Claude Fable 5.1 Cuts Cache Read Prices 75%

Claude Fable 5.1 keeps Anthropic’s $10-per-million input-token and $50-per-million output-token prices, but cuts prompt-cache reads from $1 to $0.25 per million tokens. That 75% line-item reduction can materially lower the cost of long-running agents, although independent testing shows the new model may still cost more per completed task at its highest effort setting.

What changed with Claude Fable 5.1

Anthropic describes Claude Fable 5.1 as a frontier model for demanding coding, research and knowledge work. It is available in paid Claude plans and through the Claude API, supported cloud marketplaces and Microsoft Foundry. The company retained the headline API rates used for Fable 5: $10 per million uncached input tokens and $50 per million output tokens.

The pricing change sits in prompt caching. When an application repeatedly sends the same large context—code repositories, policy documents, tool definitions or a long conversation—the provider can cache that material and charge a lower price when it is read again. Anthropic’s Claude Fable 5.1 product page says those reads now cost $0.25 per million tokens, 75% less than Fable 5’s $1 rate.

Everyone else is reporting a cheaper cache; we are explaining why the discount does not guarantee a cheaper task. The invoice depends on how much context is reused, how many new tokens enter the prompt, how much the model writes, and what effort setting the developer chooses. Claude Fable 5.1 shifts one important cost lever without reducing every component.

Claude Fable 5.1 pricing in one table

The table prevents a common misreading: Anthropic did not cut Claude Fable 5.1’s entire API price by 75%. Only cache reads fell by that amount. A new prompt still incurs the standard input price, creating a new cache still costs $12.50 per million tokens, and generated output still costs $50 per million.

Why agentic workloads benefit most

An agent often works in a loop. It reads the same repository map, instructions and conversation history; calls a tool; observes the result; then starts another reasoning step with much of that context attached again. Without caching, every repeated byte is priced as fresh input. With caching, the stable prefix can be charged at the cache-read rate.

Consider a simplified workflow with 200,000 cached tokens read forty times. At Fable 5’s $1-per-million rate, those reads cost $8. At Claude Fable 5.1’s $0.25 rate, they cost $2. That is a $6 cache-line saving before counting cache creation, uncached input, output or any platform markup.

The saving becomes less important when prompts are short, change constantly or produce large answers. Output is the most expensive published token category for Fable 5.1. A model that writes substantially more can erase the cache benefit even while the per-token rates remain unchanged.

Cheaper cache does not always mean cheaper work

Artificial Analysis evaluated Claude Fable 5.1 before release and reported a score of 66 on its Intelligence Index at maximum effort, ahead of Fable 5’s 62. However, it measured about $3.76 per index task for Fable 5.1 versus $3.14 for Fable 5—a roughly 20% increase. Its explanation was higher output-token use, not a reversal of the cache discount.

This distinction is the core purchasing lesson. Token prices are ingredients; cost per successful task is the meal. A model that finishes a complex job in fewer failed attempts may provide better value even at a higher single-run cost. Conversely, a verbose model can raise bills when users do not need the additional analysis.

Developers should measure complete workflows with their own prompts and tools. Compare success rate, retries, latency, cached-input share and output length. A synthetic leaderboard helps establish capability, but it cannot predict the exact economics of an Indian support bot, a legal-research pipeline or a software agent operating on a private repository.

What Anthropic says about performance

Anthropic reports that Claude Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1, more than double Fable 5’s 24.7%. It also highlights gains on coding and long-horizon knowledge work. These are company-reported or partner-run evaluations and should be read with the test definitions, effort settings and confidence ranges.

Independent confirmation is useful but not identical to production proof. Artificial Analysis found Fable 5.1 strongest in its index while noting that some leads over Claude Opus 5 were within confidence intervals. Hardware Busters focused on the science benchmark and price structure. Cursor’s model documentation independently lists the same $10 input, $0.25 cache-read and $50 output rates.

Lapaas Voice previously covered the original Claude Fable 5 launch and Anthropic’s later temporary free-access extension. Claude Fable 5.1 is a distinct release because it changes the model, benchmark results and API economics rather than merely changing access to the earlier version.

How teams should evaluate the upgrade

Start with workload traces, not a generic chat test. Select representative tasks and record the stable prompt prefix, cache hits, new input, output, tool calls, retries and whether the result passed human review. Run the same set on Fable 5 and Claude Fable 5.1 with equivalent effort settings.

Next, test failure modes. Anthropic says its updated cybersecurity safeguards reject benign requests less often, but teams should not treat a lower refusal rate as permission to weaken controls. Applications handling sensitive code, medical information, financial actions or personal data need their own authorization, logging and human approval boundaries.

Finally, compare Fable 5.1 with cheaper models for each stage. A smaller model may classify requests, retrieve documents or draft routine text, while a frontier model handles the hardest reasoning step. Routing work by difficulty can matter more than a single-provider discount. Lapaas Voice’s report on Claude Opus 5 pricing and positioning offers another reference point within Anthropic’s model line.

Cache design also deserves operational attention. Reuse works best when applications keep a stable prefix and append changing user data afterward. Constantly reordering tool definitions or rewriting system instructions can reduce cache hits even when the underlying information is similar. Teams should expose cache-read and cache-write tokens in their observability dashboards, set alerts for sudden hit-rate changes, and test whether prompt updates improve quality enough to justify rebuilding the cache. That engineering discipline turns the advertised rate cut into a measurable production saving.

What Claude Fable 5.1 means for India

Indian AI startups often serve price-sensitive customers while paying inference costs in dollars. A lower cache-read rate can improve margins for coding assistants, compliance tools and multilingual knowledge agents that repeatedly use large instructions or document sets. The benefit is strongest when the context remains stable across many turns.

Currency exposure and cloud-platform markups still matter. The published Anthropic API rate may differ from an enterprise’s effective cost after taxes, marketplace pricing, data-residency architecture and observability tools. Teams should calculate rupees per successful task and include engineer review time, not only tokens.

Claude Fable 5.1 makes repeated context four times cheaper to read, but it does not make every agent run 75% cheaper. The practical saving depends on cache reuse and can be offset by longer outputs or higher effort. That is the self-contained conclusion procurement and engineering teams should carry into testing.

Frequently asked questions

How much does Claude Fable 5.1 cost?

Anthropic lists $10 per million standard input tokens, $12.50 per million cache-write tokens, $0.25 per million cache-read tokens and $50 per million output tokens.

Is Claude Fable 5.1 75% cheaper than Fable 5?

No. Cache reads are 75% cheaper. Standard input, cache writes and output retain the same listed rates, so total savings depend on the workload.

Why did an independent benchmark find higher task cost?

Artificial Analysis reported that Claude Fable 5.1 used more output tokens at maximum effort. Because output remained $50 per million tokens, the extra generation outweighed the cache savings in that test.

Who should upgrade first?

Teams running long, repeated agent loops with stable context are best positioned to benefit. They should still benchmark full task cost, quality and latency before switching production traffic.