GPT-6 Sol vs Kimi K3: Three Assumptions About Open Weights, Tested Against One Bill

Buying open weights is usually a bet on three things: that they will be cheaper, that they give you control, and that they are insurance against a vendor changing its mind. Kimi K3, the 2.8-trillion-parameter model Moonshot AI released under open weights on July 27, 2026, and GPT-6 Sol, the closed model shipped on September 22, 2026, are a clean test case for all three, because the two models are close enough on the neutral board that the assumptions have to carry the argument on their own.

Here is how they fare. K3 is not cheaper per task — it is roughly twice as expensive, at $3.00 and $15.00 per million tokens against Sol's $2.00 and $10.00. The control is real but has a hardware floor that most teams will not cross. And the insurance is the one assumption that holds up completely, which makes it the only one worth paying for. What follows is the arithmetic behind each.

The two models, briefly

Kimi K3 is a sparse mixture-of-experts model with 2.8 trillion total parameters and roughly 104 billion — about 3.7% — active per token, published under a Modified MIT license with a 1,048,576-token context window. Moonshot's architectural claims are specific: Kimi Delta Attention, a hybrid linear-attention scheme the company says is up to 6.3× faster at long-context decoding, plus Attention Residuals and a Stable LatentMoE framework. It takes text and image in and returns text. The undistilled model exceeds 1 TB in size and a production deployment is described in terms of a 64-accelerator floor.

GPT-6 Sol is closed, single-identifier, and priced at a flat $2.00 and $10.00 with a $0.20 cached read and $2.50 cache writes, repricing the whole request above 272,000 input tokens to $4.00 and $15.00. It is text and image in, text out, with a 1,050,000-token window, a 922,000-token maximum input, a 128,000-token output ceiling and an April 20, 2026 knowledge cutoff.

The context windows are, for practical purposes, the same number — a 0.1% difference. That removes the most common reason to prefer one, and it means the decision rests entirely on the three assumptions.

Assumption one: open weights are cheaper. They are not.

• Input — GPT-6 Sol $2.00 per million tokens vs Kimi K3 $3.00 per million tokens

• Output — GPT-6 Sol $10.00 per million tokens vs Kimi K3 $15.00 per million tokens

• Cached input — GPT-6 Sol $0.20 per million tokens vs Kimi K3 $0.30 per million tokens; both are a tenth of the uncached input rate

• Intelligence Index, independent — GPT-6 Sol 48 at max effort vs Kimi K3 44 at max effort

• Cost per Index task, independent — GPT-6 Sol $1.06 vs Kimi K3 $2.00

• Output tokens for the index run — GPT-6 Sol 77M vs Kimi K3 160M

• Context window — GPT-6 Sol 1,050,000 tokens vs Kimi K3 1,048,576 tokens

• Weights — GPT-6 Sol closed vs Kimi K3 open, downloadable and self-hostable

• License — GPT-6 Sol not applicable vs Kimi K3 Modified MIT

• Total parameters — GPT-6 Sol undisclosed vs Kimi K3 2,800 billion, of which 104 billion are active per token

Sol is cheaper on every line of the rate card, and the gap is widest exactly where agentic workloads spend their money. The independent board agrees on direction and roughly on size: Sol costs about half as much per completed task, at four index points higher. K3 generated a little over twice as many output tokens to run the same index.

One nuance keeps this from being a rout. Moonshot claims K3 emits 21% fewer output tokens than its predecessor, and the architectural work — the attention scheme, the residual design — is aimed squarely at long-context decoding cost. On a workload dominated by very long inputs, K3's efficiency features may close part of the gap the index run shows. Nobody has published that measurement on a comparable harness, so treat it as a vendor claim with a plausible mechanism, not a result.

Assumption two: open weights give you control. Up to a hardware floor.

The control is genuine and worth stating precisely, because "open weights" is often used loosely. A Modified MIT license on a downloadable checkpoint means you can run the model on your own hardware, fine-tune it, pin a specific revision, and keep inference inside a boundary you control. None of that is available with GPT-6 Sol at any price.

What it does not mean is that self-hosting is a substitute for the API on cost. The published figures put the full-precision checkpoint at roughly 4.9 TB, or around 397 GB at 1-bit quantization, with a deployment floor in the 410 GB of combined memory range and production deployments described in terms of 64 accelerators. A team already running that class of cluster has a real option here. A team that does not is comparing an API bill against a capital purchase, and the API bill wins by a wide margin.

The honest version of the control argument is narrower than it is usually stated: open weights give you the ability to leave, not the ability to run cheaply. Those are different things, and only one of them is available to most teams.

Assumption three: open weights are insurance. This one holds.

The third assumption survives every number above, and it is the reason K3 has a market at all. Moonshot can be acquired, can change its license for future versions, can deprecate K3, can raise its price, can suspend subscriptions — and it did suspend new subscriptions within 48 hours of release to protect service quality for existing paying customers, which is a live demonstration that API access is a policy decision rather than a property.

If K3 is deprecated, the weights you downloaded still run. If Moonshot triples the API price, your self-hosted deployment is unaffected. If you need to demonstrate to an auditor that your inference stack is frozen at a known revision, a checkpoint gives you that and an API does not. None of this shows up on a cost-per-task table, and all of it is real.

The question is what it is worth. On the numbers above you are paying roughly double per completed task for that insurance, plus four index points of capability. For a workload where an API deprecation is a business interruption, that is cheap. For a workload where you would simply switch models, it is a premium for a hedge you will never exercise.

Both are on one key if you want them to be

Kimi K3 is on OrcaRouter's catalogue at Moonshot's list price, under the pass-through model that puts a Moonshot rate change live on our side the same day rather than after a reseller renegotiates. GPT-6 Sol is not in our catalogue as of this writing and is reachable through OpenAI's own API.

That combination is worth more than it sounds for this particular decision, because the two models belong to different failure modes. The reason to run K3 is not that it is cheaper — it is not — but that it is a hedge, and a hedge you cannot switch to quickly is not much of a hedge. Holding K3 behind the same endpoint as everything else, with automatic failover and a routing rule that can be changed in configuration, is what converts "we have an open-weight fallback" from a slide in a deck into something that works at 3am.

What each is actually for

• If you want the lowest cost per completed task on general work — GPT-6 Sol, and the margin is about 2×.

• If you want the highest capability on the neutral board of the two — GPT-6 Sol by four points.

• If you already operate a cluster in the 400 GB-plus memory class and need inference inside your own boundary — K3, and the calculus changes entirely, because the marginal cost of the weights is not the same as the cost of the hardware.

• If your real requirement is that your model cannot be taken away from you — K3, and this is the only argument in the comparison that Sol cannot answer.

• If you are choosing on price per token because K3 looks like the cheaper Chinese model — check the token count first. At 160 million output tokens for the index run against Sol's 77 million, the cheaper rate is buying more tokens, not a smaller bill.