Claude Fable 5.1 Review: Features, Pricing, Pros and Cons

Anthropic just shipped its most expensive, most capable chat model yet, then turned around and told most people not to use it. Its own model picker guide says it plainly: start with Claude Opus 5 for most workloads, and only reach for Claude Fable 5.1 once your evals on Opus 5 at higher effort still fall short. That's an unusually honest thing for a company to say about its own flagship, and it's the right place to start this review.

One quick note before we go further. Frontier AI models move fast. Anthropic ships new versions, pricing changes, and features every few months, so treat every number below as accurate as of today, and confirm the current lineup and pricing directly on Anthropic's site before you commit real budget to anything long term.

Claude Fable 5.1 is Anthropic's flagship model, built for demanding reasoning and long horizon agentic work: the kind of job that runs for hours, touches multiple tools and applications, and can't afford to lose the thread halfway through. It costs double what Claude Opus 5 costs per token, and Anthropic itself says most people don't need it. So who does? That's what the rest of this review is actually about.

Claude Fable 5.1 at a Glance

That table alone tells you most of the story. Everything else in this review is the detail behind those eleven rows, especially the two that actually change how much this model costs you in practice: the cache read price, and the fact that the 1M token context window carries no long context surcharge at all.

What Is Claude Fable 5.1, and Why Does It Exist?

Claude Fable 5.1 is a point release. Fable 5 launched in June 2026 as Anthropic's reasoning and coding flagship; Fable 5.1 followed on September 1, 2026, alongside an invite only sibling called Claude Mythos 5.1, which we cover separately in our Claude Mythos 5.1 review.

Anthropic positions Fable 5.1 for "demanding reasoning and long horizon agentic work," built for jobs that span multiple applications and can run unattended for extended periods, the kind of task where a model has to hold a plan in its head for hours, not seconds. It also reads diagrams, charts, and tables nested inside long documents and PDFs, useful if your workload is research or knowledge work rather than plain chat.

Here's the useful way to think about where it sits, per Anthropic's own model comparison. Sonnet 5 is the speed and intelligence balance for most production apps. Haiku 4.5 is the cheap, fast option for simple, high volume tasks. Opus 5 is the one Anthropic tells you to try first for anything demanding. Fable 5.1 is the model you reach for only after Opus 5 has genuinely let you down on a task you can measure. We cover that sibling by sibling in our Claude Opus 5 review, worth reading before this one if you haven't picked a model yet.

What's Actually New in Fable 5.1

The headline change isn't a new architecture. Anthropic hasn't published architectural details for Fable 5.1, so treat any claim about what's happening under the hood as unconfirmed rather than fact, this review sticks to what's actually documented: pricing, specs, and measured performance.

What did change:

• A much cheaper cache read rate. Fable 5's cache reads cost $1 per million tokens, the standard 10% of its base input price. Fable 5.1 drops that to $0.25 per million tokens, just 2.5% of base input. We'll get into why that matters more than it sounds like it should.

• Meaningful gains on agentic benchmarks. Anthropic's own launch numbers show Terminal-Bench-Science 0.1 more than doubling, from 24.7% on Fable 5 to 52.6% on Fable 5.1. Terminal-Bench 4.0 moved from 42.0% to 55.8%. For context, Anthropic's own chart put GPT-5.6 Sol at 22.4% on Terminal-Bench-Science 0.1, well behind both Fable models.

• The freshest knowledge cutoff in the current Claude lineup. June 2026, a month ahead of Opus 5's May 2026 cutoff, and comfortably ahead of Sonnet 5 (January 2026) and Haiku 4.5 (February 2025).

• The longest retirement commitment in the lineup. Anthropic won't retire Fable 5.1 sooner than September 1, 2027, longer than Opus 5's July 2027 commitment or Sonnet 5's June 2027 one.

What didn't change: the base input and output price stayed exactly where Fable 5 left it, $10 and $50 per million tokens. Anthropic isn't charging more for Fable 5.1, it's charging the same and making the caching math better.

What the Benchmarks Actually Say About Claude Fable 5.1

Most coverage of a new frontier model is just the vendor's launch chart retyped. That's worth nothing to you, so here's the breakdown, split by who ran the test.

Anthropic's own reported numbers (from its launch materials, not independently verified):

One conflation worth flagging: several recap articles quote a 95% SWE-bench Verified score next to Fable 5.1's launch. That figure belongs to Fable 5's June 2026 launch. Anthropic never published a SWE-bench Verified score for Fable 5.1 itself, and we couldn't find one from an independent evaluator either. If that number matters to your decision, check a live benchmark tracker rather than assume the old figure carries over.

Independent evaluators tell a more useful, and more complicated, story:

• Artificial Analysis's Intelligence Index puts Fable 5.1 at 66 at max reasoning effort, the highest score they've measured, just ahead of Opus 5 (63, also at max effort), Fable 5 (62), GPT-5.6 Sol (61), and Grok 4.6 (61, high effort). That "66" is specifically a max effort number: Artificial Analysis notes Fable 5.1's five effort settings span an 11x range in output tokens used, from 13.1 million at low effort to 143.7 million at max, and the Intelligence Index score across that range moves from 58 to 66. Effort level changes the score meaningfully, so a bare "66" without the effort setting attached isn't the full picture.

• Vals AI's composite Vals Index puts Fable 5.1 at 67.87%, their top ranked model of the 56 they track, ahead of Opus 5. On individual benchmarks, Vals AI has it at 100% on ProofBench v1.1, first place on LiveCodeBench at 90.52%, and 92.38% on MMLU Pro.

• Terminal-Bench 2.1, specifically, is where the two independent evaluators disagree the most, and it's worth naming explicitly because it's a repeat of a pattern we found on the exact same benchmark for Claude Opus 5 (Vals AI at 84.64%, Artificial Analysis at 89.1%, see our Opus 5 review). On Fable 5.1, Vals AI scores it at 85.02%, good for third place among the 64 models they've tested. Artificial Analysis scores the same model on the same named benchmark at 91.4%, which it calls the highest it has recorded. Same model, same benchmark name, a six point gap between two credible evaluators. That's not a typo on either side, it's a structural difference in how each lab runs terminal agent tasks, and it means a single Terminal-Bench 2.1 number quoted without its evaluator attached is not a complete claim.

Both evaluators agree on direction even where they disagree on magnitude: Fable 5.1 tests as Anthropic's strongest model on coding and agentic tasks, and edges out Opus 5 on general intelligence composites, just not always by the same margin depending on who's measuring and at what effort.

Claude Fable 5.1 Pricing: The Real Numbers

Here's the full per-token breakdown, verified against Anthropic's own pricing page:

A few mechanics that actually move your bill:

• The 1M token context window has no long context surcharge. A 900,000 token request bills at exactly the same per token rate as a 9,000 token one.

• US only inference costs 10% more. Setting inference_geo to "us" applies a 1.1x multiplier across every token category. The default global routing uses standard pricing.

• Batch processing is 50% off both input and output, and stacks with prompt caching discounts.

• On claude.ai, Fable 5.1 is available to Pro, Max, Team, and Enterprise plans. Max tier and premium seats can spend up to 50% of their weekly usage limits on Fable class models at no extra API cost, useful if you're already paying for a subscription rather than metering usage through the API.

For context, Fable 5.1 sits well above the rest of Anthropic's current lineup on price: Opus 5 runs $5 input and $25 output per million tokens, Sonnet 5 is $2 and $10, and Haiku 4.5 is $1 and $5. If you want the full family lined up side by side, that comparison lives in our Claude Opus 5 review, which is genuinely the better starting point for most workloads before you consider Fable 5.1 at all.

The Cache Trick That Quietly Changes the Math

Here's the detail that makes Fable 5.1's pricing more interesting than "it costs 2x Opus 5." Every other current Claude model charges 10% of its base input price for a cache read. Fable 5.1 and its Mythos twin charge just 2.5%. In dollar terms, that means a Fable 5.1 cache read costs $0.25 per million tokens, while an Opus 5 cache read costs $0.50 per million tokens, half the price, on a model with double the base rate.

That only matters if your workload actually rereads cached context, so let's work through it honestly rather than just asserting it.

Picture a coding agent that loads a 100,000 token codebase into context once, then works through a long task in 200 small steps, rereading that same cached codebase at every step while adding only a short new instruction and producing a short reply each time. Run the numbers on Fable 5.1's actual published prices: one cache write at $12.50/MTok, 199 cache reads at $0.25/MTok, plus the small per-step input and output, comes to roughly $6.93 for the whole run. Run the identical 200 step job on Opus 5, at its own actual prices (cache write $6.25/MTok, cache read $0.50/MTok, and its cheaper base rates), and it comes to roughly $10.93. On this specific, cache read heavy shape of workload, the model that costs twice as much per token ends up about a third cheaper to actually run.

That's a stylized example, not a guarantee, and it doesn't hold for every workload. Shrink the number of reused reads relative to fresh output, and Opus 5's lower base price wins again, since output tokens get no caching discount and Fable 5.1's output price is still double Opus 5's. If your workload is a stream of unrelated one-shot prompts with nothing reused, none of this applies, you never hit a cache read, so you just pay the higher base price with nothing to offset it. Anthropic's own estimate is that typical workloads see costs drop around 25% from the cache change alone, and heavily agentic ones up to around 45%, which lines up with the direction of the example above even if the exact percentage depends on how much of your workload is genuinely reused context.

If you're weighing Fable 5.1 against a rival lab rather than against Anthropic's own cheaper siblings, we've already run those numbers elsewhere. Our GPT-6 Astra vs Claude Fable 5.1 comparison covers the case where both land on the identical $10/$50 sticker price, and our DeepSeek V4 vs Claude Fable 5.1 comparison covers the much larger price gap against a budget alternative. This review stays focused on Fable 5.1 on its own terms.

Where Claude Fable 5.1 Is Strong

Where it's strong:

• Long horizon agentic work across many tool calls and hours of runtime, exactly what Anthropic built it for, rather than a general chat upgrade.

• The full 1 million token context window at standard per-token pricing, with no long context surcharge past any threshold.

• The freshest reliable knowledge cutoff in the current Claude lineup, June 2026, ahead of every other model Anthropic currently sells.

• A genuinely cheaper cost per finished job on cache read heavy, long running agent workloads, thanks to the 2.5% cache read rate instead of the usual 10%.

• The top score on Artificial Analysis's Intelligence Index at max effort and the top spot on the Vals Index, per two independent evaluators that agree on direction if not always on magnitude.

• The longest retirement runway in the lineup, not sooner than September 1, 2027, useful if you don't want to requalify a model every few months.

Where Claude Fable 5.1 Falls Short

Where it falls short:

• It's the slowest model in the current Claude lineup. Anthropic's own comparison lists its latency as "Slower" against Opus 5's "Moderate" and Haiku 4.5's "Fastest."

• Thinking is adaptive and always on, and cannot be turned off. There's no lightweight, no reasoning mode for simple requests the way some other Claude models allow.

• It costs double Opus 5 on both input and output tokens, and Anthropic's own guide tells people to try Opus 5 first. That's a real cost most buyers pay before finding out whether they needed to.

• Agentic legal research is a clear weak spot: Vals AI's Harvey benchmark scores it at just 6.67%, 18th of 55 models tested, well behind its coding and reasoning scores.

• Claude Mythos 5.1, the safeguard adjusted twin built for cybersecurity and life sciences work, is invite only through Project Glasswing. Most readers can't access that variant even if their use case calls for it.

• The cache savings only materialize on workloads that actually reuse context. A stream of unrelated prompts sees none of the benefit and just pays the higher sticker price.

How to Actually Get Claude Fable 5.1

Developers reach it with the model ID and alias claude-fable-5-1, through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, or Claude Platform on AWS. Every dateless Claude API ID from the 5.1 generation onward is a pinned snapshot, so it won't quietly change underneath you. Regular users get it on claude.ai through the Pro, Max, Team, and Enterprise plans covered in the pricing section above.

Claude Mythos 5.1, the same underlying weights with a different safeguard layer, is not generally available at all. It's gated behind Anthropic's Project Glasswing program, invite only and currently limited to vetted cybersecurity and life sciences organizations. Our Claude Mythos 5.1 review goes into what that actually means.

Who Should Use Claude Fable 5.1, and Who Should Skip It

Reach for Fable 5.1 if:

• You run agent loops spanning hours and many tool calls, where losing the thread partway through is expensive.

• You've already benchmarked Opus 5 on your own eval set, at higher effort, and it measurably falls short.

• Your workload holds a large, repeatedly reused context, a codebase, a long document set, a big system prompt, so the cheap cache reads offset the higher base price.

• You need the freshest knowledge cutoff in the lineup for something time sensitive.

Skip it, at least for now, if:

• You haven't run Opus 5 through your own evals yet. That's the cheaper test to run first, half the price per token, and it's Anthropic's own recommended starting point.

• Your workload is mostly short, one-shot requests with little repeated context. You'll pay the higher price with none of the cache upside.

• Response speed matters more than raw capability. Fable 5.1 is the slowest model Anthropic currently sells.

• You're judging cost per request rather than per finished job. The cache economics only make sense across a long running task, not a single call.

Final Verdict: Is Claude Fable 5.1 Worth It Over Claude Opus 5?

For most people, the honest answer is no, not yet. Anthropic itself says to start with Opus 5, good advice from a company that would rather sell you the pricier model. Run your own evals first. Opus 5 costs half as much per token, and for most coding, writing, and reasoning workloads, it's genuinely enough.

Where Fable 5.1 earns its price is narrower than "it's the best model." It's the right call for long horizon agentic work Opus 5 has measurably struggled with, and a noticeably better deal than its sticker price suggests on any workload built around a large, repeatedly reused context, thanks to the 2.5% cache read rate. Outside those two conditions, you're paying a flagship premium for a job a cheaper sibling could likely have handled.

That's the whole verdict, really: don't buy the flagship because it's the flagship. Buy it because you measured the cheaper model and it came up short.