GPT-6.1 Sol vs Kimi K4: The Model in the Title Has Never Been Announced

Kimi K4 is not merely an unannounced model. It is a name with no model behind it at all. Moonshot AI has never published a K4 model card, a K4 repository, a K4 API identifier or a K4 price row, and its own research index stops at Kimi K3. Artificial Analysis has no K4 page. The name entered circulation through a leaked internal banner in July, and everything since has been extrapolation from Kimi K3, which is a real and rather good model. GPT-6.1 Sol, by contrast, has been generally available on the OpenAI API since 29 September 2026 at $2.00 per million input tokens and $10.00 per million output. So the comparison that can be written honestly is Sol against the Kimi model that exists — Kimi K3.1 Pro — with an account of what K4 would have to be for waiting to make sense.

A name with no model behind it

The evidence trail is worth setting out, because the volume of K4 content online is wildly out of proportion to the evidence.

• July 2026 — a leaked internal celebration banner from Moonshot contains the phrase "push K4 to the absolute limit". This is the origin of the name.

• 29 July 2026 — The Information reports Moonshot seeking additional Nvidia Blackwell-class capacity to train a model larger than K3. The report names no model, no parameter count and no date. It is a chip-procurement story.

• 16 July 2026 — Kimi K3 ships: 2.8 trillion total parameters, 104 billion active, a sparse mixture-of-experts with 896 routed experts and 16 active, a 1,048,576-token context and a vision encoder. Weights were published in late July.

• 11 October 2026 — Chinese-language outlets report a 5–6 trillion parameter model in training, "probably" K4, attributed to no named source.

• 15 October 2026 — an unconfirmed claim that a codenamed "x4" flagship finished pre-training and is mid-to-late post-training.

• 23 October 2026 — Moonshot ships Kimi K3.1 Pro instead.

Every K4 benchmark figure, context length and price you have seen is arithmetic performed on K3's numbers. Prediction markets put K4 at roughly a 6% chance of shipping by 31 December 2026 and about 17% by 31 March 2027. There is also an unresolved context worth knowing about: allegations that Moonshot obtained GB300 access via Thailand and distilled from Anthropic's models, with a threatened Commerce Entity List action. That is background, not a K4 spec, but it is the kind of thing that moves a launch date.

The Kimi model you can actually compare

Kimi K3.1 Pro released on 23 October 2026 and is listed in our catalogue on 24 October as kimi/kimi-k3.1-pro. It costs $2.00 per million input tokens and $10.00 per million output — exactly GPT-6.1 Sol's headline rate — with cached input at $0.20 rather than Sol's $0.10. It carries a 1,048,576-token context and, unusually, a maximum output of 943,718 tokens. That output ceiling is not a typo, and it is the single most interesting number in this comparison.

Here is the pair on Artificial Analysis Intelligence Index v5.0.1, both at maximum published effort.

• Intelligence Index v5.0.1 — Sol 55 vs K3.1 Pro 46, a nine-point gap

• Cost per completed index task — Sol $1.27 vs K3.1 Pro $3.68

• Output tokens on the index run — Sol 63.7 million vs K3.1 Pro 113 million

• Output speed — Sol 56 tokens/sec vs K3.1 Pro 33 tokens/sec

• Context window — Sol 1,050,000 vs K3.1 Pro 1,048,576, effectively identical

• Maximum output in one response — Sol 128,000 vs K3.1 Pro 943,718

• Input / output price per 1M — $2.00 / $10.00 on both

• Cached input per 1M — Sol $0.10 vs K3.1 Pro $0.20

• Reasoning control — Sol runs reasoning always on, low through max with a medium default and no off switch; K3.1 Pro defaults to thinking on with selectable effort of none, low, high or max

• Input modalities — Sol takes text, images and files; K3.1 Pro takes text, images and video

• Weights — Sol closed; K3.1 Pro's checkpoint was published on 2 November 2026

The identical headline rate is a good reminder that list price is not cost. K3.1 Pro burned 1.8 times Sol's output tokens to complete the same index and cost 2.9 times as much per task. It is also the slower of the two by a wide margin at 33 tokens per second, which matters more than usual because that is steady-state generation on a model that thinks before it writes.

Where K3.1 Pro is the only choice

One number deserves its own section. A 943,718-token maximum output is roughly seven times GPT-6.1 Sol's 128,000. If your job is to read a long corpus and emit a long structured artefact in a single call — a full codebase migration, a statutory review, a book-length normalisation task — Sol will truncate at 128K and K3.1 Pro will not. That is not a benchmark difference you can tune away with effort levels; it is a hard ceiling on one side.

The cost of using it is real. At $10.00 per million output tokens, filling even half that ceiling is about $4.70 for the output alone, before any reasoning tokens. But for a class of batch jobs the alternative is not a cheaper model, it is a chunking pipeline with its own failure modes. If that is your workload, K3.1 Pro is not a compromise; it is the model that fits.

The open-weights reason to wait has already been met

Strip away the speculation and there is exactly one structural thing a K4 could deliver that GPT-6.1 Sol cannot: a downloadable checkpoint. That is the reason people watch Moonshot, and it is the reason this comparison has a live third option that the headline misses. Kimi K3.1 Pro's weights were published on 2 November 2026 under a custom licence. Three days ago, not in some future K4 release.

The repository is a 2.78-trillion-parameter checkpoint. At 8-bit precision that is roughly 2.7 terabytes of weights before any KV cache or serving overhead, and the k-quantised artefacts in the repo are the same order of magnitude. This is a multi-node deployment with serious interconnect, not a single-GPU download. The active-parameter count — 104 billion on Kimi K3 — is what determines how fast it runs once the memory is paid for, which is why a model this large can be served at a usable speed on the right cluster and not at all on the wrong one.

So the trade is explicit. Open weights remove a vendor dependency and replace it with a capital expense and an operations team. At $2.00 per million input tokens and $10.00 per million output, Kimi K3.1 Pro's hosted price is low enough that self-hosting is unlikely to win on compute alone; what it wins on is data residency, air-gapped deployment and the ability to fine-tune on traffic you cannot send to anyone. If those are requirements, they are met today by K3.1 Pro. If they are not, the open weights are a nice property you will never use.

What K4 would add on top, on the evidence available, is a larger parameter count and a better score. That is a reason to upgrade a model. It is not a reason to defer a deployment, and it is certainly not the only way to get weights you can hold.

Where the two real models land for most teams

For tool-calling agents, document QA and support flows, GPT-6.1 Sol is the better default on every measured axis except output length: higher index score, lower cost per completed task, 70% faster generation, cheaper cache reads. It has been in production long enough to have a real telemetry profile — on our seven-day figures, a 937 ms median time to first token, a 5.7% error rate and 39.5 tokens per second of output, across 382 million tokens of traffic.

For long-output batch work, Kimi K3.1 Pro is the model to reach for, and the 943K ceiling is the reason. For teams that need to own the weights, Kimi K3.1 Pro is the answer today: the checkpoint exists, it is downloadable, and it is three days old. That is the whole practical content of the K4 question, answered by a model that shipped in October.

The routing point is not decorative here. These two models have the same list price and opposite shapes: one is fast, cheap per task and capped at 128K output; the other is slow, expensive per task and effectively uncapped. That is a textbook per-request split rather than an either-or. Both sit behind one key at provider list price with no markup, so a routing rule can send long-output batch jobs to K3.1 Pro and the interactive agent loop to Sol without a second contract, or fuse them where a single answer benefits from both. The layer is what lets you defer a K4 decision without deferring the work.

So the answer to "GPT-6.1 Sol or Kimi K4" is that one of those is a decision and the other is a rumour with a parameter count attached. If K4 ships in 2027 with open weights and a score above 46, it will be worth a serious look, and the way to be ready for that is to have a stack where swapping a model is a routing change. Waiting for it is not a strategy; being able to adopt it in an afternoon is.