Anthropic has once again extended the free usage period for its top-tier artificial intelligence model, ‘Claude Fable 5.’ The decision to delay the transition to a usage-based billing system comes just as rival OpenAI launches a new model, and is seen as a move to stem user defections. However, the process has also intensified a ‘hidden cost’ controversy, with analysis revealing that Anthropic’s proprietary development tool, ‘Claude Code,’ unnecessarily consumes a massive number of tokens, driving up user expenses.
On the 12th (local time), Anthropic announced via its official X (formerly Twitter) account that access to Claude Fable 5 would be extended without additional charges for all paid plan subscribers until the 19th. Anthropic had initially planned to switch to a token-based, pay-as-you-go pricing model starting on the 8th of this month, but first postponed it to the 12th and has now delayed it again to the 19th. Consequently, subscribers to the Pro ($19/month), Max ($110/month), Team, and Enterprise Premium Seat plans can use up to 50% of their weekly usage limit on Fable 5 until the 19th. A promotion increasing the weekly usage limit for the AI coding tool ‘Claude Code’ by 50% has also been extended for the same period.
Anthropic has made it clear that after the 19th, Fable 5 will no longer be included in the weekly usage caps of its subscription plans and will only be accessible by activating a token-based ‘additional usage’ billing plan. While a usage-based model was already applied to developers and enterprises using Fable 5 via API integration, general paid subscribers previously enjoyed unlimited access under a flat-rate plan. However, with AI agent usage exploding and token consumption surging even among individual users, the company is moving to convert its entire pricing structure to a usage-based model.
Industry observers interpret this latest extension as a strategy to appease users in response to the launch of a new model by rival OpenAI. OpenAI unveiled its next-generation ‘GPT-5.6’ series on the 9th, emphasizing that its top-tier model, ‘Sol,’ delivers higher performance at up to one-sixteenth the cost of Fable 5. Indeed, in the DeepSWE benchmark, Sol Max achieved a 73% score at a cost of $8.39 per task, whereas Fable 5 cost approximately $22 to achieve a 70% score. In the Agents’ Last Exam, Sol also scored 53.6 points, 13 points higher than Fable 5 with adaptive reasoning applied.
Fable 5 was launched on June 9, but its use by foreign nationals was restricted shortly after due to U.S. government export controls. Service was resumed on July 1 after safety measures were supplemented. However, when news of the impending switch to usage-based pricing broke right after the service resumed, users reacted strongly, complaining they would have to “pay 8 to 20 times more than before.” With the launch of GPT-5.6 on top of this, Anthropic appears to have had no choice but to play the promotion extension card to prevent an immediate exodus of developers.
However, the extension is not being viewed entirely favorably. Some developers have sarcastically called the repeated delays to Fable 5’s pay-as-you-go transition an “unreachable carrot.” One developer pointed out, “With GPT-5.6, I couldn’t find a price hike date or a date it would be excluded from the subscription plan, and it felt strangely odd that OpenAI staff explanations matched the official documentation.”
Adding to the scrutiny, analysis has emerged that Anthropic’s proprietary coding tool, Claude Code, inflates user costs by consuming an unnecessarily large number of tokens. According to test results reported by the IT publication InfoQ, even in the simplest task requiring only a 22-character “OK” response, Claude Code consumed approximately 32,800 tokens before the user prompt was even delivered. In contrast, the initial load for the open-source alternative OpenCode was about 6,900 tokens, a mere one-fifth of the amount.
Claude Code’s excessive token consumption primarily stems from its 27 tool definitions. Beyond core programming tools, it includes a vast array of background agent and task orchestration features such as CronCreate, Monitor, Task series, worktree management, and push notifications, accounting for roughly 24,000 tokens. Even the system prompt alone reached about 6,500 tokens, more than three times that of OpenCode. This means costs are leaking into Anthropic’s self-imposed ‘platform overhead’ before the user even issues a real command.
The difference in cache efficiency was also stark. OpenCode’s request prefix is byte-identical, allowing it to be written once and then read almost for free thereafter. In contrast, Claude Code repeatedly rewrites a cache prefix of tens of thousands of tokens during a session, resulting in cache write volumes up to 54 times higher for the same task. Since cache writes are billed at a premium rate, the usage dashboard spikes much faster than the actual work completed.
The use of sub-agents is another factor amplifying costs. When one user assigned a relatively large task to Claude Code, seven sub-agents were immediately activated, exhausting the entire budget. When the main agent performed the same task sequentially, there was no issue. Real-world data also showed that a task costing 121,000 tokens when completed directly ballooned to 513,000 tokens—a 4.2-fold increase—when distributed across two sub-agents. This is because every time a sub-agent is launched, the entire system prompt and tool definitions are reloaded, and the parent agent must re-read all of its output.
The problem is that all this token consumption is unrelated to any improvement in output quality. In the same tests, both tools performed all scoring tasks accurately with no difference in output quality. Ultimately, users are paying several times more to achieve the same results.
Speculation is also rising that Anthropic’s actions harbor strategic intent beyond a simple promotion extension. Some in the AI industry suggest that Anthropic is deliberately creating a new industry standard where top-tier state-of-the-art (SOTA) models are excluded from flat-rate subscriptions and offered exclusively via API. If OpenAI were to align with this trend, both companies could maximize profitability while avoiding a price war. With both pursuing initial public offerings (IPOs) at a trillion-dollar scale, a price war benefits neither party.
Meanwhile, an AI token is a concept entirely distinct from a token in the virtual asset market. An AI token is the smallest processing unit used by artificial intelligence to understand and generate data. Just as Lego blocks are assembled to complete a house, tokens are combined in sequence to create sentences, images, and code. It is sometimes likened to ‘digital fuel,’ as users purchase tokens to instruct AI to perform tasks, much like buying fuel to operate a machine.