On June 9, 2026, Anthropic did something it has not done before: it shipped a single frontier model as two distinct products. Claude Fable 5 — API model ID claude-fable-5 — is the new generally available flagship, wrapped in safety classifiers that hand certain queries off to Claude Opus 4.8. Claude Mythos 5 is the same underlying model with those safeguards lifted in some areas, and it is not generally available — access is restricted to vetted partners working on cyberdefense and infrastructure.
Anthropic's framing is that the raw model is capable enough in areas like cybersecurity and biology that releasing it openly would be irresponsible. So the company split the release: Fable 5 for everyone, with guardrails; Mythos 5 for a small set of approved organizations, without some of them. Both cost the same to run — $10 per million input tokens and $50 per million output — and both sit above Opus 4.8 on nearly every published benchmark.
This guide covers the two-model split and why it exists, the full five-model benchmark matrix (with an honest read on the asterisks that change what the headline numbers mean), the safeguard architecture, pricing and the staged rollout you need to plan around, and a practical decision guide for teams. For the model this release sits above, see our Claude Opus 4.8 release guide, and for the program behind Mythos access, our Project Glasswing expansion analysis.
- 01One model, two products — the split is the headline.Fable 5 and Mythos 5 are the same underlying model. Fable 5 is generally available with safety classifiers that route cybersecurity, biology/chemistry, and model-distillation queries to Claude Opus 4.8 instead. Mythos 5 is that model with the safeguards lifted in some areas, restricted to vetted Project Glasswing partners (cyberdefenders and infrastructure providers). Anthropic's reasoning: the raw model's capabilities could cause serious damage if released openly, so it gated the most sensitive ones behind an approval process.
- 02Fable 5 leads the board on the work businesses actually do.Against Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, the new model tops agentic coding (SWE-Bench Pro 80.3% vs 69.2% for Opus 4.8), knowledge work (GDPval-AA 1932 vs 1890), spatial reasoning, tool use, legal tasks, and health. It is built for long-horizon work: planning across stages, delegating to sub-agents, and running for days. At the highest effort it reflects on and validates its own output.
- 03Read the asterisks before you quote the numbers.Anthropic's table shows the higher of the Fable 5 and Mythos 5 scores. On starred benchmarks — cybersecurity, biology, and a few others — the displayed figure is Mythos 5's. The Fable 5 you can actually deploy performs closer to Opus 4.8 there, because its safeguards fall back to Opus 4.8 on those exact topics. The clearest case: ExploitBench shows 78.0% for the restricted model, but Fable 5 made 0% progress on offensive cyber tasks in blocking mode. Do not benchmark-shop on Mythos numbers for a Fable deployment.
- 04It is a premium tier, not a free upgrade.Fable 5 costs $10/$50 per million tokens — double Opus 4.8's $5/$25. A 90% prompt-caching discount applies to input, and US-only inference is available at a 1.1x multiplier. On subscription plans it is included free from June 9-22, then requires usage credits from June 23, and is restored as standard once capacity allows. Unlike the same-price Opus 4.7-to-4.8 jump, this one has a cost decision attached: reach for Fable 5 on the hardest long-horizon work; Opus 4.8 at half the price stays a sensible default for routine tasks.
- 05Safety is the product here, not a footnote.Mythos-class traffic carries a 30-day retention requirement, is not used for training, and logs all human access. The classifiers trigger on under 5% of sessions on average. More than 1,000 hours of external bug-bounty red-teaming found no universal jailbreak (the UK AI Safety Institute made early progress toward one in an initial window). Anthropic assesses the model's misaligned behavior as low and similar to Opus 4.8 — but higher-risk than models prior to the Mythos class, which is the whole reason for the gated tier.
01 — Release OverviewOne model, two products — and a deliberate line between them.
Start with the part that is easy to miss in the benchmark excitement: Fable 5 and Mythos 5 are not two different models. They are the same model, packaged two ways. Fable 5 is the version made safe for general use — its capabilities, in Anthropic's words, exceed those of any model the company has previously made generally available. Mythos 5 is that same model with safeguards lifted in some areas, and Anthropic describes it as having the strongest cybersecurity capabilities of any model in the world.
That second sentence is exactly why the model is gated. Anthropic says Mythos-class models have reached a threshold where they present significant risks, and that without safeguards, Fable 5's capabilities in areas like cybersecurity could be misused to cause serious damage. So Mythos 5 is not on sale. It is deployed through Project Glasswing — a program for cyberdefenders and critical infrastructure providers — with trusted-access programs planned for vetted cybersecurity organizations and biomedical researchers.
Fable 5, by contrast, is the model you and your team can use today. API model ID claude-fable-5. It is live across the Claude API, claude.ai (Pro, Max, Team, and Enterprise), and the major clouds — AWS, Google Cloud, and Microsoft Foundry. When a query lands on a safeguarded topic, Fable 5 quietly routes the response to Claude Opus 4.8 instead. For the overwhelming majority of business work — coding, analysis, content, research, agentic workflows — that handoff never fires.
How the two relate to the earlier preview: Anthropic says Mythos 5 is comparable to, or somewhat stronger than, the Mythos Preview that cybersecurity teams have been using under Glasswing — while costing substantially less. So this is both a capability step and a cost step for the restricted tier, and a first public release for the general-availability tier. For the back-story on the preview and the tier that sits above Opus, see our earlier Mythos tier analysis.
Generally available, with safeguards
API ID claude-fable-5. Live on the Claude API, claude.ai (Pro/Max/Team/Enterprise), AWS, Google Cloud, and Microsoft Foundry. Safety classifiers route cybersecurity, biology/chemistry, and distillation queries to Opus 4.8.
Same model, safeguards lifted — restricted access
No public API ID. Deployed via Project Glasswing to cyberdefenders and infrastructure providers. Trusted-access programs planned for vetted cybersecurity orgs and biomedical researchers. Described as the strongest cybersecurity model in the world.
Input / output per million tokens — both models
Double Opus 4.8's $5/$25 rate card. A 90% discount applies to input via prompt caching. US-only inference is available at a 1.1x multiplier on both input and output.
Sessions where safeguards intervene
On average, fewer than 5% of Fable 5 sessions trip a safety classifier. When one does, the response is handled by Opus 4.8. Mythos-class traffic carries a 30-day retention requirement and is not used for training.
Splitting a single model into a safeguarded public product and a gated high-capability tier is a meaningful shift in how frontier labs ship. It is an admission that capability and safety can no longer be tuned with one dial for one audience. For businesses, the practical implication is simple: the model you deploy (Fable 5) is deliberately constrained in a handful of domains, and that constraint is a feature, not a defect. The headline capability you read about in the cyber and biology benchmarks belongs to a model you cannot buy. Source: Anthropic Fable 5 & Mythos 5 announcement.
02 — Benchmark AnalysisThe five-model matrix — with the asterisks read honestly.
Anthropic published a comparison across Claude Fable 5 / Mythos 5, Claude Mythos Preview, Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. Two methodology points are load-bearing. First, the table shows the higher of the Fable 5 and Mythos 5 scores, which are within one to three percentage points of each other on most benchmarks. Second, starred (*) rows are where the two diverge more — because Fable 5's blocking safeguards for cybersecurity and biology pull its score down toward Opus 4.8. On those rows, the number you see is Mythos 5; the number you would get from a Fable 5 deployment is lower.
| Benchmark | Fable 5 / Mythos 5 | Mythos Preview | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|
| Agentic codingSWE-Bench Pro | 80.3% | 77.8% | 69.2% | 58.6% | 54.2% |
| Agentic codingFrontierCode (Diamond, xhigh) | 29.3% | — | 13.4% | 5.7% | — |
| Agentic codingTerminal-Bench 2.1 | 88.0%* | — | 82.7% | 83.4% (Codex CLI) | 70.7% (Gemini CLI) |
| Knowledge workGDPval-AA (ELO) | 1932 | — | 1890 | 1769 | 1314 |
| Knowledge work (vision)GDP.pdf, no tools | 29.8% | — | 22.5% | 24.9% | 16.7% |
| Spatial reasoningBlueprint-Bench 2 | 38.6% | — | 14.5% | 36.2% | 26.5% |
| Tool useAutomationBench | 17.4% | — | 15.5% | 12.9% | 9.6% |
| Computer useOSWorld-Verified | 85.0% | 85.4% | 83.4% | 78.7% | 76.2% |
| LegalLegal Agent Benchmark | 13.3% | — | 10.4% | 2.1% | 0.0% |
| Multidisciplinary reasoningHumanity's Last Exam, no tools | 59.0%* | 56.8% | 49.8% | 41.4% | 44.4% |
| Multidisciplinary reasoningHumanity's Last Exam, with tools | 64.5%* | 64.7% | 57.9% | 52.2% | 51.4% |
| BiologyBioMysteryBench, hard | 46.1%* | 29.6% | 40.0% | — | — |
| BiologyBioMysteryBench, human solved | 83.9%* | 82.6% | 80.4% | — | — |
| CybersecurityExploitBench (Cap%) | 78.0%* | 69.0% | 40.0% | 34.0% | — |
| HealthHealthBench Professional | 66.0%* | 64.7% | 56.9% | 51.8% | — |
Methodology, per Anthropic: reported scores are within a 1-3 percentage-point difference for Claude Mythos 5 and Claude Fable 5, and the table shows the higher of the two. Starred (*) benchmarks show a larger difference due to blocking safeguards for cybersecurity and biology-related questions; on those, Claude Fable 5 performs closer to Claude Opus 4.8 because of fallbacks. Source: Anthropic Fable 5 & Mythos 5 announcement, June 9, 2026.
The clean reads first. On agentic coding, the new model is in front by a wide margin — 80.3% on SWE-Bench Pro against 69.2% for Opus 4.8, 58.6% for GPT-5.5, and 54.2% for Gemini 3.1 Pro — and the gap on the harder FrontierCode Diamond set (29.3% vs 13.4% for Opus 4.8) is even larger in relative terms. Knowledge work (GDPval-AA 1932), spatial reasoning, tool use, legal, and health all show the same shape: a clear lead over Opus 4.8 and the broader field. These are the capabilities most teams will feel day to day.
The honest caveats. Two of them. First, computer use is the one row where the new model does not lead: Mythos Preview edges it 85.4% to 85.0% on OSWorld-Verified — a tie within noise, but worth noting that the new model is not a clean sweep. Second, and more important: every starred row reports Mythos 5, the restricted model. ExploitBench is the starkest example — 78.0% for the model behind the gate, while Opus 4.8 (which is effectively what Fable 5 falls back to on cyber queries) scores 40.0%, and Anthropic separately reports Fable 5 made 0% progress on offensive cyber tasks in blocking mode. If you are evaluating Fable 5 for deployment, treat the starred figures as the ceiling of the restricted tier, not the performance you will see.
SWE-Bench Pro — five-model comparison
Source: Anthropic Fable 5 & Mythos 5 benchmark table, June 9, 2026. Deltas are vs Claude Opus 4.8.SWE-Bench Pro 80.3% · GDPval-AA 1932 · HLE tools 64.5%*
Leads the field on agentic coding (SWE-Bench Pro 80.3%, FrontierCode Diamond 29.3%, Terminal-Bench 2.1 88.0%*), knowledge work (GDPval-AA 1932), spatial reasoning (38.6%), tool use (17.4%), legal (13.3%), and health (66.0%*). Starred scores are Mythos 5; Fable 5 lands closer to Opus 4.8 on those. $10/$50 per million tokens.
SWE-Bench Pro 69.2% · GDPval-AA 1890 · $5/$25
The current GA flagship and the model Fable 5 falls back to on safeguarded topics. SWE-Bench Pro 69.2%, GDPval-AA 1890, OSWorld-Verified 83.4%, Legal 10.4%, ExploitBench 40.0%. Half the price of Fable 5 and still the sensible default for routine, cost-sensitive, or latency-sensitive work.
Terminal-Bench 2.1 83.4% · GDPval-AA 1769
Strong on agentic coding via its own Codex CLI harness (Terminal-Bench 2.1 83.4%, just ahead of Opus 4.8's 82.7%) and on knowledge-work vision (24.9%, above Opus 4.8's 22.5%). But it trails the new Claude across the board: SWE-Bench Pro 58.6%, GDPval-AA 1769, OSWorld-Verified 78.7%, legal 2.1%.
Spatial reasoning 26.5% · GDPval-AA 1314
Competitive on spatial reasoning (Blueprint-Bench 2 26.5%) and multidisciplinary reasoning with tools (51.4%), but trails the field on SWE-Bench Pro (54.2%), GDPval-AA (1314), and legal (0.0%). Terminal-Bench 2.1 via Gemini CLI: 70.7%.
03 — The Safeguard ArchitectureWhy Anthropic split the model — and what the classifiers actually do.
The split is built on safety classifiers — separate AI systems that watch the conversation and detect attempts at misuse. When a classifier fires on Fable 5, the response is not refused outright; it is handed to Claude Opus 4.8, the next-most-capable model, which answers in Fable 5's place. Anthropic reports these classifiers trigger on fewer than 5% of sessions on average, so for routine work the handoff is invisible.
Three domains are covered. Cybersecurity — exploitation, offensive cyber tasks, and agentic hacking. Biology and chemistry — currently a broad net that Anthropic acknowledges is overly broad, with narrowing planned so that legitimate biomedical work is not caught. And distillation — extraction attacks designed to siphon the model's behavior to train a competing model. The biology breadth is the one to watch if your work touches health, pharma, or life-sciences research: until the safeguards narrow, legitimate queries may get the Opus 4.8 fallback rather than Fable 5's full capability.
On how well the safeguards hold: Anthropic ran more than 1,000 hours of external bug-bounty red-teaming and reports no universal jailbreak was found, though the UK AI Safety Institute made progress toward one within an initial testing window. An external partner reported zero harmful single-turn responses across 30 public jailbreak techniques, and in blocking mode Fable 5 made 0% progress on exploitation and offensive cyber tasks. Anthropic frames the goal plainly: make any remaining jailbreaks slow and costly enough to detect and stop before they are used at scale.
Exploitation & offensive cyber
Queries on exploitation, offensive cyber tasks, and agentic hacking are blocked on Fable 5 and answered by Opus 4.8. This is why ExploitBench shows 78.0% for Mythos 5 but Fable 5 makes 0% progress on offensive cyber in blocking mode.
Broad net, narrowing planned
Currently a wide safeguard that Anthropic admits is overly broad. Narrowing is planned so legitimate biomedical research is not caught. Watch this if your work touches health, pharma, or life sciences.
Extraction attacks
Attempts to extract the model's behavior at scale to train a competing model are detected and blocked. This protects the capability investment rather than the end user, but it shapes how high-volume, repetitive extraction-style prompting is handled.
Full capability, no handoff
Coding, analysis, content, research, agentic delivery, vision, and knowledge work run on Fable 5 at full strength. For the work most businesses do, the safeguards never fire and the routing is invisible.
The number you read about in the cybersecurity and biology benchmarks belongs to a model you cannot buy. The model you can deploy is deliberately constrained in exactly those domains — and for almost every business, that constraint never touches the work.Digital Applied analysis, June 9, 2026
04 — Pricing, Access & RolloutWhat it costs, where to get it, and the credit window to plan around.
Pricing. Fable 5 runs at $10 per million input tokens and $50 per million output tokens. A 90% discount applies to input via prompt caching, and US-only inference is available at a 1.1x multiplier on both input and output. The number that matters for budgeting: this is double Opus 4.8's $5/$25 rate card. The Opus 4.7-to-4.8 jump was a same-price upgrade you could adopt without re-running your cost model. Fable 5 is not — it is a premium tier, and the decision is where its extra capability is worth twice the token cost.
Where to get it. Fable 5 is available on the Claude API, on claude.ai across Pro, Max, Team, and Enterprise plans, and through AWS, Google Cloud, and Microsoft Foundry. Mythos 5 is not on any of these — it is restricted to Project Glasswing partners, with trusted-access programs planned for vetted cybersecurity organizations (by application) and biomedical researchers (who get the biology and chemistry safeguards removed while cyber safeguards stay in place).
The rollout you need to plan around. On subscription plans, Fable 5 is included at no extra cost from June 9 through June 22. From June 23, using it on a subscription plan draws on usage credits. Anthropic says it will restore Fable 5 as a standard inclusion once capacity allows. If your team is building a Fable 5 evaluation on claude.ai, run it inside the free window and budget for credits afterward — do not assume the June launch pricing persists into July.
Included on Pro, Max, Team & seat-based Enterprise
Fable 5 is available at no extra cost on subscription plans during the launch window. The best time to run an evaluation against your real workloads.
Usage credits required on subscription plans
After June 22, using Fable 5 on a subscription plan consumes usage credits. API and cloud usage is billed per token throughout at $10/$50.
Restored as a standard inclusion once capacity allows
Anthropic says Fable 5 will return as a standard part of subscription plans when capacity catches up to demand. No firm date given.
The price step-up
Fable 5's $10/$50 is double Opus 4.8's $5/$25. The capability is higher, but so is the cost — this is a tier decision, not a drop-in upgrade.
05 — Team Decision GuideWhen to reach for Fable 5, and when Opus 4.8 is still the call.
Fable 5 is built for a specific kind of work. Anthropic positions it for ambitious, long-running, asynchronous tasks that previous models could not sustain: agents that work for days, planning across stages and delegating to sub-agents; large codebase migrations and multi-day autonomous coding sessions; and complex, multi-stage knowledge work that needs minimal oversight. Its vision is strong on the messy real-world inputs that matter in business — diagrams, charts, and tables nested inside files and PDFs. At the highest effort, it reflects on and validates its own work before returning it.
When Fable 5 earns the premium. The clearest cases are the hardest ones: a framework migration across thousands of files, a multi-day research-and-build task you would otherwise break into a sprint, a knowledge-work pipeline where the cost of a missed detail outweighs the token bill. If the work genuinely benefits from sustained autonomy and self-verification, the doubled token cost is easy to justify against the senior-hours it replaces. For teams running AI transformation programs, Fable 5 is the model to point at the work that used to need a person watching every step.
When Opus 4.8 is still the right default. For routine, high-volume, or latency-sensitive work — classification, summarization, drafting, interactive chat, most day-to-day agentic tasks — Opus 4.8 at half the price remains the sensible baseline. It is also what Fable 5 falls back to on safeguarded topics, so if your workload lives near cybersecurity or biology, you may be paying the Fable 5 premium to receive Opus 4.8 answers anyway. The disciplined approach is to run a representative sample of your real tasks on both tiers, measure the quality difference and the token spend, and route by task rather than standardizing on one model.
The agentic angle. The capability that should most interest delivery teams is the run-for-days, sub-agent-delegating, self-validating behavior — it is the same direction of travel as the dynamic-workflow patterns that shipped with Opus 4.8. If your team has already built around parallel sub-agents and adversarial verification, Fable 5 raises the ceiling on what a single long-running session can finish without supervision. The teams that invested in those patterns are the ones positioned to get the most out of this release.
Fable 5 is not a same-price upgrade you flip on everywhere — it is a premium tier you point at the hardest, longest-horizon work. The skill is routing by task: Fable 5 where sustained autonomy pays for itself, Opus 4.8 for everything else.Digital Applied analysis, June 9, 2026
06 — Caveats & RoadmapThe honest trade-offs and what Anthropic has signaled next.
Do not benchmark-shop on Mythos numbers. The single most important caveat for anyone evaluating this release: the starred cyber and biology figures belong to the restricted model. A Fable 5 deployment performs closer to Opus 4.8 on those exact tasks. If a vendor, a slide, or a comparison cites the 78.0% ExploitBench number as if it were the model you can buy, it is overstating what you will actually receive.
The price step-up is real. At $10/$50, Fable 5 is twice Opus 4.8. There is a genuine cost decision here that the Opus 4.x line did not force. Model your task distribution before standardizing on it.
Biology safeguards are broad for now. Anthropic describes the current biology and chemistry net as overly broad, with narrowing planned. Until that lands, legitimate biomedical, pharma, and life-sciences queries may get the Opus 4.8 fallback. If that is your domain, test before you commit.
Some specs are not public. Anthropic's announcement did not state Fable 5's context-window size or maximum output tokens. We are not going to guess at numbers that were not published — verify these against the model documentation before you architect around a specific context length.
The alignment picture. Anthropic assesses Fable 5 / Mythos 5's misaligned behavior as low and similar to Opus 4.8, with a 30-day data-retention requirement on Mythos-class traffic, no use of that traffic for training, and logging of all human access. The honest framing is that this model class carries higher inherent risk than models before it — which is precisely why the high-capability tier is gated rather than sold.
What is next. Anthropic signaled two directions: expanding Mythos 5 access through trusted-access programs for cybersecurity organizations and biomedical researchers, and narrowing the biology safeguards so Fable 5 is less likely to over-block legitimate work. Project Glasswing itself is expanding to roughly 150 new organizations across more than 15 countries — the context for which is in our Glasswing expansion analysis. A natural follow-up to this post is a head-to-head of Fable 5 against GPT-5.5 and Gemini 3.1 Pro on the benchmarks that decide real deployments.
Fable 5 is the strongest model you can deploy — as long as you read the asterisks and price the premium.
The headline is the split. Anthropic decided its newest model was capable enough in cybersecurity and biology that releasing it openly would be irresponsible, so it shipped two products from one model: Fable 5 for general use with safeguards, and Mythos 5 for vetted partners with some of those safeguards lifted. That governance choice is as significant as any benchmark on the page.
For teams, the practical read is clear. Fable 5 leads the field on the capabilities businesses actually use — agentic coding, knowledge work, vision, legal, health — and it is built for the long-horizon, self-validating, sub-agent work that is hard to get from a cheaper model. The two things to keep straight: the dazzling cyber and biology numbers belong to a model you cannot buy, and the model you can buy costs twice what Opus 4.8 does. Run your real workloads against both tiers in the free window, route by task, and treat the asterisks as the line between marketing and what you will ship.