Your AI coding agent's model menu moved three ways this week. One decision outlasts all of them.

On Wednesday, a model that had been switched off for nearly three weeks came back on. Claude Fable 5, the model sitting at the top of the SWE-bench Verified leaderboard, went dark on June 12 after the US Commerce Department issued an export-control order. On July 1 it returned, globally, once that order was lifted. I have watched a lot of harness news this year, and this is the first time I have seen a frontier coding model vanish and reappear by government action inside a single sprint.

That was one of three model moves in the same 72 hours. Together they say something more useful than any leaderboard.

Fable 5 came back, Sonnet 5 went cheap, and an open-weight model landed in Copilot

Three things happened between June 30 and July 2, and none of them was a new harness feature.

Fable 5 returned. Anthropic redeployed it globally on July 1 after Commerce lifted the export controls that had pulled it offline. The fix was not a new model, it was a new safety classifier, trained with the government to block the specific vulnerability-hunting technique an Amazon research report had surfaced.

When the classifier fires, the request is quietly routed to Opus 4.8 instead of refused. Fable 5 stays free up to half of weekly usage limits through July 7, then moves to metered usage credits at ten dollars per million input tokens and fifty per million out.

Second, a cheaper model went everywhere. Claude Sonnet 5 reached general availability across GitHub Copilot on June 30 and became the default in Claude Code the same week. On paper it lists around 40 percent under the Opus 4.8 flagship.

Here is the part the announcements skip. Sonnet 5 ships a new tokenizer that maps the same input to between 1.0 and 1.35 times as many tokens. A lower price on a model that counts more tokens per job is, as one unit-economics read put it this week, roughly cost-neutral against the model it replaces. Cheaper per token is not cheaper per task.

Third, the first open-weight model became selectable inside a major harness. GitHub made Kimi K2.7 Code generally available in Copilot CLI on July 1, the first open-weight option an admin can pick from the menu, arriving the same week enterprise managed settings reached general availability.

The model became the moving part of the harness, not the harness itself

For most of the last year the harness was the volatile thing and the model felt fixed. This week that flipped. The harnesses barely moved and the models did all the moving: one revoked and restored by the state, one repriced in a way that hides its real cost, one arriving from the open-weight world.

What a CEO actually owns when a model can be revoked, repriced, or replaced

For a CEO or founder, the useful frame is not tool choice, it is exposure. A single load-bearing model that a government can switch off is a supply risk, not a developer preference. A model whose cheaper tier quietly costs the same is a budget line that will surprise finance in Q3. An open-weight model on the menu is a new thing someone has to approve, secure, and watch.

The question this week is not which model won. It is who in your company is allowed to change the model, and who finds out when it changes on its own.

Every one of those moves presupposes a person: someone who decides which model each harness defaults to, who can swap it, and who gets paged when a vendor, or a government, swaps it first. Most org charts do not name that person. The role lands, silently, on whichever senior engineer happened to be reading the changelog that morning.

The model-selection inventory to run before the Q3 budget review

One move, and it fits inside a meeting. Before the next budget review, list every place an AI coding agent picks a model on your behalf: each harness, its current default, its fallback, and the pricing basis under it, whether flat, credit, or metered. Put one human name beside each row. Where the name is nobody, or on default, that is the gap this week just exposed.

The goal is not to predict the next model move. Anthropic did not predict the last one. The goal is that when the next one lands, one named person can see it, price it, and undo it, before it shows up as a benchmark nobody can run or a bill nobody modeled.