GPT-6.1 Sol vs Laguna S 2.1: Fifty Times the Output Price for Work Nobody Has Measured Side by Side

Fifty times. That is the ratio between what GPT-6.1 Sol charges for a million output tokens and what Laguna S 2.1 charges — $10.00 against $0.20 — and it is the largest price gap in this entire series of comparisons. GPT-6.1 Sol is Open​AI's 2026-09-29 release, a 1,050,000-token generalist with a published independent score. Laguna S 2.1 is Poolside's 2026-07-21 release, a 118-billion-parameter open-weight mixture-of-experts agentic coder with roughly 8 billion parameters active per token, Apache-adjacent weights on Hugging Face, and a benchmark set that Poolside produced itself. The price gap is real and enormous. Whether it is a gap worth crossing depends on a question neither vendor has answered: has anybody measured the two on the same task?

Vendor-reported figures are labeled as such throughout. Figures from Artificial Analysis are independent measurements carried on our model records. Our own serving numbers are identified as ours. Where a cell is empty, it stays empty.

What each side is

Poolside's Laguna S 2.1 is a specialist, and its design says so. It is a 118-billion-parameter mixture-of-experts model with about 8 billion parameters active on any given token, a thinking / no-thinking toggle, and a context window that reaches 1,000,000 tokens. A smaller sibling, Laguna XS 2.1, sits at 33 billion parameters with roughly 3 billion active and prices at $0.06 input and $0.12 output per million. Weights for the S variant are published on Hugging Face at poolside/Laguna-S-2.1.

The 8-billion-active figure is the number that explains the price and constrains the product at the same time. Inference cost tracks active parameters, not total; a 118B MoE that spends 8B per token is cheap to serve in a way a dense model of the same total size would not be. But an 8B-active budget is also a hard ceiling on what a single forward pass can do, and that ceiling is the reason an independent evaluator's general-purpose index is a more useful yardstick for it than a coding-specific score.

GPT-6.1 Sol is the opposite kind of object. One snapshot identifier, gpt-6.1-sol, no dated variant, a 1,050,000-token context window, 128,000 maximum output, text, image and file input, reasoning always on across five effort levels defaulting to medium, and a knowledge cutoff of April 30, 2026. It does not accept temperature or top_p; the effort dial replaces them. It is closed, priced per token, and not fine-tunable.

The dimensions where they actually differ:

• Input price — $0.10 per million for Laguna S 2.1 against $2.00 for GPT-6.1 Sol. Twenty times.

• Output price — $0.20 against $10.00. Fifty times.

• Reasoning control — a binary thinking / no-thinking toggle on Laguna S 2.1 against five graded effort levels on GPT-6.1 Sol.

• Weights — downloadable and self-hostable for Laguna S 2.1; closed for GPT-6.1 Sol.

• Context window — up to 1,000,000 tokens for Laguna S 2.1 against 1,050,000 for GPT-6.1 Sol. Effectively level.

The benchmark rows, with both sets of labels attached

Poolside's published numbers for Laguna S 2.1, all vendor-reported and none reproduced by a third party as far as we can find:

• SWE-Bench Multilingual — 78.5%.

• Terminal-Bench 2.1 — 70.2%.

• DeepSWE v1.1 — 40.4%, against 9.0% for DeepSeek-V4-Pro-Max, which Poolside frames as roughly a sixth of the active parameters producing over four times the score.

GPT-6.1 Sol's independent record, all Artificial Analysis:

• AA Intelligence Index — 51.8, rank 10 of 147, 93.2nd percentile.

• Humanity's Last Exam — 52.9%.

• Long-Context Recall — 83%.

• SciCode — 54.2%.

• Terminal-Bench 4.0 — 56.06%.

There is one place these two lists appear to meet, and it is a trap. Laguna S 2.1 reports 70.2% on Terminal-Bench 2.1. GPT-6.1 Sol is not scored on Terminal-Bench 2.1 in our records at all — its terminal figure is on version 4.0, a later and harder revision of the harness. Putting 70.2% next to 56.06% and calling it a twelve-point Laguna win would be comparing two different tests, and it is exactly the mistake this blog exists to avoid. The two models have no shared benchmark row, which means the performance half of this comparison is, at present, unmeasured.

Two further gaps are worth naming. Laguna S 2.1 has no Artificial Analysis Intelligence Index score and no independent SWE-bench Verified result. For a model published in July, that is a long time to go without an outside measurement, and it is the single largest omission in anything Poolside has published about it.

What the thinking toggle actually buys

Poolside's own numbers put a number on the toggle, and it is the most interesting thing in the release. Turning thinking on lifts DeepSWE v1.1 from 16.5% to 40.4% — a factor of roughly 2.4 — and Terminal-Bench 2.1 from 60.4% to 70.2%. Both figures are the vendor's, unreproduced.

Read against the price, that matters in a way the headline rate card does not capture. Thinking mode consumes output tokens, and output tokens are what you are paying for. A workload that runs Laguna S 2.1 with thinking on will burn materially more output than the same workload with it off — which is precisely what you would expect from a 2.4× gain on a hard agentic benchmark. So the honest comparison against GPT-6.1 Sol is not $0.20 against $10.00 on raw output. It is the thinking-enabled cost per completed task against the effort-priced cost per completed task.

That comparison still favours Laguna S 2.1 enormously, and it is worth doing the arithmetic rather than asserting it. Take a workload producing a billion output tokens a month against three billion input tokens:

• GPT-6.1 Sol — 3 billion input at $2.00 per million is $6,000; 1 billion output at $10.00 per million is $10,000. Sixteen thousand dollars a month.

• Laguna S 2.1, thinking off — $300 of input and $200 of output. Five hundred dollars a month.

• Laguna S 2.1, thinking on at roughly 2.4× output — $300 of input and about $480 of output. Under eight hundred dollars a month.

Even with the thinking penalty, that is a difference of roughly $15,000 a month at this volume. If your workload is one Laguna S 2.1 can actually do, the price question is not close, and no plausible amount of retry overhead closes it. Which pushes the entire decision onto capability — and capability is the half of this comparison with no shared measurement.

Where the fifty-times gap stops being the point

Three things pull the ratio back toward reality, and each is worth stating plainly because the number above is seductive in a way that has cost teams real money.

The first is that an 8-billion-active-parameter coder and a frontier generalist are not substitutes. Laguna S 2.1 is built for agentic software engineering — repository edits, terminal tasks, multilingual code — and Poolside markets it as such. It is not positioned for document-heavy professional work, computer use, or the long-context recall tasks where GPT-6.1 Sol carries an 83% independent Long-Context Recall score and a 1,050,000-token window. On those tasks the honest comparison is not "the cheap model wins," it is "the cheap model was not built for this." The $15,000 a month in the worked example is only a saving if the cheaper model's output is usable, and for a class of workloads nobody has demonstrated that it is.

The second is throughput and the operational surface. GPT-6.1 Sol reports p50 latency of 2,442 milliseconds and 95.57 output tokens per second over seven days of our own playground traffic, with a 6.36% error rate. We have no comparable serving panel for Laguna S 2.1 — it is not routed on OrcaRouter — so there is no number to put beside it, and the self-hosted path's latency is a function of whatever hardware you buy. On eight H100s you will not match a hosted frontier endpoint's throughput on the same concurrency; you will trade latency for unit cost, and how that trade lands depends on whether your workload is latency-sensitive.

The third is that the self-hosted path is only cheap at volume. The same $0.10 and $0.20 rates that make the arithmetic above look like a rout are also what an inference provider charging near cost would quote — which is the argument for renting the open model rather than buying hardware for it, at any volume short of the one that justifies the GPUs.

How to actually make the call

This is the rare pairing in which the right next step is not a routing change, because the two models are reached in entirely different ways.

GPT-6.1 Sol is on OrcaRouter as openai/gpt-6.1-sol at Open​AI's list price with 0% markup — $2.00 input, $10.00 output, cached input at $0.10, and the above-272,000-token tier applied exactly as Open​AI sets it. Because the routing layer passes the provider's price through rather than marking it up, a pricing change on Open​AI's side is live here the same day it takes effect. Where a route is live, automatic failover sits underneath it — which is the cheap way to find out whether a cheaper model can hold a production path, since the expensive model stays one routing rule away.

Laguna S 2.1 is not on OrcaRouter and we do not host it. It is reachable through Poolside's own API and, because the weights are published, by serving it yourself — which for this model is the more interesting option, since a 118B MoE with 8B active is genuinely small enough to run on hardware a mid-sized team can own. The right first move with this pairing is therefore not a price comparison at all: take twenty of your real coding tasks, run them through your existing GPT-6.1 Sol path, then run the same twenty through Laguna S 2.1 with thinking on and off. You now have the shared benchmark row that neither vendor has published, on the only task set that matters, and the fifty-times price difference becomes a number you either collect or decline with evidence.

The verdict this pairing earns, on the evidence that currently exists: Laguna S 2.1 is the most interesting cost-per-capability proposition in agentic coding right now, with vendor numbers that would make it the default choice if they were independently confirmed, and no independent confirmation of any of them. GPT-6.1 Sol costs fifty times as much on output and is the only one of the two whose capability anyone outside the company has measured on a shared instrument. Buy the cheap model's weights and test them, because unlike most open releases this one is small enough to test properly — and treat the fifty-times figure as the reason to run that test rather than as the answer to it.