Claude Fable 5.1 Arrives: 75% Cheaper Cache Reads—What’s Changed from Fable 5?

September 2, 2026Updated: September 2, 2026Articles

Anthropic released its new Claude model, Fable 5.1, on September 1, 2026.

Fable 5.1 improves its ability to sustain work on difficult tasks while cutting API cache-read pricing—the cost of repeatedly reading the same context—by 75%.

The clearest way to understand this update is as a combination of better performance and pricing changes that make long-running agentic work more affordable.

Key Features of Fable 5.1

Here are the main features of Fable 5.1:

Improved performance on tasks with many steps, including coding, research, and computer use
API cache-read pricing reduced from $1 to $0.25 per million tokens
Estimated cost reductions of around 25% for typical API workloads and up to around 45% for highly agentic workloads
Fewer false positives from safeguards, making legitimate cybersecurity work easier to carry out
Available on all paid Claude plans, as well as in Claude Code, on major cloud platforms, and through the API

Anthropic recommends starting with Opus 5 for most everyday work and choosing Fable 5.1 for more demanding reasoning or long-running agentic tasks. Rather than a daily model that replaces every conversation, Fable 5.1 is the top-tier option for difficult problems.

What Has Changed from Fable 5?

Item	Fable 5.1	Fable 5
Input pricing	$10 per million tokens	$10 per million tokens
Output pricing	$50 per million tokens	$50 per million tokens
Cache reads	$0.25 per million tokens	$1 per million tokens
Context window	1 million tokens	1 million tokens
Maximum output	128,000 tokens	128,000 tokens

Base input and output prices, along with the amount of context the model can handle, remain unchanged. The major changes are its ability to complete long tasks and the price of cache reads.

For a closer look at Fable 5’s role in the Claude lineup, see our guide to Claude Fable 5.

The 75% Cut in Cache-Read Pricing Is the Key

Prompt caching stores lengthy instructions or reference material that the model has already processed so it can reuse them in later interactions. Large codebases, long specifications, multiple documents, and conversation histories do not have to be processed from scratch every time.

With Fable 5.1, reusing that context costs one-quarter of what it did before. The difference may be small for a single short question, but it grows more significant when a task repeatedly uses tools while referring to the same material.

Implementation work in Claude Code that involves checking many files
Research that repeatedly investigates and verifies information across multiple sources
Work that produces documents, spreadsheets, and slides using the same rules
Agentic tasks that carry a long conversation history forward

Anthropic estimates that typical token-billed workloads could become around 25% cheaper, with savings of up to around 45% for highly agentic workloads that make heavy use of caching. Actual savings depend on how much context is written to the cache and how often it is reused.

We discuss how to compare models beyond their unit prices—including retries and human corrections—in choosing an AI model by cost per completed task. Fable 5.1 is also best judged by the total cost of finishing the work, not the price of a single response.

How Much Has Performance Improved?

In the main evaluations published by Anthropic, Fable 5.1 outperforms not only Fable 5 but also ChatGPT’s GPT-5.6 SOL.

Evaluation	Fable 5.1	Fable 5	GPT-5.6 SOL
Terminal-Bench-Science 0.1	52.6%	24.7%	22.4%
Terminal-Bench 4.0	55.8%	42.0%	37.3%
AutomationBench	31.4%	17.1%	19.6%
CursorBench	73.4%	70.5%	67.2%

Anthropic also says the model is more inclined to fix the underlying cause of a problem rather than take shortcuts that merely pass the immediate tests.

In real work, however, results vary with the tools, instructions, reference material, and reasoning effort used. Anthropic says Fable 5.1 can approach Fable 5’s performance even at lower effort settings, so reducing the effort level is worth trying when speed and cost matter.

How Does It Compare with GPT-5.6 SOL?

Fable 5.1 and GPT-5.6 SOL are both top-tier models with context windows of around one million tokens and a maximum output of 128,000 tokens. Their API pricing, however, differs considerably.

Item	Fable 5.1	GPT-5.6 SOL
Input pricing	$10	$4
Cache reads	$0.25	$0.40
Output pricing	$50	$20
Context window	1 million tokens	1.05 million tokens
Maximum output	128,000 tokens	128,000 tokens

Prices are per million tokens. GPT-5.6 SOL’s $4 input and $20 output rates are promotional prices available at least through November 21, 2026. Requests with more than 272,000 input tokens are charged at twice the input rate and 1.5 times the output rate.

For regular input and output at the rates shown above, GPT-5.6 SOL is 60% cheaper, while Fable 5.1 is 37.5% cheaper for cache reads. SOL’s price advantage is more likely to matter in short conversations or API workloads that process different information each time. Fable 5.1’s cache pricing may be more useful for tasks that repeatedly reuse the same large context.

On performance, Fable 5.1 surpassed SOL in the terminal, automation, and coding evaluations Anthropic published with this release. These comparisons were conducted in Anthropic’s environment, though; they do not mean that Fable 5.1 is better than SOL at every task.

In practice, the choice also depends on the working environment, not just the model. One approach is to use Fable 5.1 when pursuing a difficult problem over a long session in Claude Code, and GPT-5.6 SOL when implementing changes locally or progressing multiple tasks in Codex. In my own workflow—Claude for specifications, Codex for implementation—it makes more sense to use each where it works best than to have one replace the other.

Fable 5.1 vs. Mythos 5.1

Anthropic announced Mythos 5.1 alongside Fable 5.1. They share the same underlying model but have different safeguards.

Fable 5.1 is the generally available version. Mythos 5.1 has fewer safeguards and is offered through access programs for trusted organizations. It is not another option that regular Claude users can compare and select in the model picker.

Which Plans Include Access?

Fable 5.1 is available on all paid Claude plans, but not on Free. How its usage is covered differs by plan.

Plan	Fable 5.1 access
Max / Team Premium	Use up to 50% of the weekly usage limit at no additional charge
Pro / Team Standard	Usage credits are consumed from the first use
Enterprise	Included allowance or usage-based billing, depending on the contract
Free	Not available

There are no new free credits for Fable 5.1. The 50% allowance on Max and Team Premium draws from the regular weekly usage limit; it is not a separate, unlimited allowance on top of it. Fable models also consume usage more quickly, so it makes sense to reserve them for difficult work.

For more on the 50% allowance, see Fable 5’s continued inclusion in subscriptions.

Subscribers Move Straight from Fable 5 to 5.1

Fable 5.1 is offered as Fable 5’s successor. Users on plans such as Max and Team Premium, which allow up to 50% of the weekly usage limit to be spent on Fable, can use Fable 5.1 under that same allowance.

API users need to change the model ID to claude-fable-5-1. Base input and output prices remain the same as Fable 5; only cache reads become cheaper.

That said, there is no need to choose Fable 5.1 solely for short questions, text formatting, or simple conversions. Anthropic itself recommends starting with Opus 5 for most work.

What to Check in Claude Code and the API

Fable 5.1 requires Claude Code version 2.1.250 or later. If it does not appear, update Claude Code to the latest version first. The API model ID is claude-fable-5-1.

Anthropic also notes that Fable 5.1 can be inclined to rewrite an entire file even for a small change. If you only want to modify part of an existing codebase, it is safer to say explicitly, “Change only the necessary sections” and “Leave unrelated code untouched.”

AI Jiten’s Take: Making Difficult Work Affordable to Sustain Matters More Than Raw Performance

When Fable 5 arrived, I found it more useful for solving problems I had not been able to resolve before, or for building systems I would keep using, than for brief everyday conversations. That approach remains the same with Fable 5.1.

The significance of this update goes beyond higher benchmark scores. By lowering cache prices, which matter more as tasks run longer, Anthropic has made its top-tier model easier to incorporate into ongoing work rather than reserve for a one-off experiment.

The first things I would try are tasks that stalled with Fable 5 or required repeated retries with Opus 5. If you also have access to GPT-5.6 SOL, give both models the same task and compare the time to completion, rework, and usage consumed. That should help reveal the right division of labor for your own work.

Summary

Claude Fable 5.1 strengthens long-running reasoning, coding, and automation while retaining Fable 5’s base prices and one-million-token context window. The 75% cut in cache-read pricing creates the potential for improvements in both performance and cost on tasks that repeatedly reuse a large context.

Use Opus 5 for everyday work, smaller models for lighter tasks, and Fable 5.1 for difficult problems that take longer to solve. Rather than switching simply because a model is new, start with work where Fable 5.1 could reduce the need for retries and corrections.

References

Model specifications, pricing, and availability are current as of September 2, 2026. Benchmark results and estimated cost reductions are based on evaluations and estimates published by Anthropic. Actual results vary with the task and settings.