The thirty-second version
Fable 5 is not an Opus upgrade. It is a new, higher tier. Anthropic calls it a Mythos-class model: a class that sits above Opus entirely. Fable 5 is the first Mythos-class model anyone can use, made safe for general release by wrapping it in strict safety classifiers.
How the family fits together now
Naming, briefly
Fable and Mythos 5 are the same underlying model. Fable (from the Latin fabula, "that which is told", akin to the Greek mythos) is the safeguarded public version; Mythos 5 is the unrestricted version reserved for vetted cyber defenders. The safeguards are the only difference, hence the two names.
What is Claude Fable 5?
Anthropic's most capable model ever made generally available. State of the art on nearly all tested benchmarks across software engineering, knowledge work, vision and scientific research, and built to run autonomously for longer than any previous Claude.
The back-story
In April 2026 Anthropic built a model so capable in areas like cybersecurity that it would not release it publicly. That model, Claude Mythos Preview, went only to vetted cyber defenders and critical-infrastructure providers under Project Glasswing, run in collaboration with the US government (expanded to roughly 150 organisations across 15+ countries on 2 June). Fable 5 is that class of capability finally made public, with hard safety limits bolted on. The release also landed days after Anthropic publicly urged AI labs to agree a coordinated brake on frontier development, warning of recursive self-improvement, and as the company prepares for an IPO.
What it's notably good at
Software engineering
Stripe reported Fable performed a codebase-wide migration across a 50-million-line Ruby codebase in a day; by hand it would have taken a team over two months. Highest frontier score on Cognition's FrontierCode, even at medium effort. State of the art on CursorBench.
Vision
New state of the art. Extracts precise numbers from dense scientific figures and can rebuild a web app's source code from screenshots alone. Completed Pokémon FireRed using raw screenshots only, no helper harness, something earlier Claudes could not do even with one.
Memory and long context
Stays focused across millions of tokens and improves its work using its own notes. Given file-based memory in Slay the Spire, its performance gain was three times larger than Opus 4.8's.
Knowledge work
Highest score of any model on Hebbia's senior-level Finance Benchmark. First model to break 90% on Hex's analytics benchmark, a ten-point jump over Opus. Strong document, chart and table reasoning.
The safeguards (the bit that makes it "Fable")
Separate classifier AIs watch every request. When one detects a restricted topic, Opus 4.8 answers instead of Fable, and you're told when that happens. Anthropic has deliberately tuned them cautiously, so they will occasionally catch harmless requests. They trigger in under 5% of sessions; 95%+ of sessions run entirely on Fable.
1. Cybersecurity
Mythos-class models excel at discovering and exploiting software vulnerabilities, plus agentic hacking (reconnaissance, lateral movement and so on). The classifiers block both exploitation and broader offensive cyber tasks. One external partner found Fable's cyber safeguards the most robust of any model tested: zero compliance with harmful single-turn requests, with or without 30 public jailbreak techniques.
2. Biology and chemistry
Tuned very broadly for now: most bio and chem requests fall back to Opus 4.8, because Mythos-class models showed real ability in dual-use tasks such as viral design prediction. A trusted access programme will give vetted biomedical researchers a version with these safeguards removed. Anthropic says it wants to narrow the false positives quickly.
3. Distillation
Requests flagged as attempts to extract Fable's capabilities to train competing models (Anthropic has identified large-scale attempts from authoritarian countries) also fall back to Opus 4.8.
Jailbreak testing
An external bug bounty produced no universal jailbreaks in over 1,000 hours. External red-teaming organisations also failed on long-form agentic tasks, although the UK AISI made progress towards one in a brief initial window. Anthropic's stated goal is to make any remaining jailbreaks slow and costly enough to detect before they're used at scale.
New: mandatory 30-day data retention
All Mythos-class traffic now carries compulsory 30-day retention, even for business customers with previous zero-retention agreements. Anthropic says the data won't train models and is used only to defend against novel attacks and reduce false positives, with human access logged. Worth knowing if you ever route client-sensitive material through Fable.
Availability
- API:
claude-fable-5, generally available now (Claude API, Bedrock, Vertex AI, Microsoft Foundry). 1M context, 128k max output, adaptive thinking always on. - Subscriptions: included on Pro, Max, Team and seat-based Enterprise at no extra cost until 22 June; usage credits required from 23 June; Anthropic intends to restore it as standard "as quickly as we can".
- Mythos 5 (safeguards lifted): Glasswing partners only for now, with a broader trusted access programme planned. Anthropic calls it the strongest cybersecurity model in the world.
Claude Opus 4.8 recap
The everyday flagship. An upgrade to Opus 4.7 (which it followed by just 41 days), stronger across coding, agentic tasks, reasoning and professional work, at an unchanged price.
What arrived with it
- Dynamic workflows (research preview, Claude Code on Max/Team/Enterprise): Claude plans a big job, runs hundreds of parallel subagents in one session, then verifies its outputs. Demonstrated on codebase-scale migrations across hundreds of thousands of lines.
- Effort control in claude.ai and Cowork on all plans: dial effort down for speed and lighter rate-limit use, or up ("extra", "max") for harder problems. Opus 4.8 defaults to high.
- Mid-task system messages in the API: developers can update instructions mid-run without breaking the prompt cache.
- Alignment: substantially lower misaligned-behaviour rates than 4.7, on a par with Anthropic's best-aligned models, plus fixes for 4.7's comment verbosity and tool-calling niggles.
Head to head
Rows where they differ are tinted. Use the button to hide everything they share.
| Feature | Fable 5 | Opus 4.8 |
|---|---|---|
| Model class | Mythos class (new top tier) | Opus class |
| Released | 9 June 2026 | 28 May 2026 |
| API ID | claude-fable-5 | claude-opus-4-8 |
| API price (per MTok in/out) | $10 / $50 | $5 / $25 |
| Fast mode | Not offered | $10 / $50 at 2.5x speed |
| Context window | 1M tokens | 1M tokens (200k on MS Foundry) |
| Max output | 128k tokens | 128k tokens (300k via batch beta) |
| Thinking | Adaptive, always on | Adaptive (effort defaults to high) |
| Benchmarks | State of the art nearly everywhere; 10%+ ahead on some tests; lead grows with task length | Strongest Opus yet; beaten by Fable on most tests |
| Restricted-topic behaviour | Cyber, bio/chem and distillation queries fall back to Opus 4.8 (<5% of sessions) | No fallback system; standard safety training |
| Data retention | Mandatory 30 days, even over zero-retention agreements | Standard policies, zero-retention possible |
| Subscription access | Included on Pro/Max/Team/Enterprise seats until 22 June, then usage credits | Standard on paid plans |
| Sibling model | Mythos 5: same model, cyber safeguards lifted, Glasswing only | None |
| Multimodal input | Text + image | Text + image |
Third-party verdicts on Fable vs Opus
- Hex: first model to break 90% on its core analytics benchmark, a 10-point jump over Opus.
- Iambic (Sean Ward): "Claude Fable 5's reasoning is a clear step beyond Opus 4.8... works at senior research scientist grade."
- Anaconda (Peter Wang): beats Opus 4.8 on its everyday spreadsheet suite at every effort level, finishing 25 to 30% faster.
- Rakuten: at the highest effort, Fable reflects on and validates its own work; "the extra thinking pays for itself".
Timeline
- 7 Apr 2026Claude Mythos PreviewFirst Mythos-class model, invitation-only via Project Glasswing for cyber defence. $25/$125 per MTok for participants.
- 28 May 2026Opus 4.8 ships everywhere41 days after Opus 4.7. Dynamic workflows, effort control, cheaper fast mode.
- 2 Jun 2026Glasswing expandsMythos access widened to ~150 organisations across 15+ countries running critical infrastructure.
- 9 Jun 2026Fable 5 + Mythos 5 launchFable 5 generally available; Mythos 5 (safeguards lifted) to Glasswing partners. Both $10/$50.
- 22 Jun 2026Last day of included accessFinal day Fable 5 is bundled with Pro, Max, Team and seat-based Enterprise plans.
- 23 Jun 2026Usage credits requiredFable 5 moves to usage credits on subscriptions; Anthropic aims to restore bundled access when capacity allows.
Cost calculator
API list prices, June 2026. Enter a workload in millions of tokens and compare. (Subscription plans draw on allocation rather than these per-token prices, but the ratios still tell you how fast each model burns it.)
Which model, when?
Tap a job type for a recommendation.
Rules of thumb
- Default stays Opus 4.8. Half the price, no fallback surprises, dynamic workflows in Claude Code, and it's excellent. Fable is for when Opus isn't enough.
- Fable earns its 2x price on length and difficulty. Its lead over Opus grows as tasks get longer and more complex. Short routine tasks won't show the difference.
- Security work goes straight to Opus. Fable's cyber classifier would route you there anyway; skip the detour (and the false-positive risk).
- Mind the retention policy. Mythos-class traffic is kept 30 days regardless of agreements. For client-confidential material you may prefer Opus.
- Before 22 June: throw at least one genuinely hard job at Fable while it's bundled with paid plans, e.g. a codebase-wide audit or a big migration, and see if the step up matters to your work.