Claude Fable 5: New Features, Pricing, Availability, and API Changes

TL;DR: Anthropic launched Claude Fable 5 on June 9, 2026 as its first generally available Mythos-class model. The main upgrades are long-horizon autonomy, stronger software engineering, better vision, improved memory use, more dependable parallel subagents, and higher performance on complex knowledge work. API pricing is $10 per million input tokens and $50 per million output tokens. Fable 5 also introduces stricter safety routing, a new refusal stop reason, adaptive-thinking changes, and mandatory 30-day retention for Mythos-class traffic. Some requests are answered by Claude Opus 4.8 instead.

Research note: This article was verified on June 10, 2026 using Anthropic's launch announcement and Claude API documentation. Capability and benchmark statements are attributed to Anthropic or its named early-access customers. No independent benchmark is claimed here. Availability is time-sensitive, so the exact subscription dates below matter.

Jump To

Key Takeaways

  • Claude Fable 5 is available through the Claude API under the model name claude-fable-5.
  • The model is aimed at long, difficult work rather than only faster chat responses.
  • Anthropic recommends high effort for most demanding tasks, xhigh for the most capability-sensitive work, and lower settings for routine jobs.
  • Developers should expect some runs to last many minutes or hours and should use streaming, longer timeouts, progress updates, and asynchronous job handling.
  • Cybersecurity, biology, chemistry, and suspected model-distillation requests may fall back to Claude Opus 4.8.
  • All Mythos-class traffic is retained for 30 days, including traffic through first-party and third-party surfaces.
  • Subscription access is temporarily included from June 9 through June 22, 2026, with usage credits planned from June 23 unless Anthropic extends the window.

What Is Claude Fable 5?

Claude Fable 5 is the generally available version of Anthropic's Mythos-class capability level. Anthropic says Fable 5 and the restricted Claude Mythos 5 use the same underlying model. The difference is access and safety routing: Fable is available broadly with classifiers that redirect selected requests, while Mythos is limited to approved cyber defenders, infrastructure partners, and planned trusted biology users.

The launch is less about a single chat feature and more about how long the model can keep working. Anthropic positions Fable 5 for software migrations, research, document analysis, visual tasks, and multi-agent workflows that span far more tool calls and context than a normal question-and-answer exchange. The company says the performance gap grows as a task becomes longer and more complex.

The Biggest Claude Fable 5 New Features

1. Longer autonomous work. Claude Fable 5 is designed to hold a goal across extended runs, retain instructions, gather context, use tools, and verify results. Anthropic's prompting guide warns that hard requests at higher effort can run for many minutes, while autonomous jobs can continue for hours. That changes application design. A synchronous request that waits behind one browser tab may be acceptable for a two-minute answer and poor for a multi-hour code migration.

2. Stronger software engineering. Anthropic reports gains on production-style coding evaluations and says early users assigned Fable 5 codebase-wide migrations, difficult debugging, prototyping, and multi-agent engineering work. One cited Stripe test involved a migration across a 50-million-line Ruby codebase. That is an early-customer report, not a promise that every repository will see the same result. The practical signal is that Anthropic expects the model to plan, edit, test, inspect failures, and continue across larger code surfaces.

3. Better vision. Fable 5 can interpret dense technical figures, detailed screenshots, web interfaces, and noisy images. Anthropic demonstrated tasks such as reading scientific charts and rebuilding an application's source from screenshots. Its documentation also says the model is trained to use shell and crop tools when an image is flipped, blurry, or cluttered. For developers, this points to stronger screenshot-to-code, visual regression analysis, document extraction, and interface troubleshooting.

4. Memory that supports long work. Anthropic says Fable 5 stays focused across millions of tokens in long-running tasks and makes better use of its own notes. Persistent files are still an application responsibility: a model cannot remember a prior session unless the product stores and returns that information. The new capability is better use of a well-designed memory system, not automatic permanent memory for every API call.

5. More dependable subagent delegation. Fable 5 is more willing to divide a task and keep parallel workers active. Anthropic recommends long-lived subagents that preserve context across related work and communicate asynchronously with the coordinating agent. This can reduce repeated context loading, but it raises the need for task ownership, result checks, cancellation, spend limits, and clear rules about which agent may change shared files.

6. Effort becomes the main quality and cost control. The effort setting controls how much work the model spends on a request. Anthropic suggests high as the default for difficult tasks and xhigh where maximum capability matters. medium or low can suit routine jobs and interactive use. Higher effort can improve checking and reasoning, but it can also increase latency, output, and cost or lead the model to inspect more context than the task needs.

Claude Fable 5 vs Claude Opus 4.8

Area Claude Fable 5 Claude Opus 4.8
Capability tier Mythos-class, generally available Opus-class
Best fit Long, complex, agentic work General high-capability work and fallback responses
Long runs Built for multi-day goal-directed work Shorter-running behavior by comparison
Subagents More active parallel delegation Less likely to sustain large parallel plans
Vision Anthropic's strongest generally available vision model Capable, but below Fable in Anthropic's launch comparison
Thinking control Adaptive thinking and effort levels Existing Opus controls differ
Safety routing May route selected requests to Opus 4.8 Receives those fallback requests
API price $10 input / $50 output per million tokens Check current Opus pricing before migration
Data retention Mandatory 30 days for Mythos-class traffic Existing product and account terms apply

Fable 5 does not make Opus 4.8 irrelevant. Opus remains the fallback when Fable's classifiers intercept selected requests, and existing applications may prefer Opus when its behavior, price, latency, or data terms better match the workload. Evaluate with your own tasks. A benchmark win on coding or finance does not settle response time, false-positive routing, tool reliability, or total cost in a specific product.

Claude API and Prompting Changes Developers Need to Know

Fable 5 supports adaptive thinking rather than the older pattern of manually assigning an extended-thinking token budget. Anthropic's migration documentation says thinking output is summarized, extended thinking budgets are not used, and applications must handle a new refusal stop reason. Teams should update response parsing, logs, retry logic, dashboards, and fallback behavior before changing the production model ID.

Prompts may also need subtraction. Anthropic says Fable 5 follows short instructions more strongly than prior models and that overly prescriptive skills can reduce output quality. State the goal, boundaries, acceptance checks, and reason for the work. Avoid long lists of style rules unless tests show they are needed. For autonomous jobs, require progress claims to be tied to tool results, define when the agent may pause, and use a separate verifier for high-impact output.

  1. Run representative evaluations against the old and new model before routing all traffic.
  2. Add streaming and raise client, proxy, worker, and job timeouts for long requests.
  3. Support refusal responses and configure server-side or client-side fallback where suitable.
  4. Test low, medium, high, and xhigh effort against quality, latency, and token cost.
  5. Remove instructions that request hidden reasoning or a transcription of internal thought.
  6. Give long-running agents a progress channel and a cancellation path.
  7. Review the 30-day retention requirement with privacy, security, and legal owners.

Safeguards, Fallbacks, and the 30-Day Data Rule

Fable 5 uses separate classifiers for offensive cybersecurity, biology and chemistry, and attempts to extract or copy model capabilities. When a request matches those controls, it is generally handled by Claude Opus 4.8 instead of Fable. Anthropic says more than 95 percent of early Fable sessions had no fallback, which also means a small but meaningful share did. Benign security or scientific requests can be caught because the launch controls are intentionally broad.

The data policy is a separate change. Anthropic requires 30-day retention for all traffic on Fable 5, Mythos 5, and future models at the same or higher capability level, on both Anthropic products and third-party surfaces. The company says the retained data will be used for safety work, not to train new Claude models or for unrelated purposes, and that human access is logged with deletion after 30 days in almost all cases.

Do not treat that statement as a substitute for your own data review. Before sending source code, customer records, legal documents, health information, or regulated data, confirm the current commercial terms, product-specific controls, regional requirements, subprocessors, and internal policy. A stronger model can still be the wrong model for data that must not be retained for 30 days.

Claude Fable 5 Pricing and Availability

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens through the Claude API. Developers call it with claude-fable-5. It was available immediately on June 9, 2026 for the API and consumption-based Enterprise plans.

For Pro, Max, Team, and seat-based Enterprise subscriptions, Anthropic announced a temporary included-access window from June 9 through June 22, 2026. On June 23, 2026, the company plans to remove included Fable access and require usage credits, unless available capacity lets it extend the window. Anthropic says it intends to restore Fable as a standard subscription feature when capacity permits. Check the live plan page before purchasing access because this schedule can change.

Cost planning should include more than the headline token rate. Long agent runs can read a large repository, open many documents, call tools repeatedly, and generate extensive output. Prompt caching, compact context, lower effort for simple stages, bounded subagents, and explicit stop conditions can have a larger effect on the final bill than small prompt edits.

Who Should Use Claude Fable 5?

  • Good fit: software teams handling large migrations, difficult debugging, repository-wide changes, and long Claude Code workflows.
  • Good fit: research and analysis products that combine many documents, charts, tables, screenshots, and tool calls.
  • Good fit: agent platforms that can run asynchronously, report progress, preserve memory, and verify results.
  • Test carefully: security and life-science applications that may encounter classifier fallbacks.
  • Test carefully: latency-sensitive chat products where a multi-minute high-effort turn would damage the user experience.
  • May not fit: workloads that cannot accept mandatory 30-day data retention.

The practical reason to move is not simply that Fable 5 is newer. Move when your application has work that Opus 4.8 cannot complete reliably within the current process, or when Fable reduces the amount of custom orchestration needed. Keep the old route available until evaluations cover quality, cost, duration, safety fallbacks, and data handling.

Frequently Asked Questions

Is Claude Fable 5 available now?
Yes. Anthropic launched it on June 9, 2026 through the Claude API and supported Claude plans, with temporary subscription conditions described above.

What is the Claude Fable 5 API model name?
claude-fable-5.

Does Fable 5 replace Opus 4.8?
No. Opus 4.8 remains a separate model and also receives requests redirected by Fable's safety classifiers.

Is Claude Mythos 5 the same as Fable 5?
Anthropic says they share the same underlying model. Mythos access is restricted and selected safeguards are lifted for approved users.

Sources