Anthropic spent five weeks unable to decide what to do with its most powerful AI model. The result, announced July 17 and taking effect July 20, draws a clear line through its paying subscriber base.
Max and Team Premium subscribers keep Fable 5 as a permanent feature. But there is a catch on top of a catch. Fable 5 is available to them at only 50 percent of their standard weekly limits. And those standard limits are themselves shrinking on July 20, when a 33 percent bonus Anthropic had been running comes to an end. So Max users wake up on July 20 with meaningfully less of the model than they had the day before.
Pro subscribers and Team Standard subscribers lose ongoing access altogether. They get a one-time $100 credit, and then Fable 5 costs them at full API rates: $10 per million input tokens and $50 per million output tokens. That is the highest pricing Anthropic charges for any publicly available model, and double what its previous top model cost. A single involved work session can run $4 or more. The $100 credit is a buffer, not a plan.
The road to this point was chaotic. Anthropic originally planned to end included access on July 7. Then it extended to July 12. Then to July 19. Each extension was a sign that the company had not yet settled on a strategy. The real reason for the reversals was a combination of capacity problems and competitive pressure.
On the capacity side, Fable 5 is simply expensive to run. It belongs to a new class of model that requires substantially more computing per response than the previous generation. Even after Anthropic secured a deal for more than 220,000 Nvidia chips through SpaceX's Colossus data center, it still could not serve the model freely to all subscribers without degrading quality across the board.
On the competitive side, OpenAI released GPT-5.6 Sol on July 9, right in the middle of Anthropic's deliberations. According to independent benchmarking by Artificial Analysis, GPT-5.6 Sol scores just one point below Fable 5 on a widely used intelligence measure, at roughly one third of the cost. That single fact changes the negotiating position considerably. If a competitor offers near-identical results for a third of the price, keeping Fable 5 locked behind expensive pay-per-use billing becomes harder to justify.
Below the OpenAI rivalry sits a broader shift that business operators should pay attention to. Chinese AI makers have been aggressively undercutting US frontier models on price. Open-source Chinese models from DeepSeek, Zhipu, and Moonshot AI now perform close to Anthropic's previous flagship while costing 60 to 90 percent less. Some are free to download and run on your own infrastructure entirely. In June 2026, Zhipu's GLM 5.2 reached within a percentage point of Anthropic's Opus 4.8 on an agentic benchmark, at roughly a fifth of the cost. Moonshot's Kimi K3, released this week, reportedly beat Fable 5 on some front-end coding tests. These are not distant threats; they are already being used by Western companies looking to cut AI costs.
This matters for anyone evaluating AI tools for business use. The pricing gap between a premium US model and a capable Chinese alternative has grown large enough that finance teams are now part of the conversation. A workload producing 50 million output tokens a month costs roughly $2,500 on Fable 5 and around $14 on DeepSeek's flash-tier model. The gap is not a rounding error.
For the most demanding and sensitive work, US frontier models still hold real advantages: better reliability, stronger safety controls, government compliance readiness, and more consistent output on complex multi-step tasks. But the threshold for when you actually need that level is now a genuine business question, not an assumption.
For anyone currently on a Claude Pro or Team Standard plan using Fable 5 for regular work, the practical move is to assess whether Max or Team Premium is worth the step up, or whether a capable alternative at lower cost handles enough of what you need. The AI market has become competitive enough that the answer is no longer obvious.