A coding agent working through a project can resend the same files, instructions and conversation history to a model dozens of times in an hour. Each resend shows up on the invoice as a cache read. Anthropic, in releasing a new version of its Claude model on Sept. 1, cut the price of that line item by 75%.
The new model, Claude Fable 5.1, is available immediately in general release, the company said. A variant called Mythos 5.1, built from the same model with looser safety guardrails, is being offered only to invited institutions in cybersecurity and the life sciences under a program the company calls Glasswing.
The headline prices did not move. Input tokens remain $10 per million and output tokens $50 per million. The change came in cache reads, which fell to 25 cents per million tokens from $1. The company said that based on its data from August, typical customer bills would fall about 25%, and that heavy agentic workloads, where software agents run long stretches of work, could see reductions of as much as 45%.
The outside measurements immediately disagreed. Artificial Analysis, which runs standardized tests across model providers, found that Fable 5.1 at its highest reasoning setting costs about 20% more per task than the previous Fable 5 model, because it generates roughly 1.7 times as many output tokens. Cognition, the maker of a coding agent, measured the opposite result, reporting that cost per task on its programming benchmark fell to $2.68 from $5.84.
The two numbers are not necessarily contradictory. They reflect different traffic patterns. A task that mostly re-reads a long, stable context benefits directly from cheaper cache reads. A task that pushes the model to reason harder and write more tokens pays on the output side, where prices did not change.
Cache reads are the most frequently billed item in an agentic loop. When an agent plans a task, checks its work, calls a tool and recovers from an error, it repeatedly re-reads the same project context. Anthropic’s move lowers the cost of that repetition rather than the cost of entry, an approach aimed squarely at developers who run agents for hours at a stretch.
Anthropic has positioned Claude around software engineering, where this pattern is common. Its coding tools have become a point of competition with rivals, and model providers have been cutting prices steadily as they fight for the workloads that developers attach to their assistants. Analysts said the cache discount is designed to make long-running agents affordable enough that developers stop watching the meter.
The pricing structure of large language models has several tiers, and providers make different choices about where to take revenue. Input tokens carry a premium when a customer sends new text. Output tokens, the model’s generated text, are priced highest because they consume the most compute. Cache reads sit in between, and cutting their price is a bet that volume will make up for the lower margin.
Mythos 5.1, the restricted variant, is a separate signal. By releasing the same underlying model with relaxed guardrails only to vetted research institutions, Anthropic is testing demand for models that can examine sensitive material, including descriptions of exploits and biological research, without the filtering that consumer products apply. Such restricted releases have become a standard way for labs to serve security and scientific customers while keeping general-purpose models aligned with their published policies.
The debate over Fable 5.1’s economics points to a deeper shift in how AI costs are measured. Token prices, the industry’s original yardstick, matter less than the cost of completing a task. Different benchmarks produce different task costs because they use the models differently, and the gap between Artificial Analysis and Cognition shows how little consensus exists on what an agentic task actually costs.
For Anthropic, the release is also a response to competitive pressure from several directions. Rivals have matched or undercut each other on frontier-model pricing repeatedly, and open-weight models have pulled down the price of routine inference. The company is betting that software engineering, where context is long and agents run deep, rewards a provider that prices the repetition cheaply.
Enterprises running agents at scale have begun to model their AI budgets in terms of cost per completed unit of work, whether that is a reviewed pull request, a resolved support ticket or a drafted document. Cheaper cache reads change those models directly. A workflow that was uneconomical at $1 per million tokens for repeated context can look reasonable at 25 cents.
The catch, analysts noted, is that the discount applies only to tokens the customer has already paid to write into the cache. Customers who send new material, or whose workloads are short-lived, see less of the benefit. The typical 25% reduction Anthropic cites describes blended usage, and customers whose agents churn out long answers at maximum effort could see bills rise even as their per-token costs fall.
Anthropic said the pricing change applies to all customers effective with the Sept. 1 release, and that the company will publish usage guidance for teams that want to restructure their agent loops around cheaper cache reads. Third-party trackers said they would update their cost comparisons once Fable 5.1 has had time to accumulate production traffic.
The long-running question for the model industry is whether falling prices will be offset by rising usage. Anthropic is betting the answer is yes, and that developers who stop rationing context will find more ways to put agents to work. The 75% cut on cache reads is the price of that bet.
