Laguna by Poolside Review (2026): Pricing, Features & Honest Verdict

Laguna by Poolside is a pair of foundation models built from scratch for agentic coding and long-horizon engineering work. The verdict: strong benchmarks and a genuinely useful open model, wrapped in pricing that is free for now and unclear for later. Best for developers building coding agents. Price: free for a limited time via API and OpenRouter. Rating: 7.5/10.

What is Laguna by Poolside?

Laguna is a family of two AI models from Poolside, an American AI startup focused entirely on code. There is Laguna M.1, a proprietary 225-billion-parameter mixture-of-experts model with 23 billion active parameters, and Laguna XS.2, a smaller 33-billion-parameter open model with 3 billion active parameters released under the Apache 2.0 license. Both launched in 2026 and were trained from the ground up for agentic coding, meaning the kind of multi-step work where an agent plans, edits files, runs tools, and keeps going across a long task.

The headline numbers are real. Laguna M.1 hits 72.5 percent on SWE-bench Verified, the benchmark that measures how well a model resolves actual GitHub issues, and XS.2 reaches 68.2 percent. Those are serious scores for a startup going up against the frontier labs. M.1 carries a 128K context window with tool calling and reasoning, and it is aimed at high-stakes enterprise and government work. XS.2 is the more interesting story for most makers: it is small enough to download and run locally on a single GPU or even a laptop, fully offline and private.

What Are Laguna’s Key Features?

Built from scratch for agentic coding

Most coding models are general models that happen to be good at code. Poolside trained Laguna specifically for agentic workflows from the start. That shows up in long-horizon tasks where the model has to hold a plan across many steps, call tools, and recover from its own mistakes. M.1 performs at its best inside Poolside’s own coding agent, which is the environment it was tuned for. If your work is agent-driven rather than single-shot completions, that focus matters.

Laguna XS.2 runs locally and offline

This is the feature that should grab indie developers. XS.2 is Apache 2.0 licensed and available on Hugging Face, so you can download it, fine-tune it, quantize it, and serve it on your own hardware. It runs on a single GPU and even on a laptop without an internet connection. For anyone who cares about privacy, cost control, or not shipping their code to someone else’s API, a capable open coding model you fully own is a big deal.

Works with your existing tools

Both models are available on OpenRouter and through Poolside’s own API, so you can use them with the harnesses and tools you already have. You are not forced into a proprietary client to try them. M.1 supports tool calling and a 128K context window, which is enough to hold a meaningful chunk of a codebase in a single request. Slotting these into an existing agent setup is straightforward.

A real choice between open and frontier

The two-model approach is smart. XS.2 gives you a free, open, local option for everyday and privacy-sensitive work. M.1 gives you a high-end model for the hard, long, high-consequence problems. Poolside is also willing to grant access to M.1 weights for university and institutional research, which is more open than most labs at the frontier. Having both ends of the range from one team, trained the same way, is a cleaner story than mixing models from different vendors.

How Much Does Laguna Cost?

Right now, both models are free. Laguna M.1 and XS.2 are available at zero cost for a limited time through Poolside’s API and on OpenRouter, listed at $0 per million input and output tokens, and XS.2 is downloadable free from Hugging Face under Apache 2.0. That is great while it lasts, but the phrase “for a limited time” is the whole catch. Poolside has not committed to long-term token pricing for M.1, so the proprietary model could move to usage-based pricing that competes with Anthropic’s Claude or OpenAI’s models, which run a few dollars per million tokens at the frontier. XS.2 stays free forever because it is open licensed, so the safest bet for cost certainty is the open model.

Who is Laguna Best For?

Use Laguna if you build coding agents and want strong agentic performance without paying frontier prices today. Developers experimenting with autonomous coding, teams that want a local model they fully control, and researchers who can get weight access will all find something here. XS.2 in particular is a gift for anyone who wants a private, offline coding model on their own machine.

Skip Laguna, at least for production, if you need pricing you can plan around. The “free for a limited time” status on M.1 makes it risky to build a business on the proprietary model before the real pricing lands. It is also not the pick if you want a polished consumer app: this is foundation-model infrastructure, so you bring your own agent, harness, and tooling. If you want a finished product rather than a model, look elsewhere.

Best Laguna Alternatives

Anthropic Claude. The Claude family is a frontier option strong at coding and agentic work, with mature tool use and large context windows. It is more polished and more proven in production, but it is closed and priced per token. Pricing runs a few dollars per million tokens depending on the model.

Qwen Coder. Alibaba’s open coding models are the obvious comparison for XS.2, offering strong open-weight performance you can run yourself. Qwen Coder is a fair fight on the open-model side and worth benchmarking against XS.2 for your specific tasks. Free to download and run.

OpenAI Codex / GPT models. OpenAI’s coding-focused models power a lot of agent tooling and are battle-tested. They are closed and metered, but the ecosystem and reliability are hard to beat for production. Priced per token, typically a few dollars per million.

Final Verdict: Is Laguna Worth It?

Laguna is one of the more credible challengers to come out of a coding-focused startup. The benchmarks are real, the agentic training is a genuine differentiator, and XS.2 being open, local, and Apache 2.0 makes it immediately useful for anyone who wants a private coding model they own. While it is free, there is no reason not to test both models against your current stack and see how they hold up on your actual tasks.

The reservation is all about what happens when the free window closes. M.1 is excellent but its long-term pricing is unknown, so building production on it today is a bet. XS.2 sidesteps that entirely because it is open, which is why it is the safer long-term pick for most makers. We rate Laguna 7.5 out of 10: a strong, honestly impressive model family held back only by pricing that has not fully shown its hand yet.