The Free Trial Ends at Exactly the Wrong Moment
Since June 9, Anthropic has included Claude Fable 5, the publicly accessible version of its frontier Mythos-class model, at no extra charge for subscribers on Pro, Max, Team and seat-based Enterprise plans. That two-week window closes today. From tomorrow, June 23, Fable 5 drops out of standard plan limits, and anyone who wants to keep using it must do so through usage credits billed at Anthropic's API rates: $10 per million input tokens and $50 per million output tokens. That is double the cost of Opus 4.8, the model Claude will quietly fall back to once the free access expires.
The timing creates a familiar and uncomfortable dynamic. For two weeks, teams have been encouraged to throw their hardest problems at the most capable model available, building habits and workflows around its output. Today those habits collide with a meter. The free trial that drove adoption is the same mechanism that now exposes the bill, and the organizations that leaned in hardest are precisely the ones facing the sharpest adjustment. We have seen this play out across the industry in 2026, where the gap between a model's list price and its real-world consumption is where budgets quietly break.
What Fable-Class Capability Actually Costs
The headline token prices understate the real spend, because frontier agentic work is token-hungry by nature. Heavy users report that Fable 5 sessions routinely consume between 500,000 and one million tokens, and that long-running agent tasks can swallow far more. During the free window, one developer described hitting the $200-per-month Max plan's limits in under 30 minutes. Independent observers have reported single days of usage costing over $100, and a single 78-million-token session pricing out at roughly $99 at API rates. These are not pathological edge cases. They are what serious agentic coding and analysis look like when the model is doing real work.
There is a partial offset worth naming. Fable 5 tends to use roughly half the tokens of Opus 4.8 for comparable agentic work, which softens the doubled per-token price. In effect, the premium is closer to a wash than the sticker shock suggests, provided the work genuinely benefits from Fable's added capability. The danger is using the expensive model for tasks that a cheaper one would handle just as well. At highest effort, Fable reflects on and validates its own work, which is exactly the behavior that makes autonomous operation viable and also exactly the behavior that runs the token counter the fastest.
Why Anthropic Is Metering, Not Bundling
Anthropic has been candid about the logic. The company expects demand for Fable 5 to be very high and difficult to predict, which is precisely the situation in which a flat subscription becomes dangerous for a provider. If a fixed monthly fee gives every user uncapped access to the most expensive model, a small cohort of power users can consume a disproportionate share of compute and torch the unit economics. Metering the frontier tier through credits transfers that variability back to the customer, where it can be measured and controlled.
This fits a broader pattern we have tracked all year, in which AI vendors are racing to attach spend controls, hard caps and per-user metering to their most powerful offerings. The economics of frontier inference simply do not support all-you-can-eat pricing on the leading edge. The credit model also preserves a clean separation: routine work stays inside predictable plan limits, while the genuinely demanding jobs that justify Fable-class capability get priced explicitly. For a provider that has just filed confidentially for an IPO, demonstrating disciplined unit economics on its most expensive product is not a side concern. It is the whole game.
The Enterprise Playbook for Tomorrow
For technology leaders, the right response is neither panic nor indifference. The first move is an audit: identify which employees and which workflows genuinely require Fable-class capability, and which were simply using it because it was free and available. The bulk of day-to-day work, drafting, summarizing, routine coding, runs perfectly well on Opus 4.8 at half the price. Defaulting to the cheaper model and reserving Fable for the hardest problems is the single biggest lever on cost, and it requires nothing more than guidance and a sensible default.
Beyond defaults, the operational discipline matters. Teams should use Fable's effort levels deliberately, reserving the highest-effort, self-validating mode for tasks that truly warrant it, and they should fund and monitor usage credits with the same rigor they apply to any other consumption-based bill. Crucially, API and cloud-platform access through Bedrock, Vertex and Foundry is unaffected by today's change, so organizations already consuming Fable 5 programmatically face no disruption. The shock lands squarely on seat-based subscribers who treated the free trial as the permanent price. As of tomorrow, it is not, and the teams that planned ahead will be the ones not explaining a surprise invoice next month.
