Anthropic left the headline rate of Fable 5.1 identical to Fable 5, then quartered the price of a cache read. This Claude Fable 5.1 test measures what that shift actually does to an invoice, and what the sixty percent drop in interruptions unlocks along the way.
Key Takeaways
- Cache reads fall from 1 dollar to 0.25 per million, cutting total cost by 25 to 45% depending on the workload.
- The model beats Fable 5 on every published benchmark, with margins running from 3 to 28 points.
- Three API breaks await any team migrating, starting with the removal of forced tool use.
Have an AI Sum Up This Test
ChatGPTAn Unchanged Rate and a Bill That Drops Anyway
The starting point of this Claude Fable 5.1 test is a pricing grid that did not move. Ten dollars per million input tokens, fifty on output, exactly as for Fable 5.
The shift sits one line further down. Cache reads drop from 1 dollar to 0.25 per million, a fourfold cut that appears nowhere in the announcement headline.
The effect depends entirely on the shape of the workload. On multi-step agentic tasks with 90% of context served from cache, measured cost falls by roughly 20.9%, and the broader estimate lands between 25 and 45% across scenarios.
In practice, a team replaying the same system prefix across thousands of daily calls changes budget line without changing a line of code. That kind of shift shows up at month end, not on a product page.
There is a trade-off, and it deserves stating. Fable 5.1 enforces 30-day data retention and opens no Priority Tier access, two constraints that bite in regulated environments.
A zero-retention option is available for customers who ask, and the enterprise safeguards announced for the autumn promise to leave clients control over monitoring methods. That framing extends the 25% cost reduction we flagged at the model’s release.
The internal comparison stays the most instructive one. Claude Opus 5 costs 5 dollars on input and 25 on output, half the rate, which leaves a genuine question open for high-volume workloads.
Sixty Percent Fewer Interruptions, and What That Unlocks
The second change in this Claude Fable 5.1 test is not measured in dollars. The model triggers roughly 60% fewer safeguard interventions per session than its predecessor.
In daily use, the difference shows up in how often a legitimate request hits a refusal. False positives on benign security questions fall sharply, which was the most persistent complaint aimed at the family.
One capability changes status outright. The model now agrees to identify vulnerabilities in source code, where the previous version refused, opening a defensive use case that had belonged to dedicated tooling.
⚠️ Important information
Fable 5.1 triggers roughly 60% fewer safeguard interruptions than Fable 5, and it now agrees to analyse code for vulnerabilities, something the previous version refused. That is the most concrete change in daily use.
That opening fits a consistent trajectory at the vendor, which had already pushed code review onto its heaviest models by handing code analysis to Mythos 5.
On performance, the widest gains land on scientific and agentic tasks. Terminal-Bench-Science 0.1 jumps from 24.7% to 52.6%, twenty-eight points, while AutomationBench moves from 17.1% to 31.4%.
Terminal-Bench 4.0 follows the same curve, from 42.0% to 55.8%, putting the model ahead of Opus 5 and its 52.3%. That is the score carrying the claimed record on command-line agentic coding.
Everywhere else, the gains are real but modest. CursorBench 3.2.0 adds three points to 73.4%, SWE-bench Pro moves from 80.0 to 81.2, and OSWorld 2.0 rises from 36.1% to 41.7%.
General reasoning improves at a comparable pace. Humanity’s Last Exam with tools goes from 63.8% to 65.0%, and the GDPval-AA v2 professional work index climbs from 1723 to 1853 points.
The average spread therefore tells two stories depending on the task. Three points on exercises already mastered, twenty-eight on those where the previous version fell down, which describes a model closing its gaps rather than lifting its ceiling.
More articles on Horizon
- GPT-6 Astra Launches as OpenAI Claims AGI Era
- OpenAI Makes Astra’s Reasoning Harder to Follow
- Gemini 3.8 Flash Trails Claude Opus 5 on Coding
How We Ran the Numbers on This Comparison
The figures in this test come from three sources laid side by side: the pricing grids Anthropic publishes for Fable 5.1, Fable 5 and Opus 5, the benchmark scores published at the model’s release, and an agentic workload scenario with 90% of context served from cache, the one behind the measured drop of roughly 20.9%.
The 25 to 45% range covers the other workload profiles, from cache-light to cache-heavy. Whenever the vendor is the only source of a figure and no independent measurement exists yet, the body of the article says so explicitly.
Our Verdict, and the Three API Breaks to Plan For
Migrating from Fable 5 is not transparent, and that is the main warning of this test. Three behaviours change at the interface level, with no workaround.
Forced tool use is gone. The values that compelled the model to call a tool, named or otherwise, are no longer supported, which breaks pipelines built on deterministic routing.
Thinking blocks are now bound to the conversation prefix. Any rewrite of the history mid-session raises an error, ruling out the on-the-fly context manipulation some architectures practised freely.
Thinking can no longer be switched off. It is modulated through an effort parameter, which moves the control lever but removes the option of cutting it entirely for simple calls.
None of the three is a configuration flag that can be flipped back. Each one demands a code change on the calling side, which turns what looks like a point release into a small migration project for anyone running Fable 5 in production today.
The context envelope, at least, moves nothing. One million tokens of input and 128,000 of output stay as they were, with a knowledge cutoff set at June 2026, so prompt architecture carries over untouched.
Our verdict splits into two clear cases. For autonomous multi-step research and long agentic sessions with heavy tool use, Fable 5.1 earns its place and the 0.25 cache read keeps the bill bearable.
For everyday coding, refactoring and analysis at high call volume, Opus 5 remains the rational pick. Half the rate for a three-point gap on standard coding benchmarks has no equivalent in the catalogue.
Fable 5 drops out of the comparison without argument. It costs exactly the same headline rate, loses on every published benchmark and pays four times as much per cache read, which leaves nothing to defend.
The Mythos 5.1 variant stays out of reach for most teams. It is restricted to vetted partner organisations in cybersecurity and life sciences, a framing that echoes what we saw when a model in the family fixed Claude better than twenty-eight researchers.
One claim sits beyond what this test can verify. The vendor states the model produced three novel scientific findings before launch, and that category of result only checks out with time.
Three Questions to Lock In the Essentials
What actually changes in the Fable 5.1 pricing grid?
Show answer
Nothing on the headline rates ($10 per million on input, $50 on output). Cache reads drop from $1 to $0.25 per million tokens, which cuts the bill by 25 to 45% depending on the workload.
Which model remains the rational pick for everyday coding?
Show answer
Opus 5. Half the rate of Fable 5.1 for a gap of roughly three points on standard coding benchmarks.
Which API breaks await a team migrating from Fable 5?
Show answer
Three: forced tool use disappears, thinking blocks are bound to the conversation prefix, and thinking now gets modulated through an effort parameter instead of switching off.
Follow the story on Horizon.
