THE AI RANKINGS

Anthropic

Claude Fable 5.1

Provider
Anthropic
Status
Current
Context
1,000,000 tok
Price
$10 / $50 /MTok
Knowledge
2026-06

Claude Fable 5.1 is Anthropic’s current Mythos-class flagship — the tier that sits above the Opus class — released on 1 September 2026, just under three months after Fable 5. It is the first model to pass 65 on the independent Artificial Analysis Intelligence Index, scoring 66 against Opus 5’s 63 and Fable 5’s 62, and it takes the top of that index outright.

The launch’s own framing is about cost rather than capability: token prices are unchanged at $10/$50 per million, and the entire advertised saving comes from cutting cache reads 75% to $0.25 per million. That is a real cut, and it is also the claim that most needs checking — see the cost story below, because Artificial Analysis measures Fable 5.1 costing more per completed task than Fable 5 did.

Fable 5.1 shipped alongside Mythos 5.1, the same underlying model with its cyber and bio safeguards relaxed, restricted to vetted US organisations. This page covers Fable 5.1, the model you can actually buy.

Quick specs

ProviderAnthropic
TierMythos-class (above Opus)
Released1 September 2026
StatusAvailable — paid Claude plans and API accounts
API model IDclaude-fable-5-1
Context window1,000,000 tokens
Max output128,000 tokens
Knowledge cutoffJune 2026
Input price$10.00 / MTok
Output price$50.00 / MTok
Cache reads$0.25 / MTok (down 75% from $1.00)
AA Intelligence Index66 — 1st, the highest score the index has recorded
Terminal-Bench-Science 0.152.6% (Fable 5: 24.7%)
Terminal-Bench 4.055.8% (Opus 5: 52.3%)
SWE-benchNot published — no figure of any kind
Best forLong-horizon agentic work, agentic scientific research, frontier knowledge work
Limitations2x Opus 5’s token price and higher cost per task; no SWE-bench figure; verbose and slow

CHECK STATUS →

What actually changed

Anthropic positions Fable 5.1 as a step up on “coding, knowledge work, and long-running problem-solving.” Four changes matter in practice.

Agentic science is the standout. On Terminal-Bench-Science 0.1 — an agentic scientific-research eval — Fable 5.1 scores 52.6% against Fable 5’s 24.7%. That is more than double, and it is the largest single jump in the launch table. It also clears Opus 5 (29.0%) and GPT-5.6 Sol (22.4%) by a wide margin.

Cache reads got much cheaper. $1.00 to $0.25 per million tokens. On workloads that re-read a large stable prefix — long agent loops, big repository contexts — this is where the money goes, which is why Anthropic’s own estimate of the saving is larger for agentic work (up to ~45%) than for typical use (~25%).

The safeguards fire less often. Anthropic reports Claude Code users should see roughly 60% fewer cyber-safeguard interventions per session than on Fable 5, and that the biology safeguards fire 85% less often on benign elementary-biology and medical questions. The scope also shifted: Fable 5.1 “can now be used to discover software vulnerabilities — though not to develop exploits for them.” This addresses the loudest complaint about the redeployed Fable 5, whose tightened classifier tripped on routine coding.

Effort now buys more. Anthropic says that at Low or Medium effort Fable 5.1 matches or beats Fable 5 at substantially lower cost — which is the practical form of the price claim. Product defaults reflect it: High in Claude Code, Medium in Claude Cowork and on Claude.ai.

Benchmark performance

All figures below are Anthropic-reported and vendor-run, from the 1 September launch comparison, except where marked independent.

Agentic science and coding

BenchmarkFable 5.1Fable 5Opus 5GPT-5.6 Sol
Terminal-Bench-Science 0.152.6%24.7%29.0%22.4%
Terminal-Bench 4.055.8%52.3%37.3%
CursorBench 3.2.073.4%70.5%

Read the Terminal-Bench 4.0 row carefully. Anthropic’s table compares Fable 5.1’s 55.8% against Mythos 5’s 42.0%, not against Fable 5, which has no published figure on that version — so the headline “55.8 against 42.0” is a comparison with a different, restricted model. The restricted Mythos 5.1 leads the same test at 60.9%.

Reasoning, knowledge work and computer use

BenchmarkFable 5.1Fable 5
Humanity’s Last Exam (no tools)60.9%57.8%
Humanity’s Last Exam (with tools)65.0%63.8%
GDPval-AA v2 (knowledge work)1853 Elo1723 Elo
OSWorld 2.0 (partial)77.9%72.9%
OSWorld 2.0 (strict)41.7%36.1%
AutomationBench31.4%17.1%

AutomationBench nearly doubles and OSWorld moves several points on both scorings — consistent with Anthropic’s long-horizon-agent pitch. The gains on Humanity’s Last Exam are real but modest: three points without tools, roughly one with.

The independent number

Artificial Analysis scores Fable 5.1 at 66 on its Intelligence Index (adaptive reasoning, max effort, default fallback), 1st of 192 models tracked and the first score above 65 the index has recorded. Opus 5 sits at 63 and Fable 5 at 62. This is what earns the placement: a vendor table alone would cap the model in our Strong tier.

The same measurement carries the caveats. Fable 5.1 generated 140 million output tokens running the index against a median of 71 million — roughly double the field — at 66 tokens per second (slower than average) with a 285-second time to first token. It cost $8,523 to run the index, the highest Artificial Analysis has recorded.

What Anthropic did not publish

No SWE-bench figure of any kind — neither Pro nor Verified — appears in Anthropic’s launch comparison. For a model pitched on coding, from the lab whose predecessor set the SWE-bench Pro record, that is a conspicuous gap. The consequence for this site’s board is concrete: the highest SWE-bench Pro score we carry is still Fable 5’s 80.3%, a predecessor’s number, and we cannot say Fable 5.1 beats it.

There is a second gap worth naming. Anthropic’s launch table puts Fable 5.1 at 1853 Elo on GDPval-AA v2; Anthropic’s own Opus 5 System Card reports 1861 for Opus 5 on the same version. The two figures come from different Anthropic tables and are not a clean head-to-head, but on Anthropic’s published knowledge-work numbers Fable 5.1 does not clear the model that costs half as much.

The cost story: cheaper tokens, dearer tasks

This is the claim to test, because the launch leads with it.

Input / MTokOutput / MTokCache read / MTokAA cost per index task
Claude Fable 5.1$10.00$50.00$0.25$3.69 ($3.76 at max effort)
Claude Fable 5$10.00$50.00$1.00$3.08–$3.25
Claude Opus 5$5.00$25.00$0.50$2.34

Anthropic’s estimate — about 25% off a typical workload, up to about 45% off highly agentic work — is a statement about cache-heavy workloads, and on those it is credible: nothing else changed. But Artificial Analysis, running the same standard suite it ran against Fable 5, measures Fable 5.1 at $3.69 per completed task against Fable 5’s $3.08–$3.25 — about 20% more. The reason is verbosity: at roughly double the median output tokens, Fable 5.1 spends the savings and then some on its own reasoning.

Both things are true, and which one you get depends on your workload. If you re-read a large cached prefix on every turn, the cut is real money. If your bill is dominated by output tokens, expect to pay more than you did on Fable 5 — and roughly 60% more per task than on Opus 5.

Safeguards, watermarks and enterprise privacy

Fable 5.1 keeps the Mythos-class safeguard architecture but retunes it. Anthropic reports ~60% fewer cyber-safeguard interventions per session in Claude Code versus Fable 5, and biology safeguards firing 85% less often on benign elementary-biology and medical questions. The permitted scope widened in one specific direction: vulnerability discovery is now allowed, exploit development is not.

Two things are new at the platform level:

Mythos-class models remain “Covered Models” with mandatory 30-day data retention; there is no zero-data-retention option.

How to access Claude Fable 5.1

Fable 5.1 is available to anyone on a paid Claude plan or with an API account, as claude-fable-5-1 across Claude.ai, the Claude API, Claude Code, Claude Cowork and Claude Enterprise, plus Amazon Bedrock (anthropic.claude-fable-5-1), Google Cloud Vertex AI and Microsoft Foundry. Anthropic commits to retiring it no sooner than 1 September 2027.

from anthropic import Anthropic
client = Anthropic()

message = client.messages.create(
    model="claude-fable-5-1",
    max_tokens=4096,
    effort="medium",  # Anthropic: matches or beats Fable 5 at low/medium, for less
    messages=[{"role": "user", "content": "Your prompt here"}],
)

Integration notes for Mythos-class models: adaptive thinking is always on and cannot be disabled, the raw chain-of-thought is never returned, and the API default effort is high while Claude Cowork and Claude.ai default to medium. Supported features include the memory tool, code execution, programmatic tool calling, context editing, compaction, task budgets and vision.

For the restricted sibling with relaxed cyber and bio safeguards, see Mythos 5.1.

How Claude Fable 5.1 compares

vs Claude Opus 5

Opus 5 is the tier below at half the token price ($5/$25) and, on Artificial Analysis’ measurement, $2.34 per task against Fable 5.1’s $3.69. Fable 5.1 leads the independent index (66 vs 63) and clears Opus 5 comfortably on agentic science (52.6% vs 29.0%) and by three points on Terminal-Bench 4.0 (55.8% vs 52.3%). But Opus 5 has a published SWE-bench Pro figure (79.2%) and Fable 5.1 has none, and Anthropic’s own GDPval-AA v2 numbers put Opus 5 marginally ahead. Opus 5 remains our best-overall pick for buyable work; Fable 5.1 earns its premium on long-horizon agentic and scientific tasks, not across the board.

vs Claude Fable 5

Same price per token, better on every published benchmark they share, and much cheaper on cache reads. The honest catch is verbosity: about 20% more per completed task on the independent measurement. Fable 5 keeps one thing 5.1 cannot claim — a published SWE-bench Pro score, and the highest one on our board at 80.3%.

vs Claude Mythos 5.1

Same underlying model, different safeguard levels. Mythos 5.1 leads Terminal-Bench 4.0 at 60.9% against 55.8% because it is not routing or declining the same requests, but it is restricted to vetted US organisations through the Cyber Verification Program and the Life Sciences Verification Program. If you can buy it, you are buying Fable 5.1.

vs GPT-5.6 Sol

On Anthropic’s table Fable 5.1 leads GPT-5.6 Sol clearly on both terminal benchmarks (52.6% vs 22.4% on science, 55.8% vs 37.3% on coding), and it leads on the independent index (66 vs 61). Sol is far cheaper — $4/$20 on its promotional rate to 21 November — and carries a METR-flagged reward-hacking concern. See best AI models and best AI for coding for where this lands across the field.

Known limitations

No SWE-bench figure of any kind. Neither Pro nor Verified. On the site’s coding-weighted methodology this is the single largest reason Fable 5.1 does not simply take every coding pick.

Cost per task went up, not down. Despite the cache-read cut, Artificial Analysis measures $3.69 per index task against Fable 5’s $3.08–$3.25, driven by roughly double the median output tokens.

Slow. 66 output tokens per second and a 285-second time to first token on the independent measurement — at the high end for reasoning models, and a real constraint on interactive use.

Vendor-run headline benchmarks. Every figure except the Artificial Analysis index and cost measurements is Anthropic’s own, and the Terminal-Bench 4.0 comparison is drawn against Mythos 5 rather than Fable 5.

Enterprise Frontier Safeguards is announced, not shipped. A phased rollout begins in autumn 2026; it is not something you can deploy today.

Mandatory 30-day data retention. Mythos-class models are “Covered Models” — no zero-data-retention option.

Twice Opus 5’s token price. $10/$50 against $5/$25, for a model that does not lead Anthropic’s own published knowledge-work numbers.

Version history

VersionReleasedKey points
Claude Fable 5.1 / Mythos 5.11 Sep 2026AA Intelligence Index 66 (1st); Terminal-Bench-Science 52.6%; cache reads cut 75% to $0.25; no SWE-bench figure
Claude Fable 5 / Mythos 59 Jun 2026First public Mythos-class tier; SWE-bench Pro 80.3%; suspended 12–30 Jun under US export controls
Claude Opus 524 Jul 2026The Opus tier below, at half the price

Frequently asked questions

Is Claude Fable 5.1 the best AI model right now?

On the leading independent composite, yes — it scores 66 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5 (63) and Claude Fable 5 (62), and it is the first model above 65. We rank it #2 on our board, behind only the restricted Mythos 5.1, and it is the highest-ranked model you can actually buy. It is not the best value: Opus 5 costs half as much per token and less per completed task.

How much does Claude Fable 5.1 cost?

$10 per million input tokens and $50 per million output — unchanged from Fable 5 and double Claude Opus 5. Cache reads are the change: $0.25 per million, cut 75% from $1.00. Anthropic estimates roughly 25% off a typical workload and up to about 45% off highly agentic work. Batch requests are 50% off.

Is Fable 5.1 actually cheaper than Fable 5?

It depends on the workload. On cache-heavy agentic loops, yes — the cache-read cut is real and nothing else went up. But Artificial Analysis measures Fable 5.1 at $3.69 per completed benchmark task against Fable 5’s $3.08–$3.25, because it produces roughly double the median output tokens. If your bill is output-dominated, expect to pay more.

What is Claude Fable 5.1’s SWE-bench score?

Anthropic did not publish one — no SWE-bench Pro or Verified figure of any kind accompanied the launch announcement. The highest SWE-bench Pro score on our board remains Claude Fable 5’s 80.3%.

What’s the difference between Fable 5.1 and Mythos 5.1?

They are the same underlying model at different safeguard levels. Fable 5.1 ships the standard cyber and bio safeguards and is available to any paid Claude or API customer. Mythos 5.1 relaxes those safeguards for vetted US organisations through the Cyber Verification Program and the Life Sciences Verification Program, and scores higher on Terminal-Bench 4.0 (60.9% vs 55.8%) as a result.

What is the context window?

1,000,000 tokens, with up to 128,000 output tokens on the synchronous Messages API. Anthropic gives the reliable knowledge cutoff and the training data cutoff as June 2026.

Can Fable 5.1 be used for security work?

For finding vulnerabilities, yes — Anthropic explicitly permits vulnerability discovery while continuing to block exploit development, and reports around 60% fewer cyber-safeguard interventions per session in Claude Code than Fable 5. Work that needs the safeguards relaxed further runs through the Cyber Verification Program on Mythos 5.1.


Last verified 2 September 2026. Benchmark figures are Anthropic-reported from the 1 September 2026 launch comparison and vendor-run unless marked independent; the Intelligence Index score, cost-per-task, throughput and latency figures are Artificial Analysis measurements. Anthropic published no SWE-bench figure for this model. Specifications are from Anthropic’s models overview. Pricing and availability subject to change.