Stop Using Claude for Everything

The White House is threatening sanctions over a model I am about to tell you to route most of your work to. Those two facts should not coexist comfortably, and this post is about why they do.

The model is Kimi K3, from Moonshot. Anthropic says it has evidence that Moonshot distilled Claude to build it. Distillation, if the term is new to you, is training your model on another model's outputs, a free ride on somebody else's research bill. The accusation is serious enough that the Treasury is talking about sanctions, Moonshot is publicly denying it, and the two labs that never agree on anything, OpenAI and Anthropic, are suddenly aligned in warning policymakers about open-weight models. I will get to all of that. But the geopolitics is the second most interesting thing here. The most interesting thing is what happened when I stopped reading the discourse and started reading my own token ledger.

The era of picking one model and using it for everything is ending. That is the actual story, and K3 is just the release that makes it undeniable.

Half the price, most of the capability

Start with the boring numbers. On DeepSWE, the benchmark I check first for backend work, K3 Max lands within two points of the frontier: 69 against GPT 5.6 Sol at 71, with Claude Fable 5 in the same neighborhood. The costs are not in the same neighborhood. K3 Max runs the benchmark at 4.65 dollars against 9.18 for Fable 5 at high effort. Half the price for roughly the capability of the best model in the world is already a headline, and it is the least impressive number in this post.

I have written before about what benchmarks actually measure, so I will not relitigate it here. Leaderboards reward obedience, obedience is cheap to distill, and none of this tells you what it is like to work with a model. Hold the skepticism. It gets more interesting.

The ledger nobody reads

While preparing the cost comparison for this video, I broke down where Fable actually spends its money, and it was not where I expected.

Fable does not burn its budget on thinking. It burns it on orchestration. Tool calls, opening files, reading files, running commands in the terminal, the plumbing of agentic work. On terminal and agentic operations specifically, K3 came out around fifty times cheaper. In my own ledger, the step where Fable created an entire brand identity cost about the same as the step where it opened the HTML file to look at it. That sentence sounds like a typo. It is not. The most intelligent model on the planet spends frontier rates on work a model at a tenth of the price does identically.

This is genuinely good news, because it means the fix does not require abandoning anything. Keep the smart model as the orchestrator. Route the plumbing to something cheap. Most of your bill is plumbing.

Routing beats both

Fireworks published task-level data comparing K3 solo, Fable solo, and a router that assigns each task to whichever model suits it, and the routed configuration beat both solo runs on every benchmark they show, at a cost sitting just above the cheap solo run. Their conclusion, in their own words: single models are wasteful and no longer state of the art. The task-level detail explains why. These models are specialists at very different prices. K3 is sharpest on symbolic math and dev tooling. Fable wins on web work and data visualization, and the gap there is not small. Their oracle router sent 72 to 96 percent of task traffic to K3 and produced overall quality above either model alone.

Seventy-two to ninety-six percent. Sit with that range. The share of your workload that genuinely needs the frontier is the leftover slice, and it is shrinking with every open-weight release. I made this argument about agents a few weeks ago, one agent is not enough, and the model layer is now saying the same thing back. My own daily setup is still Fable orchestrating with GPT 5.6 Sol on execution, and I still think that pairing beats Fable plus K3. But the principle has stopped being a workflow preference and started being arithmetic.

The catch, because there is one

Before you cancel anything, the honest caveats.

On Artificial Analysis's new AA-Briefcase, the agentic knowledge-work benchmark, K3 jumps straight to second place, behind only Fable 5 and ahead of GPT 5.6 Sol. On the presentation-quality Elo, the closest thing benchmarks have to a taste measurement, it even edges past Fable, a result I will complicate in a moment. But the cost-per-task chart tells a different story: K3 is still roughly double the price of 5.6 Sol, sitting somewhere between Sonnet and Opus. Cheap next to Fable is not the same thing as cheap. And it is slow. K3 takes over twice as long as Fable to complete a task, longer than any model on the chart. For some work that does not matter. For plenty of work it does.

One more confession, because it is the honest frame for everything I write about Anthropic pricing. I evaluate Fable from inside a heavily subsidized 200 dollar Claude Max subscription. If I paid API list prices for it, this would be a very different post, and Fable being my favorite model would be a much harder position to defend.

The branding test

Now the part where the video stops being a benchmark recap. The new generation of models unlocked something I find genuinely remarkable: one-shotting a complete brand identity from a plain description. Typography, palette, logo system, brand guidelines, the whole document. Fable is the best design model in the world, so I gave the same two briefs to Fable and to K3, blind, same repo. Build the identity system, then build the launch landing page from it.

Fable produced the better work. The logo has taste where K3's looks like a poorly drawn SVG, the landing page demos are interactive where K3's mostly are not. But K3 was closer than anything else I have tested, and the landing page it built actually explained the product better than Fable's did.

Here is where the test stopped being about quality. The longer I scrolled the two guideline documents, the stranger they felt. The layout decisions, the section structure, the way both models chose to present type specimens, the small conventions nobody prompted for. K3 reads like a slightly worse photocopy of Fable's instincts. Discount Fable, is what I called it on camera, and I stand by the phrase. On backend work I cannot say that, because K3 is legitimately excellent there on its own merits. On design, the resemblance is uncanny enough that, for the first time, the distillation accusation stopped feeling like lab-war noise and started feeling plausible to me.

The bill for the whole experiment: 25.10 dollars for Fable across both builds, 3.94 for K3. Fable finished about thirty minutes sooner. That works out to 21.16 dollars for the speed, roughly 71 cents per minute saved. Whether that trade is worth it is not a rhetorical question. Sometimes it is. The point is that it is now a question you can answer per task, with a router, instead of once, with a subscription.

The accusation, read carefully

The White House statement is worth reading in full, and I do in the video. The claim is specific: a sophisticated internal platform for large-scale distillation against US models, rotating access methods to avoid detection, GB300 servers accessed through Thailand. The statement even goes out of its way to bless legitimate distillation before condemning the covert industrial kind aimed at stealing US technology.

Moonshot denies it, credits its own architecture work, and its defenders point out that Fable went public on July 1 and K3 shipped on July 15, which is a genuinely awkward timeline for a distillation story. My design test is suggestive, not proof; similar training data produces similar instincts, and no side-by-side of brand guidelines settles a question like this. So I will hold the position the evidence supports: I believe Anthropic has a point, I cannot prove it, and neither can you.

What I can say is that the answer changes the industry less than either side pretends. If the accusation is true, distilled models are structurally condemned to trail the frontier by a step, which is exactly where K3 sits. And true or not, the model exists, the weights are public, and the price is the price.

What gating cannot fix

The US response so far is gates. GPT 5.6 launched to an approved list. Fable and Mythos went dark for nineteen days over security concerns before coming back. I wrote about that era when it started, the US government now chooses who gets AI, and the K3 episode is its logical continuation: you cannot export-ban a model whose weights are already on every continent, and you cannot sanction bad actors into not downloading a file. Restrictions on open weights will slow legitimate research and inconvenience nobody dangerous.

The uncomfortable question underneath all of it is economic. If routers send 90 percent of the workload to open-weight models, the frontier labs keep the hardest 10 percent and the biggest training bills. Somebody has to fund the next generation, both labs have IPOs on the horizon, and the customers are learning to do this exact math. I do not know how that resolves. I do know that the only defense that has ever actually worked here is staying ahead, because a distilled model inherits your capabilities a step late, and a step is a lot if you use it well.

Where this leaves us

Fable 5 is still the best model I have ever used. That sentence has survived every test in this post, and it is also no longer a workflow. The question stopped being which model is best months ago. The question is which model is best at the step you are on, and for most steps, most of the time, the honest answer is no longer the expensive one.

Stop using Claude for everything. Use it where nothing else comes close, and route the rest.