Three labs shipped flagship models in fifteen days: GPT-5.6 Sol on July 9, Kimi K3 on July 16, Claude Opus 5 on July 24
Opus 5 leads SWE-bench Pro 79.2 to 64.6 over Sol, and ARC-AGI-3 30.2 to 7.8
Sol holds Terminal-Bench 2.1 at 91.9 percent in its top mode and still takes DeepSWE 1.1 and HealthBench Professional
Kimi K3 is a 2.8 trillion parameter open-weight model at 3 and 15 per million tokens, roughly 40 percent under Opus 5 on input
Fifteen days. That is the gap between OpenAI making GPT-5.6 Sol generally available and Anthropic shipping Claude Opus 5, with Moonshot dropping a 2.8 trillion parameter open-weight model in the middle of it. I wrote a frontier check like this in June and most of it is already out of date, so here is where the three current flagships actually stand.
Model
Lab
GA
Context
Per million (in / out) GPT-5.6 Sol
OpenAI
2026-07-09 1.05M
5 / 30
Kimi K3
Moonshot AI
2026-07-16 1M
3 / 15
Claude Opus 5
Anthropic
2026-07-24 1M
5 / 25
The specs have converged to the point where they barely differentiate anything. All three sit at or just above a million tokens of context. All three cap output around 128k. The input prices are within a factor of two of each other. Two years ago a context window was a headline; now it is table stakes, and the interesting differences have moved entirely into behavior under load.
Two timing details that get flattened in the coverage. GPT-5.6 Sol was previewed on June 26 and only became generally available on July 9, so some of the earliest benchmark tables were run against a preview build. And Sol is the top of a three-model family alongside Terra and Luna, spanning roughly 1 to 30 per million tokens depending on tier. Comparing Opus 5 to "GPT-5.6" without saying which one is close to meaningless, which is a large share of the comparisons currently circulating.
One structural note on Kimi K3, because the parameter count gets quoted carelessly. It is a mixture-of-experts model with 896 expert subnetworks, of which 16 activate for any given token. The 2.8 trillion figure is total capacity, not what runs on each token. That is why a model of that nominal size can be served at 3 and 15 per million tokens rather than at a price nobody would pay.
The separation is widest on exactly the workload I care about, which is a model running inside an agent loop that has to plan, edit, run something, read the failure, and try again.
On SWE-bench Pro, real GitHub issues graded by the repository's own tests, Opus 5 posts 79.2 percent against 64.6 for Sol. That is a 14.6 point spread on the single most-quoted coding row in the industry, and it is the biggest gap between the two models anywhere. On Frontier-Bench v0.1 it is 43.3 to 34.4, roughly a 26 percent relative lead on multi-step tool use.
The outlier is ARC-AGI-3, where Opus 5 scores 30.2 against 7.8 for the next model on the public list. A near four-fold gap on a reasoning benchmark is unusual enough that I would want independent replication before building anything on it. ARC-style tests are also the ones most exposed to training-set contamination arguments, so treat it as a strong signal rather than a settled fact.
Anthropic also claims Opus 5 beats Fable 5's best OSWorld 2.0 computer-use result at just over a third of the cost per task, which matters more for the price question than the leaderboard one. I broke the in-family numbers down separately in Claude Opus 5 Benchmarks: What the Numbers Actually Show, including the rows where the published figures do not agree with each other.
Three places, and I am not going to bury them.
Sol owns Terminal-Bench 2.1 at 91.9 percent in its top mode and 88.8 percent at base. That is the highest agentic-terminal number anyone has published, and Anthropic did not beat it last week. Sol also takes DeepSWE 1.1 and HealthBench Professional, and on broad reasoning aggregates the two models are close enough that the ordering flips depending on whose table you read (92.5 to 90.4 in Sol's favor on one widely cited set).
It is also worth weighting the independent indices differently from the vendor tables. Artificial Analysis, which runs its own harness rather than reprinting launch charts, puts GPT-5.6 Sol at 59 on its Intelligence Index against a median of 32 for reasoning models in a similar price band. That is not a coding-specific measure and it will not settle an agentic argument, but it is one of the few numbers in this comparison produced by someone with no model to sell. When an independent index and a vendor chart disagree, I go with the index and assume the vendor picked its strongest framing, which is what I would do in their position too.
Kimi K3 leads the Frontend Code Arena outright, ahead of Fable 5, which is the result that made it news the week it landed. I covered that one when it happened in The AI Week: Kimi K3 Tops the Code Arena, so I will not re-litigate it here beyond noting it has survived a further week of scrutiny.
There is also a row I deliberately will not score. GDPval-AA, the human-graded knowledge work benchmark, has Kimi K3 placing third behind the two American flagships, which is a genuinely notable result for an open-weight model. But the published Elo values for that benchmark disagree across sources by more than a hundred points depending on which version and which effort tier is being quoted. The ordering looks stable. The numbers do not, so I am giving you the ordering only.
Moonshot has said full weights land by July 27. If that holds, the practical situation is that a model competitive with the American frontier on several public benchmarks becomes something you can run on your own hardware, fine-tune, and audit.
Open weight is not the same as open source, and the distinction matters if you are making a procurement decision on it. Published weights mean you can download, run, inspect, and fine-tune the model. They do not necessarily mean you get the training data, the training code, or a licence that permits every commercial use. Read the licence before you build a product on the assumption, because "open" in model releases has covered at least four different arrangements in the last two years.
The pricing detail that goes underquoted is the cache-hit input rate of 0.30 per million tokens, with no long-context surcharge. For a workload that replays a large stable prefix on every call, which describes most agent harnesses, the effective cost sits far below the headline input price. That is a bigger practical lever than the sticker difference and it is the number I would model against before switching anything.
For a one-person studio the honest answer is that this changes less than the discourse suggests, at least today. Serving a 2.8 trillion parameter mixture-of-experts model yourself is not a laptop project, and at 3 and 15 per million tokens through a hosted endpoint the price advantage over Opus 5 is real (about 40 percent on input, 40 percent on output) but not category-changing for my volume. What it changes is the floor. Every closed model now has to justify its price against something with published weights that scores in the same range, and that pressure shows up in pricing long before it shows up in benchmarks. The second-order effect is the one I would actually watch. Anthropic held Opus 5 at Opus 4.8's exact price while closing most of the gap to its own top tier. Nobody does that in a market with no competition.
If the work is agentic coding, long tool-use loops, or anything where a silent wrong answer costs real time, Opus 5 leads on the rows that describe that work, by a wide margin on SWE-bench Pro and Frontier-Bench. If the work looks like Terminal-Bench, or sits in DeepSWE or clinical territory, GPT-5.6 Sol is still the model to beat and it is cheaper to be honest about that than to pretend the sweep was clean. If price per token is the binding constraint and you can live with a hosted Chinese endpoint or your own infrastructure, Kimi K3 is now a serious answer rather than an asterisk. The comparison I wrote in June, Claude Fable 5 vs GPT-5.5 vs Gemini 3.1 Pro, had a shelf life of about six weeks. This one will not last longer. I would pick per task rather than per quarter, keep two providers wired up, and re-check when the next flagship lands, which on the current cadence is roughly a fortnight away.