5 for interactive work
model = "gpt-5.5"

# ~/.codex/minimax.config.toml — M3 for cost-sensitive tasks
model = "minimax-m3"
model_provider = "minimax"

[model_providers.minimax]
name = "MiniMax"
base_url = "https://api.minimax.io/v1"
env_key = "MINIMAX_KEY"

# ~/.codex/batch.config.toml — M3 with tighter token budgets
model = "minimax-m3"
model_provider = "minimax"
model_auto_compact_token_limit = 80000

[model_providers.minimax]
name = "MiniMax"
base_url = "https://api.minimax.io/v1"
env_key = "MINIMAX_KEY"

Use them from CI or scripts:

# Batch processing with M3
codex exec --profile batch "update all copyright headers to 2026"

# Interactive debugging stays on GPT-5.5
codex --profile default

AGENTS.md Model Guidance

Encode routing preferences in your project’s AGENTS.md so the agent itself can inform model selection:

## Model Routing

- Routine refactoring, linting fixes, and documentation updates: use `--profile minimax`
- Multi-file architectural changes and debugging: use default GPT-5.5
- Batch operations via `codex exec`: use `--profile batch`

The Open-Weight Convergence

M3 is not an isolated event. The open-weight coding model landscape has compressed dramatically in the first half of 2026:

  • Devstral Small 2 (24B): 68% on SWE-Bench Verified, runs on a single RTX 4090 6
  • Kimi K2.7 Code: purpose-built for software engineering with dual OpenAI and Anthropic API compatibility 7
  • DeepSeek V4-Pro: competitive on agentic benchmarks at a fraction of proprietary model costs 8

The pattern is clear: quality training data and specialised architectures matter more than raw parameter count. Skywork-SWE demonstrated this empirically — a 32B model fine-tuned on 8,209 rigorously validated trajectories achieved 38.0% on SWE-Bench Verified, outperforming 72B and 671B general-purpose models without specialised SWE training 9.

For Codex CLI users, this convergence means the model_provider configuration in config.toml is no longer a one-time decision. It is a continuously tuneable parameter that should respond to the evolving price-performance frontier.

Caveats and Risk Mitigation

Before routing production workloads through M3, consider these risks:

  1. Unverified benchmarks. Until M3’s SWE-Bench Pro score appears on an independent leaderboard, treat 59.0% as an upper bound estimate 4.

  2. Tool-call compatibility. M3 supports the OpenAI tool specification 3, but edge cases in complex multi-tool orchestration may behave differently from GPT-5.5. Test your specific MCP server configurations before switching profiles.

  3. Availability and rate limits. MiniMax’s API infrastructure is newer and less battle-tested than OpenAI’s. For mission-critical workflows, configure a fallback in your CI pipeline:

codex exec --profile minimax "task" || codex exec --profile default "task"

  1. Regional considerations. MiniMax offers separate API endpoints for international (api.minimax.io) and Chinese (api.minimaxi.com) users 10. Ensure your endpoint matches your deployment region.

  2. Weight release timing. MiniMax announced open-source weights within ten days of the 1 June launch 1. Verify current availability before planning self-hosted deployments.

What to Do This Week

  1. Register for a MiniMax API key at platform.minimax.io or access M3 through your existing OpenRouter account.
  2. Create a minimax.config.toml profile using the configuration above.
  3. Run your existing codex exec batch tasks through M3 for one week and compare output quality against GPT-5.5 results.
  4. Add a PostToolUse hook that logs model, token count, and task outcome to a local CSV — you will want this data when GPT-5.6 arrives and the routing calculation changes again.

Citations

  1. MiniMax, “MiniMax M3: Frontier Coding, 1M Context, Native Multimodality — All in One Model,” MiniMax Blog, 1 June 2026. https://www.minimax.io/blog/minimax-m3 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  2. MarkTechPost, “MiniMax Releases MiniMax M3 with MSA Architecture Supporting 1M-Token Context, Native Multimodality, and Agentic Coding,” 1 June 2026. https://www.marktechpost.com/2026/06/01/minimax-releases-minimax-m3-with-msa-architecture-supporting-1m-token-context-native-multimodality-and-agentic-coding/ ↩ ↩2 ↩3

  3. OpenRouter, “MiniMax M3 - API Pricing & Benchmarks,” accessed 20 June 2026. https://openrouter.ai/minimax/minimax-m3/api ↩ ↩2 ↩3

  4. ofox.ai, “MiniMax M3 vs GPT-5.5: SWE-Bench Pro, 8x Price Gap, A/B Both via ofox (2026),” June 2026. https://ofox.ai/blog/minimax-m3-vs-gpt-5-5-coding-benchmark-2026/ ↩ ↩2

  5. ICSE 2026 Patch Correctness Study, as cited in ofox.ai comparison. SWE-bench systematic overestimation of 3.8-5.2 percentage points. ↩

  6. Pinggy, “Best Open Source Self-Hosted LLMs for Coding in 2026,” June 2026. https://pinggy.io/blog/best_open_source_self_hosted_llms_for_coding/ ↩

  7. Flowtivity, “Kimi K2.7 Code vs MiniMax M3: Open-Source AI Coding Models Compared (June 2026),” June 2026. https://flowtivity.ai/blog/kimi-k2-7-code-vs-minimax-m3/ ↩

  8. kilo.ai, “Best Open-Source & Open-Weight Coding Models (2026),” accessed 20 June 2026. https://kilo.ai/open-source-models ↩

  9. Zeng et al., “Skywork-SWE: Unveiling Data Scaling Laws for Software Engineering in LLMs,” arXiv:2506.19290, June 2025. https://arxiv.org/abs/2506.19290 ↩

  10. MorphLLM, “Codex config.toml (2026): Add Any Custom Provider in 6 Lines,” accessed 20 June 2026. https://www.morphllm.com/codex-provider-configuration ↩