Claude Fable 5.1 is now listed on AIxploria, the AI tools directory that catalogues thousands of AI sites across more than fifty categories. The new card sits at number one on the directory’s “Latest AI” feed, carries a Gold Verified badge, is priced “Paid”, and files Anthropic’s frontier model under LLM models and SuperTools. For a model that shipped on 1 September 2026, that is a fast arrival on one of the most-browsed AI directories on the web.
The listing is short, and it is not wrong. It calls Claude Fable 5.1 “Anthropic’s LLM model for coding, research, and long, complex tasks”, notes improved performance and reduced costs on agent-based workloads, and adds that it “produces far fewer false security alerts”. Every one of those claims checks out against Anthropic’s own documentation. What a directory card cannot do is tell you what the model costs per million tokens, which benchmarks moved, which three things break if you already call its predecessor, or whether it is the right model for your team at all.
That is the job of this article. Our AI models and tools hub tracks releases like this one, and we have run the same exercise on the previous arrivals in this series: Claude Academy, Gemini 3.5 Transcribe, and, only yesterday, Gemini 3.8 Flash — which happens to sit directly below Claude Fable 5.1 at number two on the same feed.
Every figure quoted here comes from Anthropic’s launch announcement, the Claude Platform model documentation and pricing pages, or the AIxploria listing itself, all checked on 3 September 2026. Where a number is our own arithmetic on those published figures, we say so.
Table of contents
- What Claude Fable 5.1 Actually Is
- Claude Fable 5.1 on AIxploria: Reading the Listing Properly
- The Benchmark Numbers Behind Claude Fable 5.1
- What Claude Fable 5.1 Costs, and Where the Saving Actually Hides
- Where You Can Actually Run Claude Fable 5.1
- The Safety Story: Fewer False Alarms, More Refusals
- Should Your Business Switch to Claude Fable 5.1?
- What a Directory Listing Cannot Tell You
- Frequently Asked Questions About Claude Fable 5.1
- References
What Claude Fable 5.1 Actually Is
Claude Fable 5.1 is Anthropic’s frontier model: the tier above Opus, sold for demanding reasoning and long-horizon agentic work rather than for everyday chat. It replaces Claude Fable 5, which shipped roughly three months earlier, and it arrives at exactly the same input and output prices.
A model built for work that runs for hours
The pitch is endurance. Claude Fable 5.1 is designed for jobs that stretch across hours and several applications at once — clearing a ticket backlog, driving a browser, working through a repository. Anthropic’s framing is that you hand it a task list in the morning and it plans the work, recovers when a step fails, and reports back. Thinking is always on, in adaptive mode, and the depth is steered by an effort setting rather than a fixed token budget.
The customer anecdote doing the rounds comes from the investment firm Millennium. A particular piece of code had an extremely rare crash — roughly one in a million runs — that nobody on the team had explained in four to five years. Claude Fable 5.1 traced it to a bug in an external library. Jane Street reports that it solves more of their internal coding problems than Fable 5 did, and Cognition said it was moving its Opus 5 traffic in Devin across on launch day.
The Mythos 5.1 twin, and why it exists
Shipped alongside it is Claude Mythos 5.1, which is the same underlying model with more permissive safeguards for cybersecurity and life-sciences professionals. It is invitation-only, distributed through Anthropic’s Project Glasswing and its Cyber and Life Sciences Verification Programs, and restricted to vetted organisations. It shares Claude Fable 5.1’s specifications and pricing exactly. If you are reading a benchmark table and see a Mythos row scoring higher, that is the safeguard configuration talking, not a different model.
How it sits against the rest of the line-up
This is the part the directory card leaves out entirely. Claude Fable 5.1 costs twice what Opus 5 costs on both input and output, and Anthropic’s own guidance is to start with Opus 5 for most workloads and reach for the frontier tier only when your evaluations on Opus 5 at higher effort still fall short.
| Model | Context | Max output | Price / MTok | Default effort | Knowledge cutoff |
|---|---|---|---|---|---|
| Claude Fable 5.1 | 1M | 128K | $10 / $50 | high | Jun 2026 |
| Claude Opus 5 | 1M | 128K | $5 / $25 | high | May 2026 |
| Claude Sonnet 5 | 1M | 128K | $2 / $10 | high | Jan 2026 |
| Claude Haiku 4.5 | 200K | 64K | $1 / $5 | not supported | Feb 2025 |
A million tokens is roughly 555,000 words on the current tokenizer, which is the specification most people quote and fewest people use. Anthropic commits to keeping the model available until at least 1 September 2027.
Claude Fable 5.1 on AIxploria: Reading the Listing Properly
AIxploria is a discovery directory, not an evaluator. Its card is how a large audience will first meet this model, so it is worth reading closely — including the parts that are easy to misread.
What the card actually says
The entry carries a Gold Verified badge, a “Paid” pricing tag, and three category tags: Latest AI, LLM models and SuperTools. It shows 128 upvotes, a trend counter reading 210,749, and a tooltip that says “This is a popular tool!”. Its suggested similar tools are Claude Fable 5, Grok 4.6 and Claude Opus 4.8. The write-up underneath is unusually detailed for a directory — it quotes benchmark figures, the cache read price and the Mythos distinction, and it is accurate on all three.
| Listing field | Value on 3 September 2026 | What it tells you |
|---|---|---|
| Position | No. 1 on the Latest AI feed | Recency, not quality |
| Badge | Gold Verified | The entry was checked by a human |
| Pricing tag | Paid | Correct: no free tier access |
| Star rating | 4.4 out of 5 | Computed from 12 votes |
| Upvotes | 128 | Launch-week attention |
| Trend counter | 210,749 | Views, not evaluations |
Read the vote count, not the star rating
Here is the detail almost nobody checks. That 4.4 out of 5 is computed from twelve votes — the underlying figure buried in the page markup is 4.4166666666667. The rating is not wrong, but it is not evidence either. Twelve people means one vote is worth 8.3 per cent of the sample, so a single disappointed rater moves the visible score. The listed alternative GPT-5.5 shows 4.6 from fourteen votes, which is the same problem with a different number on it.
We have flagged the same pattern on every AIxploria listing we have covered. Treat the badge as a visibility signal and the star as noise until the sample grows into the hundreds.
Number one on new arrivals, absent from the category
There is a quieter detail worth noting. Claude Fable 5.1 tops the Latest AI feed, but on the day we checked, the directory’s LLM models category page still showed Claude Fable 5 in the top slot and did not list version 5.1 at all. New arrivals and category rankings are two different lists updating on two different clocks. If you browse AIxploria by category rather than by date — which is how most people use a directory — you would not have found this model yet.
The Benchmark Numbers Behind Claude Fable 5.1
Anthropic published a full comparison table at launch, pitting Claude Fable 5.1 against Fable 5, Opus 5 and OpenAI’s GPT-5.6 Sol. The numbers are the vendor’s own, which is worth remembering, but they are specific and they are checkable.
Terminal-Bench-Science: the score that doubled
The headline is Terminal-Bench-Science 0.1, a test of autonomous scientific research. Claude Fable 5.1 scores 52.6 per cent against Fable 5’s 24.7 per cent — 27.9 percentage points higher, and 2.13 times the earlier result by our arithmetic on those two figures. Opus 5 records 29.0 per cent and GPT-5.6 Sol 22.4 per cent on the same test. A doubling on any benchmark is unusual within a point release.
Agentic coding, computer use and knowledge work
On Terminal-Bench 4.0, the agentic coding test, Claude Fable 5.1 records 55.8 per cent against Fable 5’s 42.0 per cent — a gain of 13.8 points on the published figures — with Mythos 5.1 reaching 60.9 per cent under its looser safeguards. On CursorBench 3.2.0 it posts 73.4 per cent, the strongest of the four models compared. AutomationBench nearly doubles, from 17.1 to 31.4 per cent.
Vision improved more quietly. The model reads charts and tables buried inside PDFs, which matters more for finance and legal teams than any coding score, and the effort level can now be changed mid-conversation.
The full published comparison
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
| Humanity’s Last Exam (no tools) | 60.9% | 57.8% | 56.6% | not published |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% | not published |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| GDPval-AA v2 (knowledge work) | 1853 | 1723 | 1824 | 1711 |
What Claude Fable 5.1 Costs, and Where the Saving Actually Hides
This is where the AIxploria card is most useful and least specific. It says costs are reduced. They are — but not in the place most buyers look.
The headline prices did not move
Input is $10 per million tokens and output is $50 per million tokens, both unchanged from Fable 5. Output therefore costs five times input, which is the ratio that quietly decides most bills. Against Opus 5 at $5 and $25, Claude Fable 5.1 is exactly double on both sides of the ledger.
The cache read price did move
Prompt caching reads normally cost 10 per cent of the base input price. On Claude Fable 5.1 they cost 2.5 per cent — $0.25 per million tokens rather than $1.00, a 75 per cent cut. For any workload that re-sends a large stable prefix on every turn, which is what an agent loop is, that is the number that matters.
Where the 25 and 45 per cent figures come from
Anthropic quotes roughly 25 per cent lower real costs on typical workloads and up to 45 per cent on heavily agentic runs. Both figures are downstream of the cache change, not of any list-price cut — which is why a team that has never configured prompt caching will see none of it. If your bill does not move after switching, caching is the first thing to audit. Our LLM API pricing comparison sets out how the tiers stack up more broadly.
| Billing line | Claude Fable 5.1 | Note |
|---|---|---|
| Input | $10 / MTok | Unchanged from Fable 5 |
| Output | $50 / MTok | Unchanged from Fable 5 |
| Cache read | $0.25 / MTok | Down 75 per cent |
| Cache write, 5 minutes | $12.50 / MTok | Standard 1.25x input |
| Cache write, 1 hour | $20 / MTok | For long-lived prefixes |
| Batch API | 50% discount | Applies to input and output |
Where You Can Actually Run Claude Fable 5.1
Access runs through the paid Claude plans — Pro, Max, Team and Enterprise — or through the API, billed by the token. The free tier does not include it, which is what the directory’s “Paid” tag is telling you.
Plans, API and the cloud platforms
On the API the model identifier is claude-fable-5-1. It is also served on Amazon Bedrock as anthropic.claude-fable-5-1, and on Google Cloud, Microsoft Foundry and Claude Platform on AWS under the bare identifier. Subscribers on a paid plan get Claude Fable 5.1 at no extra charge within their plan’s usage limits — the same limits we wrote about when the Claude iOS app started warning users about Max effort mode.
Three things break if you already call Fable 5
Anthropic lists three breaking changes for anyone migrating. Forced tool use now returns an error. Earlier models cannot read this model’s thinking blocks. And editing an earlier turn in a conversation invalidates the thinking blocks attached to it. None is exotic, but all three fail at runtime rather than at review time, which is the worst way to find them.
And five things that are simply new
The additive changes are per-message effort, turn-scoped system messages, readable progress updates between tool calls, the lower cache read price, and content provenance — an invisible watermark applied to outputs in line with the EU AI Act, with a detection API in private preview. The first three are in beta.
The Safety Story: Fewer False Alarms, More Refusals
The AIxploria description singles out “far fewer false security alerts”, and that is a real, quantified change rather than marketing.
Sixty per cent fewer false positives in security work
Anthropic reports around 60 per cent fewer false positives on cybersecurity work in Claude Code. The policy line is now explicit: the model may hunt for software vulnerabilities, but it will not write exploits for them. On biology and medical questions the improvement is larger still — roughly 85 per cent fewer false positives on elementary questions, with research-level queries routed to the Opus models instead.
But refusal rates are higher overall
The trade-off is on the listing’s own “cons” list, and it is honest: refusal rates run higher than on earlier Claude models, because blocking classifiers sit in front of sensitive cybersecurity and life-sciences queries. Anthropic reports comparable refusal rates to Mythos 5, Sonnet 5 and Opus 5 on genuinely malicious requests, and says the model is its most robust yet on external prompt-injection benchmarks. If your work sits near either sensitive domain, budget for refusals and design a fallback path.
What the science results suggest
The launch materials include three concrete research results: protein binder designs achieving ten times higher binding affinities with roughly a 50 per cent hit rate across twelve targets, against a typical 10 to 15 per cent; a Venus elevation map at 2 to 3 kilometre detail where 10 to 20 kilometres was previous practice; and GPU kernel optimisation delivering up to a 2.5 times speed-up. These are not benchmarks, and they are not reproducible from the announcement — but they do explain the Terminal-Bench-Science jump.
Should Your Business Switch to Claude Fable 5.1?
Probably not wholesale, and Anthropic effectively says so itself. Its documented guidance is to start with Opus 5 and move up only when your own evaluations at higher effort fall short.
Start low, not high
The single most useful habit is to start at low or medium effort, where Claude Fable 5.1 already matches its predecessor at materially lower cost, and raise the level only where measurement shows headroom. Claude Code defaults to high effort; Claude Cowork and the consumer apps default to medium. Because effort drives token spend directly, the default you inherit is a budget decision nobody made deliberately.
Which workload belongs on which model
| Workload | Sensible default | Why |
|---|---|---|
| Long autonomous coding runs | Claude Fable 5.1 | Best published agentic scores; cache saving compounds |
| Multistep research and analysis | Claude Fable 5.1 | The Terminal-Bench-Science gap is the largest of any test |
| Dense PDF, spreadsheet and slide work | Claude Fable 5.1 | Improved reading of charts and tables inside documents |
| General business questions and drafting | Claude Opus 5 | Half the price for work that does not need the frontier tier |
| High-volume production traffic | Claude Sonnet 5 | One fifth the input price; ample for routine tasks |
| Classification and simple extraction | Claude Haiku 4.5 | Cheapest and fastest; frontier reasoning is wasted here |
The honest downside
Claude Fable 5.1 is slower than the Opus tier by Anthropic’s own latency rating, it is priced for heavy work rather than casual chat, and it refuses more. For a team whose AI spend is mostly chat and drafting, moving everything to the frontier tier is a straightforward way to double a bill for no measurable gain. Our write-up of the Claude 5 family covers how the tiers were originally positioned.
What a Directory Listing Cannot Tell You
The Claude Fable 5.1 card on AIxploria does its job well. It is accurate, it is detailed, and it appeared within two days of launch. But three limits are structural, not faults of this particular entry.
A directory measures attention, not fitness
A trend counter of 210,749 and 128 upvotes measure how many people looked, not how many deployed successfully. The Latest AI feed ranks by recency. The star rating rests on a dozen votes. None of those signals is designed to answer “is this the right model for my workload”, and none of them should be asked to.
The category comparison is apples to oranges
The same category page that will eventually list Claude Fable 5.1 also lists open-weight models, consumer chat products and single-purpose utilities. Placing a frontier model priced at $50 per million output tokens next to a free tool is exactly the comparison most buyers are least equipped to make. That gap is why we keep writing these up rather than linking to the card.
What we checked, and what we did not
Everything in this article comes from Anthropic’s published announcement and documentation or from the live AIxploria listing on 3 September 2026. We have not independently reproduced any benchmark. Vendor-published scores are a starting point for your own evaluation, never a substitute for it.
Frequently Asked Questions About Claude Fable 5.1
Is Claude Fable 5.1 free?
No. It requires a paid Claude plan — Pro, Max, Team or Enterprise — or an API account billed by the token. The free tier does not include it. Subscribers get it at no extra charge within their plan’s usage limits.
What is the difference between Claude Fable 5.1 and Mythos 5.1?
They are the same model with different safeguard levels. Claude Fable 5.1 is generally available. Mythos 5.1 keeps fuller cybersecurity and life-sciences capability and is limited to vetted organisations through Anthropic’s verification programmes.
Does Claude Fable 5.1 replace Claude Opus 5?
No. Anthropic positions Opus 5 as the sensible default and Claude Fable 5.1 as the tier above it, for demanding reasoning and long-horizon agentic work. Opus 5 remains available at half the price.
How large is the Claude Fable 5.1 context window?
One million tokens, with a maximum output of 128,000 tokens per synchronous request. A million tokens is roughly 555,000 words on the current tokenizer.
Why does Claude Fable 5.1 refuse some requests?
Blocking classifiers filter sensitive cybersecurity and life-sciences queries, and refusal rates run higher than on earlier Claude models. The model may look for software vulnerabilities but will not write exploits for them.
When was Claude Fable 5.1 released?
1 September 2026, alongside Claude Mythos 5.1. Anthropic commits to keeping it available until at least 1 September 2027.
References
Introducing Claude Fable 5.1 and Claude Mythos 5.1
Claude Fable 5.1 model overview, Claude Platform Docs
What’s new in Claude Fable 5.1
Claude Fable 5.1 and Claude Mythos 5.1 system card
Effort levels on the Claude API
Claude Fable 5.1 listing on AIxploria
Claude Fable 5.1 is now available on AWS
Anthropic’s Claude Fable 5.1 promises better coding and research for less
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.