Claude Fable 5.1 and Mythos 5.1: what actually shipped

Estimated reading time: 12 minutes. Last verified: September 1, 2026.

Key takeaways

• Claude Fable 5.1 and Claude Mythos 5.1 launched together on September 1, 2026. Anthropic says they are the same underlying model with different safeguard levels.
• Fable 5.1 is the generally available model for coding, knowledge work, and long-horizon agents. API id: claude-fable-5-1.
• Mythos 5.1 is the restricted twin for vetted cyberdefense and life sciences work (more permissive safeguards). Access is through Anthropic’s trusted access programs, currently skewed to US organizations.
• Headline API price is unchanged from Fable 5: $10 / 1M input and $50 / 1M output. Cache reads drop to $0.25 / 1M (75% cheaper than Fable 5). Anthropic estimates ~25% lower cost on typical workloads and up to ~45% on highly agentic ones.
• Vendor-reported scores put Fable 5.1 ahead of Fable 5, Opus 5, and GPT-5.6 Sol on several agentic coding and science benches. Treat those as Anthropic’s eval setup, not independent audits.
• For most builders: start on Claude with Fable 5.1 or stay on Opus 5 if cost/latency matter more than peak agent performance. Mythos is not a “better Fable you can flip on in the dashboard.”

Table of contents

1. What Fable and Mythos are
2. Specs that matter
3. Fable vs Mythos (the actual difference)
4. Pricing and where the savings come from
5. Benchmarks (vendor-reported)
6. Safeguards, privacy, and EU watermarking
7. How it compares to Opus 5 and GPT-5.6 Sol
8. Who should care
9. FAQ

---

What Fable and Mythos are

Model naming at Anthropic got noisier in 2026. Opus, Sonnet, and Haiku still sit in the general lineup. Fable and Mythos are the Mythos-class tier: higher capability, heavier safety tooling, and a deliberate split between public and gated access.

For practitioners, the useful definition is:

Claude Fable 5.1 is Anthropic’s generally available Mythos-class model for hard coding and long-running agent work. Claude Mythos 5.1 is the same model with more permissive cybersecurity and life sciences safeguards, limited to vetted organizations.

That separates three things people keep mixing up:

• Claude (the product): apps, Claude Code, Cowork, API, and cloud marketplaces.
• Fable 5.1 (this public model): what most Pro / Max / Team / Enterprise and API users can select today.
• Mythos 5.1 (gated twin): not a separate intelligence tier. It is Fable with different policy gates for cyber and biology R&D.

Fable 5 shipped June 9, 2026. Fable 5.1 is the September 1 incremental release: stronger long-horizon performance, cheaper cache reads, and tighter but less noisy safeguards.

---

Specs that matter

Verified against Anthropic’s September 1, 2026 announcement (claude-fable-and-mythos-5-1), the Claude Fable product page, and the Claude Platform models overview.

Spec	Claude Fable 5.1	Claude Mythos 5.1
Role	Generally available Mythos-class model	Same model, trusted-access safeguards
API model id	claude-fable-5-1	Program / account gated (not a public flip)
Context window	1M tokens	Same
Max output	128K tokens	Same
Thinking	Adaptive (always on)	Same
Default effort	High in Claude Code; Medium on Claude.ai and Claude Cowork	Same class of controls for approved users
Input / output price	$10 / $50 per 1M tokens	Starts at the same $10 / $50
Cache reads	$0.25 per 1M tokens	Same price shape where billed by token
Knowledge cutoff (docs)	June 2026	Same class
Cloud availability (Fable)	Claude API, AWS, Google Cloud, Microsoft Foundry / Azure	Limited trusted programs

If your workload is mostly short chat turns, you will feel the sticker price before you feel the 1M context. The product intent is multi-hour coding agents, tool loops, and research-style tasks that keep a large working set warm.

---

Fable vs Mythos (the actual difference)

Anthropic is unusually blunt here: same model, different safeguards.

	Fable 5.1	Mythos 5.1
Who can use it	Broad GA (API + Claude products + major clouds)	Vetted orgs via trusted access
Cyber	Can help find software vulnerabilities; still blocks exploit development and several dual-use tasks (often routed to Opus)	More permissive cyber safeguards for defensive work once CVP includes Mythos
Biology / life sciences	Benign medical / elementary biology false positives cut sharply; research biology still routed away	Life Sciences Verification Program (LSVP) for professional R&D
Claude Security product	N/A as the gated brain	Anthropic says Claude Security is now powered by Mythos 5.1

Cyber Verification Program (CVP): today gives reduced cyber safeguards on certain Opus / Sonnet models for defensive security. Anthropic says Mythos-class access is coming to CVP soon.

Life Sciences Verification Program (LSVP): invite-only beta for life sciences professionals, built with US government partnership. First participants are enrolled; broader enrollment is planned.

Practical rule: if you are shipping a normal SaaS coding agent, you want Fable 5.1. If you are a vetted cyberdefense or life sciences team blocked by Fable’s classifiers, you apply for Mythos. You do not “upgrade to Mythos” for a free quality bump.

---

Pricing and where the savings come from

Last verified: September 1, 2026 against Anthropic’s announcement and pricing notes.

Token type	Fable 5	Fable 5.1
Input	$10 / 1M	$10 / 1M (unchanged)
Output	$50 / 1M	$50 / 1M (unchanged)
Cache reads	$1 / 1M (prior Fable 5 rate Anthropic cites)	$0.25 / 1M (75% cut)

Anthropic’s indexed cost chart (four weeks of August 2026 usage at default effort) claims:

• Typical workload across Enterprise, Claude Code, and API: about 25% cheaper than Fable 5
• Highly agentic, context-heavy work: up to about 45% cheaper

That only holds where usage is billed by token and cache reads dominate. Seat-based Claude Pro / Max plans are a different meter: Fable 5.1 is available on those plans, but your bill is still plan limits first.

For agent stacks that re-read large repos every turn, cheaper cache reads matter more than shaving a few points off output price. If your traces show low cache hit rates, expect closer to sticker $10/$50.

---

Benchmarks (vendor-reported)

Numbers below are from Anthropic’s September 1, 2026 post. Fable 5.1 was evaluated with production safeguards on. Where safeguards intervened, Anthropic scored zeros or swapped in Opus models for cyber / biology tasks. That understates Fable on some suites.

Benchmark	Fable 5.1	Fable 5	Opus 5	GPT-5.6 Sol
Terminal-Bench-Science 0.1	52.6%	24.7%	29.0%	22.4%
Terminal-Bench 4.0	55.8% (60.9% Mythos 5.1)	42.0%	52.3%	37.3%
GDPval-AA v2 (knowledge work)	1853	1723	1824	1711
OSWorld 2.0 (computer use)	77.9% partial / 41.7% strict	72.9% / 36.1%	75.4% / 39.6%	—
Humanity's Last Exam	60.9% no tools / 65.0% with tools	57.8% / 63.8%	56.6% / 63.6%	—
AutomationBench	31.4%	17.1%	26.9%	19.6%
CursorBench 3.2.0	73.4%	70.5%	70.0%	67.2%

The science-agent jump (24.7% → 52.6% on Terminal-Bench-Science) is the loudest claim. Anthropic also lists partner anecdotes (Jane Street, Millennium) about long-running debugging and trading intuition. Useful as direction, not as your production SLA.

Effort settings matter for cost. Anthropic says Low / Medium effort on Fable 5.1 can match or beat Fable 5 High at lower spend. Defaults: High in Claude Code, Medium on Claude.ai and Cowork.

---

Safeguards, privacy, and EU watermarking

Enterprise Frontier Safeguards (EFS)

EFS is Anthropic’s answer to “we need ZDR-like privacy without dropping misuse detection.” Customer data stays in customer-controlled cloud infra; human review defaults to the customer. Rolling out in phases from fall 2026 across Claude Code, Enterprise, Platform, Bedrock, Google Agent Platform, and Microsoft Foundry. Eligible customers can run Fable 5.1 with zero data retention until EFS lands.

More precise cyber and biology gates

Relative to early Fable 5:

• Biology false positives on elementary / medical questions: about 85% fewer firings (Anthropic, Aug 2026 update carried into 5.1)
• Cyber interventions in Claude Code: about 60% fewer per session on average
• New allowance: Fable may help discover vulnerabilities, still not build exploits; several dual-use cyber tasks still route to Opus

Distillation hardening

New API accounts (from launch day forward) cannot manually edit prior thinking while keeping a multi-turn transcript, a common distillation trick. Existing accounts are grandfathered for now; Anthropic says future models will apply it broadly.

EU AI Act watermark

Models released after August 2, 2026 get an invisible text watermark for EU transparency rules. Detection API is in private preview for eligible orgs. No user or conversation PII in the mark, per Anthropic.

---

How it compares to Opus 5 and GPT-5.6 Sol

	Fable 5.1	Opus 5	GPT-5.6 Sol (Anthropic’s table)
Best fit	Peak agentic coding / long jobs	Strong general enterprise coding at half Fable’s list price	OpenAI frontier reference in Anthropic’s chart
List API price	$10 / $50	$5 / $25 (Platform overview)	Not our verification target here
Latency (docs)	Slower	Moderate	—
Safeguard theater	Heavy Mythos-class classifiers	Lighter than Fable	Different vendor policy
When to pick it	Evals beat Opus at High effort; cache-heavy agents	Default for most production agents	If your stack is already OpenAI-native

Anthropic’s own models overview still says: start with Opus 5 for most workloads; move to Fable 5.1 when demanding reasoning / long-horizon agents need it, or when Opus evals stall.

Versus ChatGPT / GPT-class models and DeepSeek: Fable is priced like a premium agent brain, not a cheap completion API. Cost-sensitive batch jobs still belong on cheaper models. Versus Cursor: CursorBench 73.4% is Anthropic’s harness result; re-run your own repo evals inside Cursor before changing the org default model.

---

Who should care

Teams running multi-hour coding agents. Cheaper cache reads plus higher Terminal-Bench / CursorBench scores are the real product pitch. Pair with Claude Code or your own tool loop.

Enterprises waiting on ZDR. EFS plus interim ZDR eligibility is the privacy story, not a new chat UI.

Cyberdefense and life sciences orgs. Mythos 5.1 is for you if Fable’s classifiers block legitimate R&D. Apply via CVP / LSVP; expect US-first access.

Most product teams on Opus 5. Stay put until your evals show Fable 5.1 wins enough tasks to justify 2× list price and slower latency.

Not ideal for: unbounded consumer chat, low-cache one-shot completions, or anyone assuming Mythos is a public “pro” toggle.

---