Claude Fable 5 is not a value pick. It is a premium escalation model.
Current access status: Anthropic suspended access to Fable 5 and Mythos 5 on June 12, 2026 after a US government directive. Do not route production work to either model until Anthropic restores access and your account confirms availability.
If access returns, use Fable 5 for rare, high-stakes long-horizon reasoning: autonomous coding work that Opus cannot finish, vision-heavy reconstruction, scientific or analytical tasks, and expensive decisions where a better answer is worth a visibly higher API bill.
Use Opus 4.8 first for most premium Claude work. Fable is 2x Opus pricing.
Use Sonnet 4.6 for routine production coding, agent workflows, summarization, and high-volume automation where cost matters.
Compare GPT-5.5 when OpenAI ecosystem fit, cached-input economics, or long-context pricing changes the workload math.
Do not default to Fable for cybersecurity, biology, or chemistry workflows unless fallback/refusal behavior is acceptable or your organization has approved trusted access.
The Cost Table
Prices are per 1 million tokens unless noted.
| Model | Standard input | Standard output | Batch input/output | Practical read |
|---|---|---|---|---|
| Claude Fable 5 | $10 | $50 | $5 / $25 | Mythos-class public model with safeguards; access currently suspended |
| Claude Opus 4.8 | $5 | $25 | $2.50 / $12.50 | Practical premium Claude baseline |
| Claude Sonnet 4.6 | $3 | $15 | $1.50 / $7.50 | Default Claude production lane |
| GPT-5.5 standard, short context | $5 | $30 | $2.50 / $15 | OpenAI premium standard lane |
| GPT-5.5 standard, long context | $10 | $45 | $5 / $22.50 | Similar input cost to Fable, lower output cost |
| GPT-5.5 Pro, short context | $30 | $180 | $15 / $90 | Much more expensive than Fable |
For a 100K-input / 20K-output request:
| Model | Approx cost |
|---|---|
| Claude Sonnet 4.6 | $0.60 |
| Claude Opus 4.8 | $1.00 |
| GPT-5.5 standard short context | $1.10 |
| GPT-5.5 standard long context | $1.90 |
| Claude Fable 5 | $2.00 |
| GPT-5.5 Pro short context | $6.60 |
The simple rule: Fable is 2x Opus, about 3.3x Sonnet, 2x GPT-5.5 short-context input, and about 1.7x GPT-5.5 short-context output. Against GPT-5.5 long context, Fable matches input pricing and is slightly higher on output. Against GPT-5.5 Pro, Fable is cheaper, but that does not make it a value model.
What Fable 5 Actually Is
Anthropic launched Fable 5 as a generally available version of its Mythos-class model, with safeguards added for public release. On June 12, 2026, Anthropic said it was disabling Fable 5 and Mythos 5 for all customers to comply with a US government directive. Read the launch availability claims as historical until Anthropic restores access.
Mythos 5 is the same underlying capability class without those Fable safety classifiers in some areas, but it was restricted to Project Glasswing and approved trusted-access customers before the suspension.
The API shape matters:
- Fable 5 and Mythos 5 use a 1M token context window by default and support up to 128K output tokens.
- Fable 5 was launched as generally available through the Claude API and major cloud platforms, but access is currently suspended.
- Mythos 5 was limited availability before the suspension.
- Fable 5 can return
stop_reason: "refusal"as a successful HTTP 200 response. - Anthropic documents fallback paths to retry on another Claude model.
- Fable 5 and Mythos 5 carry 30-day retention and are not available under zero-data-retention terms.
That last point is a deployment blocker for some teams. If your workload requires zero data retention, Fable should not be in the path until your legal/security review says otherwise.
How To Read The Benchmark Claims
Vendor-reported claims
Anthropic says Fable 5 has its biggest lead on longer and more complex tasks, including software engineering, knowledge work, vision, scientific research, memory, and long-context work. Launch coverage summarized Anthropic’s guardrail data as roughly 95% of Fable sessions running without Opus fallback, which implies a less-than-5% fallback rate on average. Treat that as a launch-average vendor claim, not a prediction for cyber, bio, chemistry, or model-development workloads.
Use those as Anthropic claims, not independent rankings.
Customer-reported claims
The launch post includes customer and partner reports from coding, finance, legal, spreadsheet, and research workflows. These are useful because they describe real workloads, but they are still selected launch testimonials. Do not turn them into blanket claims like “Fable is best for all coding.”
Independent / not yet verified
What is still missing:
- Public SWE-bench-style numbers that are directly comparable to existing Opus, Sonnet, GPT, Kimi, GLM, MiniMax, and Gemini rows.
- Third-party coding evals with reproducible prompts and token accounting.
- Real fallback/refusal frequency by domain, not just launch-average session rates.
- Cost-per-task tests: bug fix, refactor, architecture review, long-context analysis, vision reconstruction, and agentic coding run.
- Trusted-access clarity for organizations that need cyber or life-science capability without broad false positives.
Until that exists and access returns, Fable belongs in the “try only after Opus fails” bucket.
The Guardrail Problem
Fable’s public release includes classifiers around cybersecurity, biology, chemistry, and distillation-like use cases. Anthropic says flagged requests may fall back to Opus 4.8 or refuse, depending on the route.
This has already created friction. TechCrunch reported cybersecurity researchers seeing broad guardrail triggers, including complaints that code review and secure-code prompts could be downgraded. WIRED reported a separate June 11 walkback: Anthropic moved away from invisible safeguards for some frontier-AI-development cases and said users would be told when a prompt is refused or rerouted.
The buying implication is simple: if your workflow lives near cyber, bio, chemistry, model development, or safety research, do not buy Fable on benchmark claims alone. Test the actual prompts you need, log fallback behavior, and decide whether Opus 4.8 fallback is acceptable.
Decision Framework
Step 1: Start with Sonnet 4.6
Use Sonnet for routine coding, summarization, tool use, support automation, lightweight agent work, and high-volume production. At $3/$15 per million tokens, Sonnet is already expensive enough that you should have a reason before leaving it.
Step 2: Escalate to Opus 4.8
Move to Opus when Sonnet fails on architecture, multi-file reasoning, complex debugging, hard review, or long-horizon agent tasks. Opus remains the practical premium Claude baseline at $5/$25.
Step 3: Escalate to Fable 5 only after Opus fails and access returns
Use Fable when the expected value of a better answer exceeds the 2x Opus cost. Examples:
- A multi-day migration plan where one missed dependency is expensive.
- Screenshot-to-code or vision reconstruction where lower models lose structure.
- Scientific or financial analysis where the answer will be reviewed but the reasoning lift is valuable.
- Long-horizon autonomous coding where Opus stalls or loops.
Step 4: Compare GPT-5.5
OpenAI can win when:
- Your stack is already OpenAI-native.
- Cached-input pricing dominates the bill.
- Long-context GPT-5.5 standard pricing fits better than Fable.
- You need GPT-5.5 behavior or Codex tooling more than Claude behavior.
Step 5: Avoid Fable when the constraints do not fit
Avoid Fable for:
- Zero-data-retention requirements.
- Sensitive customer data without explicit approval.
- Cybersecurity, biology, or chemistry workflows where fallback breaks the job.
- Predictable low-cost scaling.
- Routine app code, summarization, and automation.
What To Test Before Buying
Before routing production workloads to Fable, run a small logged eval:
| Test | What to record |
|---|---|
| Representative prompts | The exact prompts your team would send in production |
| Fallback/refusal behavior | stop_reason, classifier labels, fallback model, and whether the fallback answer is usable |
| Cost per successful task | Full input, output, cache, batch, retry, and review cost |
| Data policy fit | Whether 30-day retention and no zero-data-retention option are acceptable |
| Opus comparison | Whether Opus 4.8 fails in a way Fable actually fixes |
| OpenAI comparison | Whether GPT-5.5 cached-input, long-context, or Codex tooling fits better |
Quick Answers
Is Fable 5 a value model?
No. Fable 5 is a premium escalation model at $10 input / $50 output per million tokens. It is 2x Opus 4.8 pricing and should not be treated as part of the value stack.
Should I use Fable before Opus?
No. Start with Sonnet 4.6 for routine production and Opus 4.8 for premium Claude work. Fable only makes sense after Opus fails in a way that matters, and only after Anthropic restores access.
Is Mythos generally available?
No. Mythos 5 was limited to Project Glasswing and approved trusted-access customers before Anthropic suspended access to both Mythos 5 and Fable 5 on June 12, 2026.
Does Fable support zero data retention?
No. Anthropic says Mythos-class models require 30-day retention and are not available under zero-data-retention terms. That is a blocker for workloads that require ZDR or stricter contractual handling.
What should security, bio, and chem teams test first?
Test fallback and refusal behavior before spend. Use representative prompts, record stop_reason, fallback model, classifier labels if exposed, and whether the answer remains usable. For now, route production work to approved alternatives because Fable and Mythos access is suspended.
Practical Verdict
Fable 5 is worth testing only after access returns, your current premium model is the bottleneck, and the work is valuable enough to pay for better reasoning. It is not the first Claude model to reach for, and it is not the default coding model.
The sane ladder is:
- Sonnet 4.6 for daily production.
- Opus 4.8 for premium Claude work.
- Fable 5 for rare tasks where Opus still fails, after access is restored.
- GPT-5.5 when OpenAI economics or tooling fit better.
Sources
- Anthropic: Claude Fable 5 and Claude Mythos 5 (Archive) — launch, availability update, safeguards, benchmark and customer claims.
- Anthropic: Statement on the US government directive to suspend access to Fable 5 and Mythos 5 (Archive) — June 12 access suspension and retention context.
- Claude API docs: Introducing Claude Fable 5 and Claude Mythos 5 (Archive) — API behavior, refusals, fallback, context, output, availability, and retention.
- Claude API docs: Pricing (Archive) — Fable, Opus, Sonnet, batch, cache, and context pricing.
- OpenAI: API pricing (Archive) — GPT-5.5 standard, long-context, batch/flex, priority, and Pro pricing.
- TechCrunch: Cybersecurity researchers aren’t happy about the guardrails on Anthropic’s Fable (Archive) — early security-community backlash.
- WIRED: Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI Researchers Using Claude (Archive) — visible-safeguard walkback after backlash.
Archive note: Wayback evidence for the original launch sources was collected on June 11, 2026 UTC, then the Anthropic launch and suspension pages were rechecked on June 13, 2026 UTC. Listed source archives are exact 200 text/html captures. The save endpoint timed out or returned 429 for some sources, so the evidence uses existing CDX captures where save-confirmation headers were unavailable.
Related links
- /compare/models/premium/ — Premium model cost comparison
- /models/claude-opus-4-5/ — Opus baseline guide
- /value/smart-spend/ — Spend strategy before premium escalation
- /posts/ai-coding-subscription-limits-may-2026/ — GPT-5.5 and Claude lane context
Last verified: June 13, 2026. Pricing, safeguards, and access status change quickly; recheck provider docs before routing production workloads.