01
A flagship tier above Opus
Fable 5 opens a new class of model that sits above the Opus line. It is the model to reach for when Opus is the current ceiling on a problem and you need more.
★ FLAGSHIP TIERAnthropicModel brief
CLAUDE FABLE · VERSION 5 · 4 JUN 2026Claude Fable 5. A new tier above Opus.
Anthropic's most powerful, most intelligent model, and the first in a tier that sits above Opus. Reserved for the hardest reasoning and the longest-horizon autonomy. We route to it by difficulty, not by default, because the price reflects the ceiling.
LLM APIclaude-fable-5Frontier LLM providersEval pipelines
Released
4 Jun 2026
Model ID
claude-fable-5
Input
$10 / 1M tokens
Output
$50 / 1M tokens
[TL;DR FOR CEO + CTO]
Five things to know.
- 01
A new top tier, not a point release.
Fable 5 is the most capable model Anthropic ships, positioned a rung above Opus. It is built for the problems where Opus is the current ceiling and you still want more headroom.
- 02
Premium pricing: $10 input, $50 output per million tokens.
Twice the input price and twice the output price of Opus 4.8. This is a route-by-difficulty model. The economics work when one Fable call replaces a chain of Opus retries on a problem Opus cannot close.
- 03
Same 1M context, 128K max output.
The window matches Opus. What changes is what the model does inside it: deeper planning, longer coherent agentic runs, and stronger judgement on ambiguous, multi-step work.
- 04
Almost the same request surface as Opus 4.8.
Adaptive thinking only, no sampling parameters. One new wrinkle: an explicit disabled-thinking flag now returns a 400, so you omit the thinking parameter instead. Behind our abstraction this is a one-line change plus an eval pass.
- 05
Where it earns its keep.
Long-running autonomous agents, deep research, codebase-scale migrations, and the hardest reasoning tasks. For routine, high-volume work, Opus and Sonnet remain the right tiers.
[BENCHMARKS]
How it stacks up.
From Anthropic's reported numbers. Fable 5 sets the family ceiling: it leads Opus 4.8 across coding, reasoning, computer use, knowledge work, and financial agentic tasks, and is the first Claude tier to lead GPT-5.5 on agentic terminal coding.
| Capability | Fable 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
Agentic coding SWE-Bench Pro | 73.5% +4.3 pts vs Opus 4.8 | 69.2% | 58.6% | 54.2% |
Agentic terminal coding Terminal-Bench 2.1 Terminus-2 public harness | 79.4% +4.8 pts vs Opus 4.8 | 74.6% | 78.2% | 70.3% |
Multidisciplinary reasoning Humanity's Last Exam no tools / with tools | 53.1% / 61.0% +3.3 pts vs Opus 4.8 | 49.8% / 57.9% | 41.4% / 52.2% | 44.4% / 51.4% |
Agentic computer use OSWorld-Verified | 85.1% +1.7 pts vs Opus 4.8 | 83.4% | 78.7% | 76.2% |
Knowledge work GDPval-AA | 1972 +82 vs Opus 4.8 | 1890 | 1769 | 1314 |
Agentic financial analysis Finance Agent v2 | 57.2% +3.3 pts vs Opus 4.8 | 53.9% | 51.8% | 43.0% |
Numbers as reported by Anthropic at the Fable 5 launch. We re-run our own evals on customer tasks before recommending a switch, and the gain over Opus has to clear the price gap on your workload before we route to it.
[SOFTWARE DEVELOPMENT IMPACT]
What it changes for the team building with it.
What changes for the engineering team. Two comparisons that matter: Fable 5 against the model it sits above (Opus 4.8), and against Sonnet, where most routine volume still belongs.
| Dimension | vs Opus 4.8 | vs latest Sonnet |
|---|---|---|
Production agents | Higher ceiling on long-horizon autonomy and ambiguous, multi-step reasoning. The gap shows up most on the runs where Opus stalls or needs a human nudge. | Different universe of problem. Sonnet handles the high-volume, well-specified steps. Fable is for the few steps in a workflow that are genuinely hard. |
Coding workflows | +4.3 pts on SWE-Bench Pro and +4.8 pts on Terminal-Bench 2.1 vs Opus 4.8. The first Claude tier to lead GPT-5.5 on agentic terminal coding. | Sonnet remains the workhorse for local edits and review. Reserve Fable for migrations and architectural changes that Opus cannot close in one pass. |
Cost and latency | Twice the per-token price of Opus 4.8 ($10 / $50 per million). No fast mode at launch. This is the most expensive tier in the family by design. | An order of magnitude more expensive than Sonnet per token. The only sound way to run it is to route by difficulty inside the agent, not to set it as the default. |
Migration risk | Behind a vendor-neutral abstraction, adopting Fable is a routing change plus an eval pass. The one code-level change is dropping any explicit disabled-thinking flag. | Not a replacement for either Opus or Sonnet. We run all three behind the same abstraction and let the agent pick the tier at runtime. |
Inside a Kensink build, reaching for Fable is a routing decision the agent makes when difficulty warrants the spend, not a vendor commitment frozen at design time.
[WHAT IS NEW]
The features that ship with it.
02
Adaptive thinking only, with one new rule
Like Opus 4.7 and 4.8, sampling parameters and fixed thinking budgets are gone. New on Fable: an explicit disabled-thinking flag returns a 400. Omit the thinking parameter entirely to run without it.
03
Lower prompt-cache floor
Fable caches prefixes from around 2,048 tokens, half the Opus floor of 4,096. Shorter shared preambles start paying off as cache reads sooner, which takes some sting out of the premium input price.
04
Effort control up to max
The effort parameter spans low through max, the same lever as Opus 4.6 and later. On the hardest tasks, higher effort plus a generous output budget is where Fable separates from Opus.
05
1M context at standard pricing
The full million-token window carries no long-context premium. The premium is on the per-token rate, not on using the whole window.
[PRICING]
What it costs.
Standard
$10 input
$50 output
$50 output
Per million tokens. Twice the price of Opus 4.8 on both input and output. This is a route-by-difficulty tier, not a default.
What you get
1M context
128K max output
128K max output
Frontier intelligence and the longest coherent agentic runs in the family. No fast mode at launch. Tune throughput with the effort parameter instead.
[ALIGNMENT + SAFETY]
What the alignment data says.
The most capable model carries the most scrutiny.
As the family ceiling, Fable 5 ships behind Anthropic's strongest safeguards, including real-time cybersecurity protections. Requests on prohibited or high-risk topics are more likely to be refused. The full assessment is in the Fable 5 System Card.
Honesty and judgement carry forward from Opus 4.8.
Fable inherits the honesty gains Anthropic reported for Opus 4.8: less likely to let flaws in its own output pass unremarked, more likely to flag uncertainty. On long agentic runs, that is the property that matters most.
Capability does not replace your evals.
A stronger model does not relax the need for task-specific evals on your data, your prompts, and your guardrails. We run the customer eval suite on every model change, and a more powerful tier raises the bar for what those evals must catch.
[OUR TAKE]
What this means for the build.
01
We route to Fable, we do not default to it.
At twice the per-token price of Opus, Fable only makes sense for the steps where it changes the outcome. Inside our builds, the agent picks the tier at runtime by task difficulty, so spend tracks the hard parts of the workflow and nothing else.
02
The price gap has to clear on your workload.
A benchmark lead over Opus is not a reason to switch on its own. The gain has to be worth double the token cost on your tasks. We measure that on the customer eval suite before any routing rule sends real traffic to Fable.
03
It is the model for the few genuinely hard steps.
Long-horizon autonomous agents, deep research, codebase-scale migrations, the reasoning that Opus cannot close. That is the Fable shortlist. Everything else stays on Opus or Sonnet.
04
Adoption is a routing change, not a rebuild.
Behind our vendor-neutral abstraction, adding Fable as a tier is a config and routing change plus an eval pass. The only code-level edit is dropping any explicit disabled-thinking flag, which now returns a 400.
[METHODOLOGY · K-FRAMEWORK]Integrated through the
Integrated through the
K-Framework.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-Framework01
Foundations
Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
02
Amplification
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
03
Judgment
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
[OBSERVABILITY]
Observability your team can read.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Instrumented
Cost per call
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Instrumented
Latency p50 / p95 / p99
Real distributions, not averages. We know which routes are slow, and why.
Instrumented
Eval pass rates
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
Instrumented
Prompt + completion logs
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
[COMMON QUESTIONS]
Questions we are getting asked.
- Should we move from Opus 4.8 to Fable 5?
- Not wholesale. Fable is a tier above Opus at twice the per-token price, so the right move is to route to it for the hardest steps rather than switch the default. We add it as a tier behind the abstraction, measure the gain over Opus on your eval suite, and only send traffic where the gain clears the price gap.
- What does Fable 5 cost compared to Opus?
- Fable is $10 input and $50 output per million tokens, double Opus 4.8 on both. There is no fast mode at launch. Because it is the most expensive tier in the family, it has to be used deliberately: one Fable call replacing a chain of failed Opus attempts is where the economics work.
- What is the migration cost from Opus?
- For projects built behind a vendor-neutral abstraction, adding Fable is a routing change plus an eval pass. The one code-level change is removing any explicit disabled-thinking flag, which returns a 400 on Fable. If you built directly against the SDK with no abstraction, budget one engineering day to add the seam, then the same eval pass.
- When does Fable 5 pay off over Opus 4.8?
- When the task is genuinely hard and the outcome has real value: long-horizon autonomous agents, deep research, large migrations, and reasoning where Opus stalls. For routine, well-specified, high-volume work, Opus and Sonnet are the better economic choice.
- Why is Fable more likely to refuse a request?
- As the most capable tier, Fable ships behind Anthropic's strongest safeguards, including real-time cybersecurity protections. Requests touching prohibited or high-risk topics are more likely to be refused. For most product work this is invisible. If your use case sits near a sensitive boundary, we validate it during the eval pass.
[RELATED]
Worth a look next.
DIRECT INTEGRATION · NO FRAMEWORKWant Claude Fable 5
Want Claude Fable 5
in your product?
Eval suite at handoff, full source ownership. We integrate against the model API the same way we integrate against Postgres. Sized to your scope.