GPT-6.1 Sol vs Kimi K3.1 Pro: Same $2/$10, Three Index Points, and a 2.9x Gap in What a Task Costs

Kimi K3.1 Pro and GPT-6.1 Sol have the same price tag, which is the least interesting thing about them. Moonshot released K3.1 Pro on October 23, 2026 at $2.00 per million input tokens and $10.00 per million output; OpenAI released Sol on September 29 at the identical pair. On Artificial Analysis, K3.1 Pro scores 55 and sits fourth of 221 models; Sol scores 52. The models are three points apart in capability and almost three times apart in what it costs to finish a task — $2.11 for K3.1 Pro against $0.72 for Sol. That inversion, not the price tag, is the decision.

The measurements below come from Artificial Analysis, which runs both models under a published configuration and accounts for the tokens each one spends, and from the two vendors' own documentation and launch claims, read on October 27, 2026. Vendor figures are labelled as such. The distinction matters more than usual here, because the two companies are making opposite bets about how a model should spend tokens.

The price is the same. The invoice is not.

Per-token pricing is a unit price, and a unit price only becomes a bill when you multiply it by consumption. The two models consume very differently. Artificial Analysis's cost-per-task accounting puts K3.1 Pro at $2.11 to complete one Intelligence Index task and Sol at $0.72 — a 2.9x difference on identical headline rates. The cause is output volume. K3.1 Pro is an always-reasoning model that emits substantially more tokens per answer, and output is the expensive half of the card at $10.00 per million.

Moonshot's own release material acknowledges the shape of this. The model is a post-trained version of Kimi K3 with heavily expanded reinforcement-learning compute, and the company describes the gains as coming from longer, more deliberate reasoning rather than from a larger base. Longer reasoning is exactly what produces more output tokens. The capability gain and the cost increase are the same fact measured two ways.

• Input / output — $2.00 / $10.00 per million on both models.

• Cached input — Kimi K3.1 Pro $0.20 at a 90% discount; GPT-6.1 Sol $0.10 at 95%.

• Cost per index task — $2.11 vs $0.72 on Artificial Analysis v4.3.2.

• Independent index — 55 vs 52; K3.1 Pro ranks fourth of 221, Sol just behind.

• Output speed — 47.9 tokens per second vs 69.2, both Artificial Analysis measurements.

• Context and output cap — 1,048,576 input for K3.1 Pro; 1,050,000 input and 128,000 output for Sol.

• Architecture — 2.8 trillion total parameters with 104 billion active for K3.1 Pro; OpenAI does not publish Sol's parameter count.

• Weights — K3.1 Pro is proprietary today, with Moonshot promising open weights by November 4; Sol is proprietary with no opening announced.

What three index points buy

A three-point gap on a composite is meaningful but not decisive, and the component scores are where the shape of the difference lives. Artificial Analysis lists K3.1 Pro at 67 on Terminal-Bench 4.0, the highest of any model it has measured, against 56.1 for Sol. On GDPval-AA, a professional-work evaluation, K3.1 Pro sits at 1,846 Elo, third overall and 312 points above its own predecessor. On AutomationBench-AA it improved from 58% to 73%. On Humanity's Last Exam it moved from 47% to 52%. The hallucination rate on AA-Omniscience fell from 53% to 44%.

Those are the strongest independent numbers either model has. Sol's independent profile is flatter: a 52 composite with no single evaluation where it leads the field, and a cost-per-task figure that is the best in its price band. The honest reading is that K3.1 Pro is the more capable model on several agentic and professional-work evaluations, and Sol is the more efficient one. Neither statement is a verdict, because capability only converts to value at the volume you run.

Moonshot's own launch claims go further than the independent board, and should be read as claims. The company reported top scores on six of fourteen tests it ran, including SWE Marathon at 51.7 and HLE-with-tools at 70.0. Its Terminal-Bench 4.0 figure is 65.7%, slightly below the 67% Artificial Analysis measured — a rare case where the vendor's number is the conservative one, which is worth noting precisely because it is unusual. The weak spot is Terminal-Bench Science 0.1 at 38.1, against 68.1 for GPT-6 Astra. K3.1 Pro is not the model for hard scientific reasoning, and Moonshot does not claim it is.

The speed and verbosity bill

Artificial Analysis measures K3.1 Pro at 47.9 output tokens per second and calls it notably slow; Sol runs at 69.2. Combined with the output-volume difference, that means a long agent loop on K3.1 Pro takes both more wall-clock time and more money per step. For interactive coding, where a developer waits on each turn, the latency difference is the one users will complain about first. For batch evaluation, where nothing waits, only the cost matters.

There is a faster tier. Moonshot sells a Highspeed variant at $4.00 input and $20.00 output with cache-hit pricing unchanged, which doubles the token rate and doubles the price. That is the same trade Sol offers with its Fast mode at 2x. In both cases the fast tier is for latency-sensitive work, not for volume, and the arithmetic rarely favours it for batch jobs.

November 4, and what open weights actually change

Moonshot says full weights will be published by November 4, 2026. If that happens, K3.1 Pro would become the highest-scoring open-weight model on the Artificial Analysis index, narrowing the gap between the best open and the best proprietary models from roughly twelve points to three. That is a genuine event for anyone who needs to self-host, fine-tune, or run inference inside a compliance boundary.

It is not an event for most API callers. Open weights change who can run the model and under what terms; they do not change the hosted price, the speed, or the capability. A team that calls an API and has no data-residency constraint gets nothing from the weights release except a second sourcing option and, eventually, more providers competing on the same checkpoint. That competition is real and usually shows up as price pressure, but it is a months-long effect, not a November one.

There is a second, more specific caveat in Moonshot's release. The company says it deliberately removed cyber and exploit-related training data to avoid extending the open frontier of offensive cyber capability, and that a checkpoint which scored 90.5% on xdev-bench-hard was reduced to 81% as a result. That is a safety decision with a measurable capability cost, disclosed by the vendor. If your work is security tooling, it is the most important sentence in the launch material — more important than the index score.

Three workloads, three answers

• High-volume batch coding, no latency requirement — GPT-6.1 Sol. At $0.72 per index task against $2.11, the efficiency gap dominates a three-point capability difference unless your task is specifically one where K3.1 Pro's agentic scores translate into fewer retries. Measure the retry rate before assuming they do.

• Long-horizon agentic work where a wrong answer is expensive — Kimi K3.1 Pro. The Terminal-Bench 4.0 lead and the GDPval-AA placement are the closest thing either model has to evidence of sustained multi-step competence, and the per-task premium is small relative to the cost of a failed trajectory.

• Interactive coding where a human waits — GPT-6.1 Sol, on speed. 69.2 tokens per second against 47.9 is a visible difference in an editor, and the cost advantage is a bonus rather than the reason.

• Scientific or mathematical reasoning — neither, at the frontier. K3.1 Pro's 38.1 on Terminal-Bench Science 0.1 is a documented weakness, and Sol's 57.0 on the same suite is respectable but well below the 68.1 OpenAI reports for GPT-6 Astra. This is the one workload where the tier above both models is the correct answer.

Running both is cheaper than choosing wrong

Both models are on OrcaRouter on one key at 0% markup over each provider's list price, with Sol at $2.00 and $10.00 and K3.1 Pro at the same pair, cached input passed through at $0.10 and $0.20 respectively. That makes the decision reversible: a workload can be benchmarked against both on the same traced traffic, and the loser removed with a routing change rather than a contract negotiation.

Model fusion is the one feature that earns its place on this particular pair rather than as a generic plug. The two models fail in different directions — Sol is terse and fast, K3.1 Pro is deliberate and stronger on long agentic trajectories — so a panel that asks both and reconciles their answers is a reasonable design for high-stakes calls where the 2.9x cost difference is immaterial against the cost of being wrong. For everything else, routing by workload is the cheaper pattern: K3.1 Pro for the agentic path, Sol for the high-volume path, one endpoint for both.

The ranking that is not the index

If the question is which model is better, K3.1 Pro wins by three points and leads on the agentic evaluations that matter for long-running work. If the question is which model to put behind a high-volume production path this month, Sol wins on cost per completed task by a factor of 2.9 and on speed by a factor of 1.4, and the three-point capability gap does not close either advantage.

The one date worth diarising is November 4. If the weights land, K3.1 Pro becomes the reference open-weight model and the calculus for self-hosted teams changes. If they slip — and Moonshot has given a date, which is more than most — then the calculus stays exactly where it is today: the same sticker price, a different bill, and a choice that should be made on measured task cost rather than on the number in the price column.