The most interesting line in Poolside's pricing sheet for Laguna S 2.1 is not the model's benchmark scores. It is this: $0.10 per 1M input tokens for a context window that stretches to one million tokens. That means you can load an entire large-scale monorepo, a full incident trace, or a multi-file migration plan into a single coherent prompt for roughly ten cents. That changes the math on some of the most expensive, time-consuming work your senior engineers currently do by hand.

Poolside released Laguna S 2.1 on July 21, 2026, positioning it as the highest-capability open-weight agentic coding model available outside a proprietary API. The architecture is a 118B-parameter Mixture-of-Experts design with only 8B parameters activated per token, which keeps inference cost down while giving the model the reasoning depth of a much larger dense network. The 1M-token context is exclusive to the S 2.1 tier; other Laguna variants top out at 262,144 tokens. This is not a general-purpose chat model. It is purpose-built for long-horizon agentic work: multi-step reasoning, tool calling, thinking mode, and the kind of cross-repo analysis that currently requires a senior engineer to manually stitch together context across three terminal windows and two code search tabs.

Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-Bench Multilingual. SWE-Bench Multilingual is the more operationally meaningful number for most teams because it measures real-world GitHub issue resolution across multiple programming languages. A 78.5% resolution rate on that benchmark, for a model competing against systems with two to eight times as many active parameters, is a meaningful signal that the architecture's efficiency is real. Terminal-Bench 2.1 is newer and worth watching. It tests autonomous terminal task completion, which maps directly to the kinds of workflows your platform and infra engineers are trying to automate: provisioning scripts, CI/CD diagnosis, migration runners.

The phrase "weight class" matters here. Poolside is not claiming to beat GPT-5 or Gemini Ultra on absolute performance. They are claiming best-in-class efficiency at 8B activated parameters, competing with models like DeepSeek-V4-Flash and NVIDIA Nemotron 3 Ultra. For teams that care about cost-per-task rather than top-of-leaderboard bragging rights, that is the right comparison.

Most teams defaulting to closed-source models are paying a hidden tax on large-context work: chunking, stitching, re-prompting, and senior engineer review of outputs that missed context from a different chunk. That cost rarely shows up on an AI spend dashboard. It shows up in cycle time. Here is a concrete breakdown for three high-value use cases where Laguna S 2.1's 1M-token window materially changes the economics:

A typical database schema migration across a 300-file codebase requires loading roughly 600,000-800,000 tokens of context if you include the schema, affected files, migration history, and test suite. On a model capped at 128K context, you run this in four to six chunks, with a senior engineer reviewing handoff points.

At five major migrations per quarter, that is roughly $10,000 in recovered senior engineer time, not counting reduced error rates from context fragmentation.

Post-incident analysis across distributed systems means correlating logs, traces, code history, and runbooks. A comprehensive incident context package for a complex outage can easily exceed 500,000 tokens. Today, most teams are either truncating that context or having a senior SRE manually synthesize it. With a $0.10/1M input token endpoint, loading 500K tokens costs five cents. The model then runs thinking mode over the full trace and surfaces root cause candidates across the entire codebase slice. The senior SRE moves from analyst to decision-maker.

For platform teams maintaining shared libraries across twenty or more downstream services, a single breaking change audit requires scanning dependency trees, interface contracts, and call sites simultaneously. This is exactly where manual stitching is most expensive and most error-prone. A 1M-token window can hold the entire relevant graph in one pass. The cost of that single pass on Laguna S 2.1 is under $0.15 in input tokens. The cost of a senior engineer doing it manually across two days is closer to $2,400.

For engineering leaders building internal cost models, here is how Laguna S 2.1 sits against the most commonly used closed-source alternatives at the high-context tier:

The Laguna S 2.1 pricing is not just cheaper than closed-source alternatives at comparable tasks. It is an order of magnitude cheaper for high-context work, and it ships as an open-weight model under the OpenMDW-1.1 license, available via Hugging Face, OpenRouter, and the Poolside API. That matters for your negotiating leverage.

The benchmark scores will get updated. Pricing will shift. What does not change is the structural advantage of an open-weight, multi-endpoint model:

The "pool harness" reference is worth unpacking. Poolside's agent framework is built to wrap around this model, but because Laguna S 2.1 is open-weight, you can deploy the same model weights behind your own tooling, run it on your own infrastructure, or benchmark it head-to-head against other providers on your specific task distribution. No other provider at this performance tier gives you that flexibility. For platform and infra teams, this is not a minor detail. The ability to shift the same workflow across providers based on latency, cost, or quality without rewriting your agent stack is the kind of optionality that compounds over time. It also gives you a credible walk-away position when negotiating enterprise contracts with closed-source providers.

A large context window does not automatically produce better outputs. It produces more expensive, slower outputs if you use it indiscriminately. Three failure modes to build against:

The right adoption posture is a tiered routing model: routine coding tasks on a cheap fast model (under $1/1M input), medium-complexity multi-file work on a 200K-context model, and Laguna S 2.1 reserved for the long-tail tasks where full-repo context is genuinely load-bearing.

Use this framework to make the CFO case for allocating budget to a Laguna S 2.1 pilot:

For most platform teams running five or more major context-heavy tasks per week, the payback period on a Laguna S 2.1 integration is measured in weeks, not quarters.

Poolside's bet on agentic, long-horizon coding workflows is also a signal about what engineering talent looks like in the near term. The engineers who will extract the most value from a model like Laguna S 2.1 are not the ones who can write the best prompts. They are the ones who can design routing architectures, build evaluation harnesses, and make judgment calls about when to trust an agent's output and when to checkpoint it. Those engineers are not generalists. They are AI-native practitioners who understand model behavior, cost tradeoffs, and workflow orchestration as core competencies. The individual team running a monorepo migration might shrink from eight engineers to two, but those two are operating at dramatically higher leverage. The companies that figure out how to hire and retain that profile will move faster on every engineering front, not just the ones touched by Laguna S 2.1. The context window just got a lot bigger. The question is whether your team is structured to use it.