The Frontier Got a Price Tag, Then a Kill Switch
Anthropic shipped the priciest model any laboratory sells, then the government switched it off. Closed frontier models are rented capability, repriced or revoked at will. Open weights cannot be.
On June 9, Anthropic released Claude Fable 5, the first publicly available model from its restricted Mythos tier, a class previously reserved for cyber-defense partners and a small group of researchers. The company priced it at $10 per million input tokens and $50 per million output tokens, twice the rate of Opus 4.8 and the most expensive among major AI models then available.
The pricing carried a structural signal that mattered more than the sticker. Fable would not live inside the flat monthly subscription. Anthropic included it on Pro, Max, Team and seat-based Enterprise plans at no additional cost only through June 22, after which continued use would draw on metered usage credits. The best model the company sells would be billed against the work it produces rather than folded into a fixed fee.
A parallel shift had already reached developers. On June 1, GitHub moved Copilot to usage-based billing, retiring its flat premium-request allotments in favor of metered AI Credits priced at one cent each. The published plan prices held steady; the volume of work they covered did not. The change ended a quiet subsidy that had let heavy users consume far more compute than their fee implied, and it drew sharp criticism from developers.
Two pricing decisions at two companies described one movement. Frontier intelligence, sold for three years as an all-inclusive monthly fee, is separating into two tiers. The commodity tier of smaller and mid-range models stays bundled and continues to fall in price. The frontier tier occupies a metered slot priced against the labor it displaces rather than the compute it consumes, and it does not collapse back into the bundle. Flat-rate access to the best model is ending.
Speed becomes a budget line
The stakes rise as models take on longer work. METR, a research group that measures the length of task an AI agent can complete unaided, has documented that this time horizon has doubled approximately every seven months since 2019, extending from work that took seconds to work that occupies a skilled professional for hours. As the horizon lengthens, human supervision moves from approving each step toward setting direction and reviewing results at milestones. The plan, build, test and ship loop that once required a person at every stage now requires one at the checkpoints.
That compression rewards whoever can buy more of it. Firms operating with large teams, deep capital and mature engineering do not gain a marginal improvement; they attach a continuous, tireless model fleet to an organization that was already ahead. The distance between fast and slow companies widens not in proportion to talent but in proportion to budget. Spending on intelligence is becoming the variable that sorts the market.
The off switch
The metered-frontier story acquired a second dimension on Friday. At 5:21 p.m. Eastern on June 12, three days after Fable reached the public, Anthropic received an export-control directive from the Commerce Department's Bureau of Industry and Security instructing it to suspend all access to Fable 5 and the related Mythos 5 for any foreign national, citing national-security authorities. Complying with an order written around foreign nationals required disabling the models for every customer worldwide, which Anthropic did while its other models remained available. In a public statement, the company said it disagreed, attributing the action to a single narrow jailbreak finding and cautioning that recalling a deployed model on that basis would, applied evenly across the industry, halt frontier releases altogether.
The most capable model on the open market went dark by government decree, worldwide, within an afternoon. The cause was not a price increase. A party that was neither the buyer nor the vendor decided the capability should stop. The frontier is not only expensive. It is revocable.
What cannot be switched off
The sorting mechanism of budget and the new fragility of revocation share a precondition. Both operate only on capability that a firm rents. One resource in the market does not behave like a rental. A trained model is, in the end, a file. Once a sufficiently capable open-weight model exists, it propagates: organizations download it, adapt it and run it on hardware they own, incurring no recurring charge. A model already resident on a company’s own machines presents no vendor for a directive to reach, no page to switch to unavailable and no invoice. A file that has been copied a thousand times cannot be recalled.
That property carries a complication worth stating plainly. Several of the strongest open-weight models now originate in Chinese laboratories (DeepSeek, Alibaba’s Qwen and Z.ai’s GLM), so the choice of which open weights to trust raises its own questions of provenance and security. The distinction holds regardless. Trust concerns which file to run, not whether the file can be taken back.
A measured lag
Open weights remain behind the closed frontier, and the size of the gap is documented. Epoch AI finds that the best open-weight models have trailed the leading closed models by an average of roughly four months since the start of 2026, a margin that has widened slightly rather than narrowed against the 2023 to 2025 trend. Four months behind the frontier nonetheless sits well ahead of what most production work demands. Summarization, extraction, classification, the majority of coding and most internal tooling do not require the leading edge; they require capability that is sufficient, owned and unmetered. The frontier commands its premium on the difficult exceptions and on iteration speed, which is why well-funded firms will continue to pay for it. For the remaining majority of work, the open floor keeps clearing the threshold that matters, even where it never reaches the frontier itself.
The precedent
The pattern has a precedent at industrial scale. Through the 1990s, serious computing belonged to proprietary Unix and to Windows, and Linux trailed both. Over the following two decades Linux became sufficient, owned and unmetered, and it displaced the incumbents across the most demanding workloads. Every system on the TOP500 list of the world’s fastest supercomputers has run Linux since 2017, and the operating system underpins most web servers and every major public cloud. The proprietary moat was not breached so much as out-iterated by software that anyone could copy.
The more instructive half of the precedent concerns where the value settled. Open code did not impoverish the companies built on it; it relocated the profit. In July 2019, IBM closed its acquisition of Red Hat for roughly $34 billion, the largest open-source acquisition on record, for a company whose flagship product was available to download at no cost. Reviewing the deal, the trade publication CIO Dive observed that IBM was buying people rather than intellectual property. The kernel was a commodity; the team that could harden it, support it and deliver it to large enterprises commanded the price.
The analogy has limits, and they matter. Linux improved through a distributed volunteer effort, with more than 15,000 kernel contributors revising the code continuously, and the software was inexpensive both to copy and to advance. Frontier models break that symmetry. Copying weights costs nothing, but training the next and better open model demands substantial compute and scarce expertise, and current open weights issue from a small set of corporate laboratories rather than a broad commons. Linux also required roughly three decades to claim its territory, whereas the open-model gap is measured in months, a faster clock that can run in either direction.
The work the precedent demands
Closing the remaining gap once relied on distillation, the practice of training an open model on the outputs of a closed one and following in its wake. That route is narrowing. The gains that now separate the leading models reside increasingly in reinforcement-learning environments and agent scaffolding, components that do not appear in any output a competitor can copy. Open-weight programs that depend on imitation will stall, while those that invest in the harder infrastructure of training environments, evaluation suites, agent harnesses and post-training pipelines will not. The four-month lag becomes a durable position only for the organizations willing to fund that work.
The durable advantage, on this reading, was never the model. A rented capability can be repriced by its vendor, out-bid by a competitor or, as June 12 demonstrated, switched off by a government. A capable team and the open weights it controls can be none of those things. Whether the open ecosystem converts its narrow lag into a lasting foothold will depend less on the next closed release than on how much capital flows into the unglamorous work of perfecting open models. Fable 5 stayed offline through Friday evening, its landing page reading “temporarily unavailable,” as Anthropic and the Commerce Department remained at odds over whether the order was justified.