Laguna XS 2.1 — AI Model Review & Pricing
Open-weight agentic coding model from Poolside with 70.9% SWE-bench Verified and a 256K token context window, priced at $0.06 input / $0.12 output per million tokens. Designed for long-context code generation and terminal-based coding agents.
Specifications
| Attribute | Value |
|---|---|
| Lab | Poolside AI |
| Tags | Coding Agentic Open Weight Fast |
| Release Date | 2026-06 |
| Context Window | 262,144 tokens |
| Input Price / 1M | $0.06 |
| Output Price / 1M | $0.12 |
| Input Modalities | Text |
| Output Modalities | Text |
Strengths
- 70.9% SWE-bench Verified (self-reported)
- 256K token context window (~256K tokens)
- $0.06/M input / $0.12/M output — 50-500x cheaper than leaders
- 33B total params with 3B activated per token for efficient inference
- Local deployment and open-weight availability
Weaknesses
- Text-only modality; no vision or multimodal support
- Self-reported benchmarks with patching adjustments (not independently verified)
- Sub-frontier performance vs Qwen3.6-35B-A3B on several tasks
- LMRank score 7.0 reflects limited benchmark diversity
Best For
- Agentic coding with tool-calling workflows
- Long-context code generation exceeding 100K tokens
- Self-hosted or air-gapped CI/CD pipelines
- Terminal-based coding agents and local DevContainers
In Depth: Laguna XS 2.1
Benchmark Performance
Laguna XS 2.1 coding model (#102). For context, Claude Sonnet 5 (rank #6) — but costs 17× more per output token, making XS 2.1 a practical alternative for budget-constrained agentic workflows.
Poolside reports 70.9% SWE-bench Verified, 63.1% Multilingual, 47.6% Pro, and 37.5% Terminal-Bench 2.0 — all self-reported with patching adjustments. These figures target real-world coding agent scenarios rather than synthetic QA. Its 256K context window supports large codebases without chunking. Buyers should treat benchmarks as vendor claims pending independent audit.
Pricing & Value
Laguna XS 2.1 costs $0.06 per million input tokens and $0.12 per million output tokens — lower than any Claude Sonnet 5 or GPT-5.5 Pro pricing tier.
Versus GPT-5.5 Pro ($30/M in, $180/M out), XS 2.1 output tokens are 1,500× cheaper. Even against Claude Sonnet 5 ($2/M in, $10/M out), XS 2.1 cuts output costs by 83%. Normalized per LMRank point, Sonnet 5 costs ~$1.06 per point (output) vs XS 2.1 at ~$0.017 per point — a 62× value gap. Poolside targets developers who run many low-cost agent loops.
Who Should Use This
Laguna XS 2.1 is for developer teams who need open-weight coding agents at low cost. Solo coders: run it locally for nightly PR reviews. Startups: deploy it in CI pipelines without API spend shock.
- Solo developer: local agent for code review and refactoring — trades raw accuracy for infinite runs
- Startup CTO: long-context CI/CD agent analyzing entire commit histories — missing multimodal understanding
- Enterprise architect: self-hosted agent for proprietary codebases — requires in-house MLOps for deployment
- Anti-persona: teams needing multimodal code reviews (diagrams, screenshots) should choose a vision-capable model
Release & Version History
Laguna XS 2.1 arrived June 2026 as Poolside’s third-generation coding model, following Laguna XS 2.0 (2025) and the original XS (2024). Each release tightened the open-weight focus on agentic coding.
Laguna XS 2.0 introduced Mixture-of-Experts and the 256K context window. Version 2.1 refines the MoE routing with sigmoid gating, adds FP8 KV cache, and ships with native interleaved thinking for tool-calling agents. The 33B-total/3B-activated parameter count stays unchanged, but the Muon optimizer improves training stability. Poolside positions this as a step toward smaller, faster agents rather than a frontier-chasing flagship.