Kimi K3 and Qwen3.8-Max caught the frontier. Now read the licence.

Four frontier releases inside eleven weeks, and two of them shipped their weights.

Model	Released	Weights	Context	List price /M in-out	AA Intelligence Index
Claude Fable 5	9 Jun 2026	closed	1M	$10 / $50	top-2 overall
GPT-5.6 Sol	9 Jul 2026	closed	-	$5 / $30	61
Kimi K3	16 Jul (weights 27 Jul)	Modified MIT	1M	$3 / $15	60
Qwen3.8-Max	3 Aug 2026	open, restricted	-	$2 / $6	58

Kimi K3 is a 2.8T-parameter MoE with roughly 104B active, Kimi Delta Attention and native vision. Qwen3.8-Max is 2.4T total, and notable because it is the first time Alibaba has put a Max-tier model into open weights at all rather than reserving that tier for the paid API.

Three points off the top, in ascending order of how much they should change your decisions.

1. The capability argument is close to over, but the benchmarks are messier than the table

On the aggregate indices these four are inside a few points of each other. Moonshot reports 88.3 on Terminal-Bench for K3, within about half a point of Sol. Alibaba claims 86.1 on OSWorld-Verified for Qwen3.8-Max against 85.0 for Fable 5 and 83.2 for Sol. Fable 5 still sits at or near the top of most published coding and agentic leaderboards, at 80.3% on SWE-Bench Pro.

Treat all of that with suspicion, in both directions. The harness is part of the product now. Moonshot's own comparison table mixes results run under Kimi Code, Claude Code, Codex and mini-SWE-agent, which is not a like-for-like comparison of models, it is a comparison of model-plus-scaffold pairs. K3 in particular is sensitive to whether you preserve thinking history between turns, so the same weights can score materially differently depending on how your agent loop is written.

The practical consequence: a leaderboard delta of two points tells you nothing about your workload. Ours is grounded retrieval over Italian and English regulated documents with tool calls into RPA and internal APIs. I have never once seen a public benchmark predict the ranking on that. Build a hundred-case eval from your own traffic. It is a week of work and it outranks every table above, including mine.

2. For half my clients the closed models were excluded before the benchmarks

This is the part that reframes the whole comparison. When the deployment is a bank running on EU-owned bare metal, air-gapped in some cases, Fable 5 and Sol are not options that lost on price. They are not options. No amount of benchmark lead fixes "the weights live in someone else's datacentre, owned by a company under US jurisdiction."

So the comparison that actually matters to me is not open versus closed. It is Kimi K3 versus Qwen3.8-Max versus GLM versus DeepSeek, on hardware I can point at. What changed this summer is that the price of choosing that constraint dropped close to zero. Two years ago sovereignty cost you a visible capability gap and you had to argue it was worth it. This summer it costs you a couple of points on an aggregate index that I have just argued you should not trust anyway.

That is the real story of Kimi K3 and Qwen3.8-Max, and it is bigger than either model.

3. The licence is now the harder constraint than the weights

Here is what I have been telling clients since 3 August, and it is the thing nobody puts on the leaderboard.

Qwen3.8-Max is open-weight with commercial restrictions attached. Any entity whose aggregate revenue exceeds US$50 million over any consecutive twelve-month period needs a separate commercial licence if it is operating a model-as-a-service or an AI work assistant business. Internal use is fine, provided the model, its outputs and its capabilities are not made available to third parties.

Read that against a real deployment shape. A bank serving its own staff internally: fine. A consultancy like mine embedding those weights into a client-facing platform: that is precisely the restricted case, and both the revenue threshold and the "AI work assistant" wording are close enough to what we build that the answer is a licence conversation, not a git clone.

Kimi K3 shipped under a Modified MIT licence in a native MXFP4 checkpoint, which is a very different procurement conversation.

Neither licence is unreasonable. Alibaba is running a platform play, not a charity, and saying so out loud is more honest than the usual "open source" label on terms that are not OSI-anything. But it means the diligence order has inverted. Licence terms first, serving economics second, benchmarks third. I have watched a team burn three weeks integrating a model their legal department was always going to reject.

4. And $2 per million tokens is not what self-hosting costs

The price column is a trap when you intend to run the weights yourself. You are not buying tokens, you are buying the ability to serve a 2.4T or 2.8T-parameter mixture of experts.

MoE sparsity helps a lot with compute, and MXFP4 checkpoints help a lot with memory, but 104B active parameters still means a serious multi-GPU node before you serve a single user, and the whole expert set has to be resident. This is not a model you drop onto the two A6000s already in the rack. Below real sustained volume, per-token API pricing wins on arithmetic, every time. My ~40% TCO argument has always been conditional on sustained scale, and these two releases push the break-even point up, not down.

Which is why the shape I actually deploy has not changed:

• A mid-size open model as the workhorse, self-hosted. Qwen3.6-27B-class and GLM handle the large majority of production traffic at a fraction of the serving cost.
• A frontier open-weight model for the hard tail, self-hosted where volume justifies the node, rented from an EU provider where it does not.
• Frontier closed models via API for the non-sensitive work only, where the client's own classification policy permits it. Public content, internal tooling, code that touches no customer data.
• Everything behind one OpenAI-compatible gateway, so which model serves which route is a config change and not a migration.

That last point is the one worth internalising from this summer. OpenAI cut Sol's API pricing by more than 20% on 21 August for three months. DeepSeek raised V4 Pro's prices sharply and introduced peak and off-peak billing. Model prices are now a marketing surface that moves monthly in both directions. Any architecture where a price change or a licence change forces a rewrite is the wrong architecture, whichever model happens to be winning this week.

Sources: Artificial Analysis, Kimi K3 benchmark caveats, Claude Fable 5 benchmarks, GPT-5.6 Sol, Qwen3.8-Max licence restrictions.