Claude Fable 5 Review: Pricing, Benchmarks, and Verdict

This Claude Fable 5 review examines Anthropic's most capable widely released model without treating launch demonstrations as a universal verdict. Fable 5 is designed for difficult reasoning, coding, vision, and long-horizon agent work. It also costs twice as much per API token as Claude Opus 5 and applies unusual safety and retention rules.

The short version: Fable 5 is a credible specialist for projects that exceed the reliable working range of cheaper models. Most users should start with Sonnet or Opus, measure the result, and escalate only when Fable produces a meaningfully better accepted outcome.

Claude Fable 5 Review: Quick Verdict

Fable 5 is Anthropic's quality-first model for ambitious, extended work. Its one-million-token context window, output limit of up to 128,000 tokens, adaptive thinking, memory support, tool use, code execution, and task budgets make it suitable for agent workflows that must plan, act, check results, and continue across many steps.

At launch, Fable led the Artificial Analysis Intelligence Index and its agentic work evaluations. Anthropic and early-access partners also reported gains on complex software, finance, vision, and research tasks. Those results support a structured trial, not a promise of better economics on routine work.

The tradeoffs are material. API pricing is $10 per million input tokens and $50 per million output tokens. Safety classifiers can refuse or reroute some requests. Fable traffic requires at least 30 days of retention and is unavailable in zero-data-retention configurations. For many teams, Claude Opus 5 will be the better default.

What Is Claude Fable 5?

Claude Fable 5 is a Mythos-class model that Anthropic made generally available with additional safety classifiers. Mythos is the company's capability tier above Opus. Fable 5 and the restricted Claude Mythos 5 use the same underlying model, but Mythos removes some safeguards and is limited to approved programs.

Anthropic launched Fable 5 on June 9, 2026, briefly suspended access after an export-control directive, and restored global access on July 1 after the controls were lifted. It is now available through Claude's paid applications, the Claude API as claude-fable-5, Amazon Bedrock, Google Cloud, and Microsoft Foundry. It is not available on Claude's Free plan.

Claude Fable 5 Features That Matter

Long-Horizon Agents and Coding

Anthropic positions Fable for tasks that can run for days rather than a single chat turn. The model can plan work, call tools, use code execution, maintain memory, compact context, delegate to subagents in a suitable harness, and validate outputs. The API also supports task budgets for controlling how much work an agent may spend on a goal.

Long runs can compound early mistakes. Use staged plans, isolated worktrees, automated tests, spending limits, and review checkpoints. Syntax Dispatch's Claude Code vs Codex comparison covers product-level workflow choices around the models.

One-Million-Token Context and Vision

Fable accepts text and image inputs, provides a one-million-token context window by default, and supports up to 128,000 output tokens per request. Anthropic highlights document reasoning, dense scientific figures, screenshot-to-code work, and long-running memory as strengths. Test recall, citations, extracted numbers, and omitted constraints on the actual materials it will receive.

Adaptive Thinking and API Behavior

Adaptive thinking is always enabled for Fable 5. Developers use the effort parameter to control reasoning depth; disabling thinking is not supported. Raw chain-of-thought is not returned, although the API can provide a summarized reasoning block.

This behavior makes migration more than a model-name swap. Test latency, token consumption, multi-turn context handling, tool-call parsing, and any logic that assumed a disabled-thinking mode.

Claude Fable 5 Pricing and Access

Fable 5 costs $10 per million input tokens and $50 per million output tokens through the API. Anthropic lists a 90% input-token discount for prompt caching; current cache rates should be checked before budgeting. Standard token prices are twice those of Opus 5 and five times the introductory prices listed for Sonnet 5.

Subscription access is more complicated than a simple plan badge:

• Max plans and premium Team or eligible premium Enterprise seats include Fable for up to 50% of the regular weekly usage limit. Fable draws from that shared limit faster than other models. After the Fable allowance is reached, users can switch models or use paid usage credits.
• Pro plans and standard Team seats can access Fable with pay-as-you-go usage credits from the start; it is not included in their normal limits.
• Usage-based Enterprise and API customers pay standard metered rates.

Compare cost per accepted task, reviewer time, retry rate, latency, and regressions. Fewer failed attempts on a difficult migration can justify a premium; extra reasoning on a routine summary cannot.

Claude Fable 5 Benchmarks: What the Evidence Shows

Artificial Analysis scored Fable 5 at 64.9 on its Intelligence Index at launch, placing it first in that snapshot and highest on five of ten component benchmarks. It also reported a GDPval-AA Elo of 1932 for professional work and a 53% result on Humanity's Last Exam.

The same independent analysis reveals the caveat. Safety guardrails triggered a fallback to Opus 4.8 on 9% of the Humanity's Last Exam tasks, and the full run cost about $2,200. Its strong knowledge result came mainly from higher accuracy rather than uniquely low hallucination. These findings show frontier capability and significant evaluation expense at the same time.

Anthropic reports that Fable's advantage grows with task length and complexity. Partner examples include large codebase migrations, difficult coding benchmarks, finance analysis, vision tasks, and extended game environments. These are vendor-selected results and partner reports, not neutral proof for every workload.

Give Fable and a cheaper baseline the same brief, tools, context, tests, and stopping rules. Score accepted output, human intervention, unsafe actions, tokens, time, and recovery from feedback.

Safety Fallbacks and Data Retention

Fable's safeguards are part of the product, not a footnote. Classifiers cover selected cybersecurity, biology, chemistry, and advanced AI-research requests. A declined API request can return HTTP 200 with stop_reason: "refusal" rather than a conventional error. Integrations must detect that response explicitly.

Anthropic supports server-side, SDK, or manual fallback to another Claude model. A request refused before producing output is not billed, and Anthropic documents a prompt-cache credit when retrying on a fallback model. Users in Claude applications may simply see that the conversation switched models.

Fallback complicates reproducibility. If a workflow silently includes another model's responses, the measured system is not Fable alone. Log the responding model and classifier outcome.

Fable is also a Covered Model. Prompts and completions require at least 30 days of retention across supported platforms, and the model is not available to zero-data-retention organizations. Teams handling sensitive source code, regulated data, or strict deletion obligations should complete a privacy and contractual review before enabling it.

For broader deployment controls, see Syntax Dispatch's AI agent security guide.

Claude Fable 5 vs Opus 5 and Sonnet 5

Fable 5 sits above Opus 5 in Anthropic's lineup, but that does not make it the automatic choice.

• Choose Fable 5 for the hardest long-running coding, research, document, or multi-agent work when quality gains can outweigh higher cost and safety fallbacks are compatible with the task.
• Choose Opus 5 for demanding daily work that needs strong reasoning and agents at half Fable's API token price.
• Choose Sonnet 5 for higher-volume coding, analysis, and tool use where speed and cost matter more than the last increment of capability.

Routing by task is usually stronger than standardizing on the most expensive model. Start with Sonnet or Opus, escalate tasks that fail a clear acceptance check, and record whether Fable actually improves the result.

Who Should Use Claude Fable 5?

Fable is worth testing for engineering teams planning large migrations, researchers analyzing extensive mixed-format evidence, and organizations building long-horizon agents with strong tests and human oversight. It may also help when Opus repeatedly stalls on a well-specified, high-value task.

It is a weaker fit for casual chat, short summaries, high-volume extraction, latency-sensitive applications, zero-data-retention environments, or teams without a way to verify outputs. It should not receive broad autonomy merely because it performed well on a leaderboard.

Run a bounded pilot with representative tasks, predefined acceptance, an Opus 5 baseline, and tracking for actual model identity, refusals, cost, latency, and human correction.

Conclusion

This Claude Fable 5 review finds a powerful but specialized frontier model. Its long-horizon agents, one-million-token context, vision, memory, and strong benchmarks make it a serious option for work that cheaper models cannot complete reliably.

Fable's constraints are equally important: premium pricing, plan-dependent access, classifier-driven refusals and fallbacks, and mandatory retention. Those are operational design inputs, not minor terms to discover after deployment.

Treat Fable as an escalation tier. Begin with Opus or Sonnet, define acceptance tests, and use Fable when higher capability reduces failed attempts or human review enough to justify the cost.