Claude Fable 5.1 Review
When Anthropic shipped Claude Fable 5 in June, I wrote that the price doubling was the headline, not the benchmarks. Ten dollars per million input tokens and fifty per million output is twice Opus 5, and I said at the time that the interesting question was not whether the model was better but whether it was better by enough to justify running it on anything that loops.
Fable 5.1 landed yesterday, September 1, alongside Mythos 5.1. The base price did not move. It is still $10 in, $50 out.
The cache read price moved a lot. It went from $1.00 per million to $0.25 per million, a 75% cut.
That sounds like a footnote. It is not. For anyone running agentic workloads, where the same large context gets re-read on every single turn of a loop, cache reads are frequently the majority of the bill. Anthropic's own claim, based on four weeks of internal August usage, is that this lowers typical bills by about 25% and highly agentic ones by up to 45%.
So this is a release where the model got better and the thing that was stopping people from using it got cheaper. Let me go through both halves.
What Fable 5.1 and Mythos 5.1 Actually Are
Same structure as last time, which is worth restating because it still confuses people.
Fable 5.1 and Mythos 5.1 are the same underlying model with different levels of safeguards. Fable 5.1 is the generally available version with Anthropic's production safeguards on. Mythos 5.1 is the same brain with reduced restrictions on cybersecurity and biology research work, available only through verification programs: the Cyber Verification Program for defensive security professionals, and the Life Sciences Verification Program run in partnership with the US government. Both are currently US-only, with international expansion described as under discussion.
The API model ID is claude-fable-5-1. It shipped day one on Bedrock, Vertex AI, Azure, the Claude API, plus Claude Code, Claude Enterprise, Claude Cowork, and Claude.ai. No waitlist, no staged rollout, which continues to be the release pattern that separates Anthropic from most of the field.
I wrote a longer piece on what the Mythos tier means when it was introduced, and the structure has held: capability is not the gate, verification is.
The Benchmarks That Moved
Here is the table from the announcement, with Fable 5 and Opus 5 for comparison.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% |
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% |
| AutomationBench | 31.4% | 17.1% | 26.9% |
| Humanity's Last Exam (no tools) | 60.9% | 57.8% | 56.6% |
| GDPval-AA v2 | 1853 | 1723 | 1824 |
Mythos 5.1 scores 60.9% on Terminal-Bench 4.0, five points above the safeguarded version, which is the clearest public number anyone has put on what production safeguards cost in raw capability on a coding benchmark.
Read the shape of that table rather than the individual numbers. The gains cluster hard in agentic terminal work and agentic science. Terminal-Bench 4.0 went from 42.0 to 55.8, which is a 33% relative improvement. Terminal-Bench-Science more than doubled. AutomationBench nearly doubled. Meanwhile Humanity's Last Exam moved three points and CursorBench moved three points.
That is not a model that got smarter. That is a model that got better at not giving up halfway through a long task. Anthropic's own framing is that it is better at sustained multi-step work and better at diagnosing root causes rather than applying shortcuts, and the benchmark spread backs that up.
One thing worth flagging because it is circulating badly: the "Fable 5.1 scores 95% on SWE-bench Verified" number you will see in roundups comes from third-party leaderboards, not from Anthropic. Anthropic reports SWE-bench Pro at 81.2 in the system card. Both numbers can be true and they measure different things, but attribute them correctly.
There is also an honest asterisk that cuts the other way. Both Fable models were scored with production safeguards switched on, and on tasks where a classifier intervened they took zeros. The published Fable scores are conservative rather than inflated. That is an unusual thing for a lab to do to its own launch numbers and it deserves credit.
What "Better at Sustained Work" Feels Like
Benchmarks are the part of a review I trust least, so here is the part I actually did.
I gave it the kind of task that used to break agents: a real refactor across a codebase, spanning several files, where the correct answer requires reading code you were not pointed at. Specifically, tracing why a build was failing three layers away from where the error surfaced.
The Fable 5 behavior on this class of task was decent and had a characteristic failure. It would find a fix, apply it, watch the error change shape, and then start chasing the new error rather than backing up and asking why the first fix did not work. You end up with four plausible-looking patches and a codebase that is worse than when you started.
5.1 backs up more. Twice in a day it did the thing I actually wanted, which was stop, say the earlier assumption was wrong, and re-read the file it had skimmed. That is the "root causes rather than shortcuts" claim showing up in behavior rather than in a chart, and it is the single most useful improvement in the release for anyone doing agentic coding day to day.
It is not magic. It still confidently wrote a test that asserted the buggy behavior once, which is the most annoying failure mode any coding model has. But the direction is right.
The Cache Price Cut Is the Actual Story
Now the part that changes what you can afford to build.
Most Claude models price cache reads at 10% of the base input rate. Fable 5's input is $10 per million, so cache reads were $1.00. Fable 5.1 prices cache reads at $0.25 per million, which is 2.5% of base input, not 10%. That is a different multiplier from the rest of the lineup, deliberately.
Why this matters more than the raw percentage suggests: in an agentic loop, the expensive thing is not generating tokens. It is re-reading the same enormous context on every turn. Your system prompt, your tool definitions, the files the agent has already loaded, the conversation so far. Fifteen turns into a session, you are paying to re-read a context that has barely changed, fifteen times.
I broke the arithmetic down in detail in the guide to what AI agents actually cost, and the conclusion there was that cache reads dominate the bill for anything that loops, which is why prompt caching is not an optimization you do later, it is the thing you architect around from the start.
Anthropic's numbers: about 25% lower bills for typical workloads, up to about 45% for highly agentic ones, based on four weeks of their own August usage. Take vendor-supplied savings numbers with the usual salt, but the mechanism is real and the direction is not in dispute. A 75% cut on the line item that dominates agentic spend produces a large cut in agentic spend.
The strategic read is more interesting than the discount. Anthropic is repricing the specific thing that makes long-running agents expensive, at the same moment it ships a model whose biggest gains are in long-running agentic work. That is not a coincidence, it is a bet. They are betting that the workload that grows is agents running for hours, and they are removing the cost objection to running the expensive model on it.
Should You Switch From Fable 5?
If you are already on Fable 5: yes, immediately, and there is nothing to think about. Same base price, better model, cache reads cost a quarter of what they did. Change the model ID.
If you are on Opus 5 and stayed there because Fable doubled the price: this is the release where that calculation gets genuinely close. Opus 5 is still half the base rate. But if your workload is agentic and cache-heavy, Fable 5.1's effective cost per session may land nearer to Opus 5 than the sticker price implies, and Fable 5.1 beats Opus 5 by 3.5 points on Terminal-Bench 4.0 and by 23 points on Terminal-Bench-Science.
Do the arithmetic on your own traffic instead of trusting either of us. Pull a week of usage, split it into fresh input, cache writes, cache reads, and output, and reprice it both ways. If cache reads are over half your input volume, Fable 5.1 has probably just become your cheapest good option. If you are doing mostly one-shot calls with short prompts, none of this touches you and Opus 5 remains the sane default.
And if you are routing between models rather than picking one, which is what most serious production setups do now, this changes the routing table rather than the winner. Model routing with fallbacks exists precisely so a price change like this is a config edit instead of a migration.
The Safeguards Changes Are Better Than Expected
Three numbers from the announcement that matter if you have ever been refused something reasonable.
Cybersecurity filtering has 60% fewer false positives, and vulnerability discovery is now permitted while exploit generation is not. Biology safeguards trigger 85% less often for benign requests on elementary and medical topics. And Anthropic describes it as their most robust model to date on external prompt injection benchmarks.
The false positive numbers are the ones I care about. A safety system that refuses legitimate security work is not safe, it is just unhelpful in a way that pushes people to models with worse safeguards. Cutting false positives by 60% while keeping malicious-request refusal rates comparable is the right trade, and it is measurable in a way that "we take safety seriously" is not.
If you are building anything that takes untrusted input into a model, the prompt injection claim is worth verifying against your own threat model rather than taking on faith. I wrote about defending against prompt injection and the honest position has not changed: a more robust model reduces your exposure and does not eliminate it, and architecture is still doing most of the work.
One more change that will bite someone: new API accounts cannot manually edit prior context in multi-turn conversations. This is an anti-distillation measure, aimed at blocking documented model extraction techniques. If you had a legitimate reason to rewrite conversation history, and there are a few, check whether your account is affected before you upgrade.
The Bigger Picture
Two releases in a row now, Anthropic has shipped a model where the headline capability gains are concentrated in agentic work and the pricing changes are aimed at making agentic work affordable. The consumer-facing chat benchmarks moved three points. The "can this thing run unattended for two hours" benchmarks moved thirty percent.
That tells you what they think the product is. Not a chatbot that answers questions. A worker that runs a loop.
Whether that bet pays off depends on something outside Anthropic's control, which is whether the rest of us build systems that can actually supervise a model running unattended for two hours. Cheaper cache reads remove one obstacle. They do not remove the observability problem, which is still the thing that turns a promising agent into an incident nobody can reconstruct.
The model got better at not giving up. We still have to get better at watching it.
Frequently Asked
When was Claude Fable 5.1 released? September 1, 2026, alongside Claude Mythos 5.1, on the Claude API, Bedrock, Vertex AI, Azure, Claude Code, Claude Enterprise, Claude Cowork, and Claude.ai.
How much does Claude Fable 5.1 cost? $10 per million input tokens and $50 per million output tokens, unchanged from Fable 5. Cache reads are $0.25 per million, down 75% from $1.00.
What is the difference between Fable 5.1 and Mythos 5.1? The same model with different safeguard levels. Mythos 5.1 has reduced restrictions on cybersecurity and biology research and is available only through the Cyber Verification Program or Life Sciences Verification Program, currently US-only.
Is Fable 5.1 better than Opus 5? On agentic and terminal benchmarks, yes, by a clear margin. It is also twice the base price. Whether it is better for you depends on how cache-heavy your workload is.
What is the model ID? claude-fable-5-1.
Did SWE-bench scores actually hit 95%? Anthropic reports SWE-bench Pro at 81.2 in the system card. The 95% figure comes from third-party leaderboards measuring SWE-bench Verified, which is a different benchmark.
If you run agents, go pull last week's usage and reprice it before you do anything else. That number, not the benchmark table, is what should decide this for you.
Top comments (0)