Laguna M.1 — AI Model Review & Pricing
Open-weight agentic coding model from Poolside with 74.6% on SWE-bench Verified, 225.8B parameters (23.4B active), and a ~256K context window. Priced at $0.20 input / $0.40 output per million tokens via OpenRouter, with Apache 2.0 license and weights on Hugging Face.
Specifications
| Attribute | Value |
|---|---|
| Lab | Poolside AI |
| Tags | Coding Long Context |
| Release Date | 2026-04 |
| Context Window | 262,144 tokens |
| Input Price / 1M | $0.20 |
| Output Price / 1M | $0.40 |
| Input Modalities | Text |
| Output Modalities | Text |
Strengths
- 74.6% on SWE-bench Verified (vendor-reported)
- ~256K context window for long agentic coding sessions
- Apache 2.0 license with open weights on Hugging Face
- 50x cheaper input than Claude Fable 5 per million tokens
Weaknesses
- LMRank #91 with 7.4/10, far behind top models
- All benchmarks vendor-reported without independent verification
- Known training instabilities in the MoE architecture
- Requires substantial hardware for local deployment
Best For
- Agentic coding workflows with long-context requirements
- Open-weight code generation for self-hosted deployment
- Cost-sensitive SWE-bench tasks at scale
- Research and experimentation with MoE architectures
In Depth: Laguna M.1
Benchmark Performance
Laguna M.1 coding model (#91), trailing Claude Fable 5 (9.9/10, #1) by 2.5 points. Its 74.6% on SWE-bench Verified is vendor-reported and unverified, but positions it as a capable open-weight option for agentic coding.
Laguna M.1 achieves vendor-reported scores of 74.6% on SWE-bench Verified, 63.1% on SWE-bench Multilingual, 49.2% on SWE-Bench Pro (Public), and 45.8% on Terminal-Bench 2.0 (mean pass@1 over 4 runs, temperature=1.0, thinking enabled, ~256K context). These benchmarks target agentic coding tasks but lack independent verification.
Pricing & Value
Laguna M.1 pricing is $0.20 per million input tokens and $0.40 per million output tokens via OpenRouter, with a limited free tier. That is 50x cheaper than Claude Fable 5 on input and 125x cheaper on output.
At $0.20 input / $0.40 output per million tokens, Laguna M.1 costs 50x less input than Claude Fable 5 ($10.00 in / $50.00 out). Per LMRank point, Laguna M.1 costs $0.08 per point ($0.60 total per M tokens ÷ 7.4), versus $6.06 per point for Claude Fable 5 ($60.00 ÷ 9.9). That is 75x better price-per-point for buyers prioritizing cost over absolute rank.
Who Should Use This
Open-source developers get full model weights for self-hosted agentic coding. Cost-conscious teams save 50x on input versus Claude Fable 5. MoE researchers study a 225.8B sparse transformer.
- Open-source developers: Full model weights on Hugging Face for self-hosted agentic coding, but requires substantial hardware for local deployment.
- Cost-conscious teams: 50x cheaper input than Claude Fable 5, ideal for high-volume SWE-bench tasks, though rank #91 limits general-purpose use.
- MoE researchers: 225.8B total / 23.4B active parameters for studying sparse architectures, but known training instabilities may affect reproducibility.
- Enterprise buyers needing verified benchmarks: Avoid if independent validation is required—all scores are vendor-reported and unverified.
Release & Version History
Laguna M.1, released April 2026, is Poolside's first open-weight MoE model for agentic coding, following the smaller Laguna XS.2 and the dense Malibu 2.2. It offers a ~256K context window under Apache 2.0.
Released April 28, 2026, Laguna M.1 is a 225.8B parameter sparse MoE transformer with 23.4B active parameters per token. It follows the smaller Laguna XS.2 (33.4B total, 3B active) and the earlier dense model Malibu 2.2. Licensed Apache 2.0 with weights on Hugging Face, it targets agentic coding with a ~256K context window.