Claude Fable 5 is Anthropic’s generally available frontier model for hard reasoning, long-running agent work, coding, document analysis, and tool use. It’s powerful, expensive, and not a drop-in choice for every workload: as of 2026, API pricing is $10 per million input tokens and $50 per million output tokens, with adaptive thinking always on and 30-day data retention.
Claude Fable 5 at a glance
Anthropic announced Claude Fable 5 on June 9, 2026, with general availability listed in its Claude Platform documentation from that date. The API model ID is claude-fable-5, which matters if you’re migrating from another Claude model and want deterministic deployment names rather than marketing labels.
The company positions Claude Fable 5 as its most capable widely released model for demanding reasoning and long-horizon agentic work. Plain English: you pick it when the task needs planning, multi-step execution, code or data inspection, and tolerance for complex instructions. You probably don’t pick it to rewrite a two-line email.
Access is broad, though the launch was unusually turbulent. Anthropic says the model is available to Claude Pro, Max, Team, and Enterprise users, and to developers through the Claude Platform API plus marketplaces including AWS, Google Cloud, and Microsoft Foundry. After a U.S. export-control action on June 12, 2026, Anthropic took Fable and Mythos down for all users; access was restored on July 1, 2026, and Anthropic said on July 2 that Fable was available globally for all users after redeployment.
If you want the product-news angle rather than the implementation view, the earlier coverage of Anthropic’s public Fable release is a useful companion. Here, the focus is practical: what it costs, where it fits, and where it can bite you.
What is Claude Fable 5 best used for?
The strongest case for Claude Fable 5 is work that punishes shallow completion. Think repository-wide refactors, agent-driven QA, large document review, multi-tool research flows, complex planning, and PDF or spreadsheet interpretation where charts and tables are part of the evidence.
Anthropic says Fable 5 can work for days at a time in an agent harness such as Claude Code or Claude Managed Agents. Treat that as a capability claim, not a license to walk away. Long-running agents need budgets, checkpoints, logs, and a human with the authority to stop them.
For software teams, the model’s appeal is obvious. It supports programmatic tool calling, code execution, memory tool use, context editing via the context-management-2025-06-27 beta header, compaction, vision, effort controls, and task budgets through the task-budgets-2026-03-13 beta header. If your engineers are comparing AI coding systems more broadly, pair this with a look at terminal-native AI coding tools and how they handle local context, diffs, and review loops.
Document-heavy teams get another advantage. Anthropic says Fable understands diagrams, charts, and tables nested in files and PDFs, which is exactly where weaker models often hallucinate relationships or miss footnotes. Legal, finance, policy, and security teams should still ask it to quote source snippets and page references. Trust, but pin it down.
Pricing, token math, and the cost trap
Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens in 2026. US-only inference is available at 1.1x pricing, making it $11 per million input tokens and $55 per million output tokens. Prompt caching retains a 90% input-token discount, so repeated context can become much cheaper if you design for reuse.
Here’s the calculation many teams skip. Suppose an agent run reads 600,000 input tokens and produces 80,000 output tokens. Without caching, that’s $6 for input and $4 for output, or $10 total. If 500,000 of those input tokens are cached at the 90% discount, the input bill becomes roughly $1.50: $1 for cached input plus $1 for the remaining uncached 100,000 tokens. The same output still costs $4, so the run falls to about $5.50.
Output is the silent budget killer. A verbose agent that explains every micro-decision can cost more than a concise one doing the same useful work, because output tokens are priced five times higher than input tokens. Honestly, if your workflow doesn’t need visible reasoning summaries, logs, and long prose, constrain the response format aggressively.
| 2026 item | Standard pricing | US-only inference | Practical note |
|---|---|---|---|
| Input tokens | $10 per 1M tokens | $11 per 1M tokens | Caching can reduce eligible input cost by 90% |
| Output tokens | $50 per 1M tokens | $55 per 1M tokens | Most likely source of runaway spend |
| Example: 600k input, 80k output | About $10 without caching | About $11 | Before any cached-input savings |
| Example with 500k cached input | About $5.50 | About $6.05 | Assumes the same 80k output tokens |
For teams already wrestling with inference bills, cost control should be designed before rollout, not after the CFO notices a spike. The practical tactics in cutting AI API costs without losing quality apply especially well here: cache stable context, cap outputs, route easy jobs to cheaper models, and measure cost per completed task rather than cost per prompt.
How to deploy it without burning money
Start with routing. Claude Fable 5 should sit at the top of your model ladder, not under every button in your app. Use it for the cases where cheaper models fail: ambiguous bug reports, multi-file code changes, long PDFs, messy data, or agent plans with dependencies.
A sensible production rollout has a few non-negotiables:
- Define which tasks require Fable, and send routine summarization or classification elsewhere.
- Set task budgets before launch, especially for agents that can call tools repeatedly.
- Cache stable prompts, policies, schemas, repository maps, and product documentation.
- Limit output length by asking for structured JSON, patch files, tables, or concise decision logs.
- Record model ID, beta headers, token counts, tool calls, refusal rates, and final task outcome.
- Add human approval for destructive actions such as deleting files, merging code, changing cloud settings, or emailing customers.
The edge case nobody likes to discuss is a “successful” agent that spends too long being careful. Adaptive thinking is on by default for Fable and Mythos, and disabling thinking isn’t supported. That can improve hard-task performance, but it also means you need budgets and timeout behavior for jobs that would be better answered by a cheaper, faster model.
Engineering organizations should connect model adoption to quality practices, not treat it as a magic developer multiplier. The more relevant question is how AI changes review, testing, and ownership; the analysis of engineering quality skills in an AI-heavy workflow is worth reading before you let an agent edit production code.
Safety, refusals, and data retention
Claude Fable 5 includes safety classifiers and safeguards. Anthropic’s platform documentation says some cybersecurity and biology queries can be declined or routed to Claude Opus 4.8. That has two consequences: some legitimate security work may need clearer framing, and some high-risk requests won’t run on this model even if you’re a paying customer.
Security teams should not ignore that nuance. If you’re building defensive workflows, describe authorization, scope, asset ownership, and intended use in the prompt and surrounding application state. For broader context on AI-driven security pressure, the piece on AI finding weak spots faster explains why safeguards and audit trails now matter as much as raw model capability.
Data retention is another hard constraint. Anthropic says Claude Fable 5 and Claude Mythos 5 carry 30-day data retention and are not available under zero data retention. For some companies, that’s fine. For regulated workloads involving sensitive client data, unpublished M&A material, health data, defense contracts, or strict internal policies, it may be a blocker.
Raw chain-of-thought is not returned for Fable or Mythos. Thinking output can be summarized or omitted. That’s good for safety and product clarity, but it means your audit strategy should rely on inputs, outputs, tool logs, citations, intermediate artifacts, and tests rather than expecting hidden reasoning transcripts.
Claude Fable 5 versus Mythos, Opus, and cheaper models
Anthropic’s documentation says Claude Mythos 5 shares Claude Fable 5 capabilities without the safety classifiers and is available through Project Glasswing. That detail comes from platform documentation and should be treated as a specialized-access point, not a normal upgrade path for most developers.
For mainstream teams, the real comparison is simpler: should you use Fable, an Opus model, or a cheaper Claude tier? Fable is the premium general-availability choice when the task is long-horizon, tool-heavy, or reasoning-dense. Opus 4.8 may appear in routing for certain cyber or biology cases, according to Anthropic’s docs, and may still be the right benchmark if your workload is coding-focused; the Claude Opus 4.8 coding comparison gives useful context on that class of evaluation.
Cheaper models remain the better answer more often than vendors like to admit. If a task can be solved with a short prompt, limited context, and no tool loop, Fable’s pricing is hard to justify. Use it where failure is expensive, not where prestige feels nice.
Agent builders should also compare the harness, not just the model. Browser agents, coding agents, and repository agents all fail in different ways: bad context selection, unsafe actions, stale pages, tool loops, and overconfident summaries. If your use case is web automation, the guide to AI browser agents in 2026 helps separate model power from product reliability.
Should you use Claude Fable 5 now?
Use Claude Fable 5 if you have a clearly valuable task that benefits from long context, tool use, visual document understanding, or multi-step reasoning. Don’t use it as the default model for every customer chat, notification draft, or short classification job. At this price, casual use gets expensive fast.
The June 2026 suspension and July 2026 restoration also argue for operational caution. If access to a model can be affected by export-control decisions, your architecture needs graceful degradation: alternate models, queueing, fallbacks, and clear customer messaging when a premium capability is unavailable.
My practical recommendation is conservative. Run a two-week pilot with 20 to 50 real tasks, not toy prompts. Track success rate, human review time saved, total token cost, refusal rate, latency, and rollback incidents. If the model saves senior staff hours or completes work cheaper models can’t handle, scale it. If it merely writes prettier summaries, route down.
FAQ
Is Claude Fable 5 available through the API?
Yes. Anthropic’s 2026 platform documentation lists the API model ID as claude-fable-5, with availability through Claude Platform and marketplaces including AWS, Google Cloud, and Microsoft Foundry.
How much does Claude Fable 5 cost?
In 2026, pricing is $10 per million input tokens and $50 per million output tokens. US-only inference is priced at 1.1x, and prompt caching keeps a 90% discount on eligible input tokens.
Can Claude Fable 5 show its chain of thought?
No. Anthropic says raw chain-of-thought is not returned for Claude Fable 5 or Mythos 5; thinking output can be summarized or omitted.
Does Claude Fable 5 support zero data retention?
No. Anthropic’s platform documentation says Fable 5 and Mythos 5 have 30-day data retention and are not available under zero data retention.
Why was Claude Fable 5 taken down in June 2026?
After a June 12, 2026 U.S. export-control action reportedly blocked foreign nationals from using Fable 5 and Mythos 5, Anthropic took the products down for all users. Access was restored on July 1, 2026.
EN 
FR 
ES