Anthropic
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)
Released 
June 2026

Anthropic
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)

Anthropic's Mythos-class model for multi-day autonomous coding and knowledge work that plans, delegates, and self-verifies.

Modality:
Text
Image
PDF
model ID
anthropic/claude-fable-5

Output Speed *

68.64
tok/s

Intelligence Index *

49.7
/ 100

Context Window *

1000000
tokens

Input price

60
Anytoken

Output price

300
Anytoken
Claude Fable 5: Long-Horizon Autonomous Agents for Coding and Knowledge Work Claude Fable 5 is Anthropic's most capable widely released model and the public entry point to its Mythos-class tier, positioned above the Opus line. Built for the hardest coding and knowledge-work problems, it can run inside an agent harness for extended sessions, planning across stages, delegating to sub-agents, and checking its own work. The longer and more complex the task, the larger its lead over other Claude models. It suits teams running large-scale refactors, autonomous engineering sessions, and document-heavy analytical work rather than everyday chat. Integrate Claude Fable 5 via the AnyAPI.ai API

Performance

Where Fable 5 Separates: Long, Hard, Multi-Step Tasks

Claude Fable 5's strength is sustained, tool-driven work rather than single-shot answers. On Anthropic's agentic SWE-Bench Pro it posted 80.3%, ahead of Opus 4.8's 69.2%, and independent reporting cites 95.0% on SWE-bench Verified. On the hardest coding tasks the gap widens sharply: on FrontierCode Diamond it reportedly separates from competing models by a large margin. This pattern matters because the model's advantage compounds on tasks that require planning, self-verification, and recovery over many steps. In production, that translates to fewer failed autonomous runs on large refactors and multi-hour agent sessions, where a wasted run costs more than the token premium.

Benchmarks

Independent Signals on Reasoning, Coding, and Depth

Third-party evaluations reinforce Fable 5's positioning as a frontier reasoning model. Price Per Token, drawing on Artificial Analysis data, reports an Intelligence score of 62.1, a Coding score of 76.5, and GPQA around 92.6. Vals reports SWE-bench Verified at 95.0%, ahead of Opus 4.8 at 88.6% and GPT-5.5 at 82.6%. On GPQA-style graduate reasoning without tools, editorial testing put Fable 5 roughly 14 points ahead of GPT-5.5. Note that most coding figures are vendor- or aggregator-reported and depend heavily on the configured effort level, so benchmark numbers are only meaningful alongside their test conditions.

Output Speed

*
68.64
tok/s

Intelligence Index

*
49.7
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
93
%

HLE *

Adherence to multi-step structured instructions.
56
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Claude Fable 5 accepts text, image, and file inputs and returns text only. It carries a 1M-token context window as the default on the Claude API — no beta header, no long-context surcharge — paired with a 128,000-token maximum output. The two most consequential specs are the always-on adaptive reasoning (thinking cannot be disabled, though effort is configurable) and the safety classifiers that can decline flagged requests and route them to a less capable model. Both change how you architect requests: budget for thinking tokens, and plan explicit fallback handling for refusals.
Verified Specifications — 
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
1000000
 tokens
Maximum output tokens
128000
Reasoning
Yes
Knowledge cutoff
June 2026
Pricing (standard)
60
 AnyTokens in
 / 
300
 AnyTokens out

Comparison

Claude Fable 5 vs Claude Opus 4.8: When the Premium Pays Off

Fable 5 and Opus 4.8 are the natural comparison because Opus 4.8 is the model Fable 5 falls back to when safety classifiers decline a request, and both target demanding reasoning and coding. They share the 1M-token context window and 128K max output. The practical decision is capability versus cost: Fable 5 leads on agentic coding (80.3% vs 69.2% SWE-Bench Pro) and its lead grows on the longest, hardest tasks, while Opus 4.8 sits at roughly half the per-token output cost and remains strong on well-defined work.

Dimension
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)
Claude Opus 4.8 (Adaptive Reasoning, Max Effort)
Context window *
1000000
tokens
1000000
tokens
Output speed *
68.64
tok/s
0.00
tok/s
Intelligence Index *
49.7
42
Input pricing
60
AnyToken
30
AnyToken
Output pricing
300
AnyToken
150
AnyToken
Knowledge cutoff *
June 2026
May 2026

Choose Claude Fable 5 when tasks are long-running, ambiguous, or highly multi-step — large refactors, overnight autonomous agents, and frontier-difficulty debugging — where a failed run wastes far more than the token premium. Choose Claude Opus 4.8 for high-volume or well-scoped work where its lower per-token cost and faster turnaround matter more than the marginal capability gain, and where you would already be satisfied with its results.

Limitations & Trade-offs

Where Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) falls short

1
High latency and slow output. Independent aggregators measured Fable 5 at roughly 59 tokens per second with a time to first token above 50 seconds, and Anthropic's own docs rate its latency as slower than Sonnet and Opus. Always-on reasoning adds further overhead. This makes Fable 5 unsuitable for real-time chat, interactive UIs, or any latency-sensitive path; route those to a faster Claude model and reserve Fable 5 for asynchronous, long-horizon jobs.
2
Premium pricing on a per-token basis. Fable 5 sits at the top of the Claude family, roughly double Opus 4.8's per-token cost and several times Sonnet's. Output is billed at five times input. For high-volume or simple, well-defined tasks the premium is difficult to justify against Opus 4.8 or Sonnet. The economics only favor Fable 5 when its capability lead prevents failed runs or reduces iteration on genuinely hard work.
3
Safety classifiers can decline requests. Fable 5 ships with strong safeguards; flagged queries in areas like cybersecurity, biology, and model distillation are refused and can be routed to a less capable model such as Opus 4.8. Anthropic states this triggers in under 5% of sessions on average, but it still requires new response handling. Integrations must plan for refusals and fallback retries — you are not billed for a request refused before any output is generated.
4
No output modalities beyond text and no way to disable reasoning. Fable 5 returns text only; it does not generate images or audio despite accepting image and file inputs. Adaptive thinking is always on and cannot be switched off, so every request incurs thinking-token cost and added latency. Workloads that need deterministic, minimal-latency, short responses — or non-text outputs — are a poor fit and belong on other models.

Best-Fit Workloads

Where this model earns its place

01

Multi-day autonomous coding agents


Fable 5 is built for agent harnesses that run for extended periods — planning across stages, delegating to sub-agents, writing its own tests, and recovering when steps fail. Its top SWE-Bench Pro score (80.3%) and widening lead on the hardest FrontierCode tasks support use in overnight refactors and unattended engineering sessions. Anthropic cites a codebase-wide migration on a 50-million-line Ruby repository completed in a day. The trade-off is latency and cost, so scope these as asynchronous jobs rather than interactive tooling.

02

Whole-codebase refactoring and review


The 1M-token context window at standard pricing lets Fable 5 hold an entire mid-sized repository, and the 128K max output supports large diffs and long-form generation in a single request. Combined with strong agentic-coding benchmarks and vision-based self-checking of outputs against goals, this fits large-scale migrations, cross-cutting feature work, and performance analysis. Because the full window is the default with no long-context surcharge, whole-repo reasoning is practical — but uncached full-window calls are expensive, so use prompt caching aggressively.

03

Document-heavy analytical knowledge work


Fable 5 reads diagrams, charts, and tables nested inside files and PDFs, and posted the top score on Hebbia's senior-level Finance Benchmark with strong document-based reasoning. That suits finance, legal, and analytics workflows that combine long source documents with multi-step reasoning. The 1M context and file input let you feed dozens of documents at once. Note the January 2026 knowledge cutoff and that flagged content may be refused, so keep source material current and plan for fallbacks.

04

Complex scientific and research reasoning


Anthropic positions Fable 5 as strong on scientific research and reasoning, and it can extract precise numbers from detailed scientific figures using vision. It performs well on graduate-level reasoning benchmarks, separating from GPT-5.5 by a meaningful margin on tool-free GPQA-style questions. This fits research synthesis, data-analysis pipelines, and computational workflows. Because biology and chemistry topics are among the categories that safety classifiers may reroute, expect occasional refusals in sensitive research domains.

Pricing in anytokens via AnyAPI
Input
60
Output
300
Cache write
75
Cache read
6

Integration

Access Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) via AnyAPI.ai

Access Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Claude Fable 5 is Anthropic's most capable generally available model, built for long-horizon coding and knowledge work. It excels at multi-day autonomous agent runs, whole-codebase refactors, and document-heavy analysis where a task requires planning, sub-agent delegation, and self-verification across many steps. Its lead over other Claude models grows as tasks get longer and harder.

Claude Fable 5 has a 1,000,000-token context window and a maximum output of 128,000 tokens per request. On the Claude API the 1M window is the default — no beta header is required and there is no long-context surcharge — so whole-repository and multi-document reasoning is practical, with cost and latency being the real constraints rather than window size.

No. Claude Fable 5 uses adaptive thinking that is always on and cannot be disabled; only the effort level is configurable. Every request incurs thinking-token cost and added latency, so you should budget thinking tokens for long jobs and avoid Fable 5 for latency-sensitive, short-response workloads.

Fable 5 is a Mythos-class model shipped with strong safety classifiers. Requests flagged in areas like cybersecurity, biology, and model distillation may be declined and routed to a less capable Claude model such as Opus 4.8. Anthropic reports this triggers in under 5% of sessions on average, and you are not billed for a request refused before any output is generated.

Fable 5 leads Opus 4.8 on agentic coding, scoring 80.3% versus 69.2% on Anthropic's SWE-Bench Pro, with the gap widening on the hardest, longest tasks. Opus 4.8 costs roughly half as much per token and is a better default for high-volume or well-defined work. Reserve Fable 5 for large refactors and autonomous runs where a failed attempt costs more than the premium.

* Benchmark data source: Artificial Analysis artificialanalysis.ai