Claude Fable 5 is not 'Opus 5': Why Anthropic created a new tier
On June 9, 2026, Anthropic announced Claude Fable 5.
Let me give you the conclusion first.
The essence of Fable 5 is not performance improvement. It is that Anthropic has begun selling 'model capability' and 'access rights' as separate products.
The simultaneously announced Fable 5 and Mythos 5 are the exact same model under the hood. The only difference is the configuration of their safeguards. For the same capability, who it is opened to and to what extent has branched as a product—this is a first in the history of AI.
Until now, Claude has stubbornly maintained the three-tier structure of Haiku → Sonnet → Opus. If performance improved, they could have just advanced the number to 'Opus 4.8'. Instead, they created a new tier called the 'Mythos class' above Opus and gave it a different name. This change in naming is a declaration of a structural change.
In this article, I will present three reasons why it is 'Fable' rather than a continuation of Opus, and I will organize the grounds for why this is a reproducible model rather than a one-off demo, as well as its impact on general users, engineers, and management respectively.
━━━━━━━━━━━━━━━
1. Why 'Fable' instead of 'Opus 5'
━━━━━━━━━━━━━━━
◆ Reason 1: The way performance improved was not continuous
It is not about improving accuracy on short tasks, but rather that 'the gap widens as tasks become longer and more complex'—this is the characteristic emphasized in the official announcement. It scored 80.3% on SWE-Bench Pro for agentic coding (Opus 4.8 was 69.2%). This is a qualitative change that cannot be expressed just by advancing a number.
◆ Reason 2: Models and access rights have been separated
・Fable 5: Public release version. Queries in high-risk areas (cybersecurity, biology, etc.) are automatically routed to Claude Opus 4.8.
・Mythos 5: For approved organizations. A version with some restrictions lifted. Provided via Project Glasswing (a joint cyber defense framework with the US government and major tech companies).
I repeat, the contents are identical. The Sonnet/Opus naming system can only express 'levels of capability.' A new namespace was needed for a world where distribution policies branch for the same capability.
◆ Reason 3: There is a backstory of it being 'too strong to release'
The Mythos class was already released in limited form as Mythos Preview this April. The reason was that 'even though it wasn't designed for cybersecurity, its ability to discover vulnerabilities in major OSs and browsers was too high.' Anthropic first distributed it to the defense side of over 150 organizations (AWS, Apple, Google, Microsoft, JPMorgan Chase, etc.) and released it this time after waiting for safeguards to be put in place.
Fable 5 is not an 'improved version of Opus,' but rather the 'release of a model that had been withheld from the public.' From its origin, it is not an extension of the existing series.
━━━━━━━━━━━━━━━
2. Is this a reproducible model?
━━━━━━━━━━━━━━━
Clearing Pokémon FireRed with only screenshots, playing 'Factorio' autonomously—flashy demos are a hot topic, but a demo is not proof of reproducibility. There are three grounds on which I judge this to be 'not a one-off'.
[Evidence 1: Anthropic itself quantitatively demonstrates that the general version = the unrestricted version]
This is the most important point. Safeguards are triggered in less than 5% of sessions. According to Anthropic's initial data, in the over 95% of sessions where they are not triggered, Fable 5's performance is effectively identical to the unrestricted Mythos 5. In other words, the version that general users interact with is not a 'special configuration for demos.' The provider itself guarantees with numbers that the same capabilities are delivered regardless of who uses it.
[Evidence 2: Practical use cases are backed by 'line counts and days']
In early testing, Stripe reported that months of engineering work were compressed into a few days. The official announcement cites a case where a migration of an entire 50-million-line Ruby codebase was completed in one day (which would have taken an entire team over two months if done manually). The point is that it is being discussed in terms of scale, not just 'it got faster.'
[Evidence 3: Independent third-party verification points in the same direction]
Ethan Mollick of the University of Pennsylvania reported being able to digest multi-page specifications after up to 12 hours of continuous operation. It achieved the highest frontier score in Cognition's FrontierCode (a coding evaluation with production-quality criteria), and data analytics firm Hex reported, 'The first to exceed 90% in complex analysis benchmarks, a 10-point jump over Opus.' It was available on AWS (Bedrock), Google Cloud, and Microsoft Foundry from the day of the announcement, which is circumstantial evidence that pre-verification had been underway.
There are also points to keep in mind. Pricing is reported at $10 per million tokens for input and $50 for output, about twice that of Opus 4.8, and there are a certain number of false positives due to conservative safeguards. 30-day data retention for safety monitoring purposes is mandatory for all traffic in the Mythos class (it is explicitly stated that it is not used for training and is deleted after 30 days). The current assessment is that 'reproducibility is shown by the numbers, but it is a model where the application domain must be chosen.'
━━━━━━━━━━━━━━━
3. Who needs it, and when? — Thinking in three layers
━━━━━━━━━━━━━━━
This is not a 'model everyone should use.' Since costs have increased, we must consider which use cases justify using the new tier. I will organize this into three layers.
◆ General layer: The upper limits of 'researching and creating' change
[Use Case 1: 'Heavy research' at life milestones]
Comparing home purchases, insurance reviews, inheritance, and job change conditions. Research that involves cross-referencing multiple systems and documents has typically resulted in AI 'losing context midway' in the past. With Fable 5, which is strong at long-term autonomous research, it is possible to entrust it with 'research that takes several days'—crossing dozens of documents to detect contradictions while creating comparison tables.
[Use Case 2: Prototyping personal projects]
Generating a working game from a single prompt, building a custom browser-based CAD editor to design 3D-printable models—the official demo shows that the barrier from 'idea to working product' has been lowered another notch. This is about changing the prototyping costs for personal side hustles and small businesses.
How to use: Conventional models are sufficient for daily questions. Call on this only for 'research or production that would take a human several days.'
◆ Engineer layer: The 'granularity of delegation' changes
[Use Case 1: Legacy migration and large-scale refactoring]
The 50-million-line migration case shows a shift from 'function-level completion' to 'delegation by migration project.' If you can have it handle the entire sequence—investigating dependencies, step-by-step rewriting, and test verification—the premise for the cost of paying off technical debt changes.
[Use Case 2: Long-duration multi-agent workflows]
On harnesses like Claude Code, you can run a loop of planning, delegation to sub-agents, and self-verification for an entire day. I personally build orchestrator-type multi-agent configurations using AWS Step Functions + Bedrock, and when the worker runtime limit is extended, the placement of HITL (human-in-the-loop approval points) changes from 'checking in detail' to 'delegating broadly and approving at key points.'
How to use: Sonnet-class models are sufficient for daily coding assistance. Incorporate Fable 5 into your model routing as a top-tier lane dedicated to 'complex, long, and high-failure-cost tasks.' The design where Fable 5 itself falls back to Opus 4.8 for high-risk queries is an official implementation example of this routing philosophy.
◆ Management layer: 'AI procurement strategy' changes
[Use Case 1: In-house decision-making for high-value knowledge work]
The official announcement emphasizes performance in specialized fields such as finance and law. Due diligence, contract review, market research—how much of the 'work previously outsourced to external experts at high rates' can be taken over by AI + internal review? Re-evaluating this line becomes a practical agenda.
[Use Case 2: Integration into cyber defense posture]
The origins of the Mythos class are in cybersecurity. In Project Glasswing, financial institutions and medical networks have used it to discover and fix vulnerabilities, and Anthropic intends to expand the availability of Mythos 5 through its Trusted Access Program. Assuming that 'the attacker side will eventually have equivalent capabilities,' whether to incorporate this class into the defense side becomes a new point of discussion for security investment.
How to use: Instead of a company-wide uniform rollout, adopt a 'model portfolio' approach: allocate it specifically to tasks that (1) have high failure costs and (2) are high-unit-cost in terms of labor, while routing routine tasks to cheaper tiers.
━━━━━━━━━━━━━━━
4. Summary
━━━━━━━━━━━━━━━
Returning to the conclusion from the beginning: The essence of Fable 5 is not performance improvement, but the commercial separation of 'capabilities' and 'access rights'.
・General users: 'research and production taking several days', Engineers: 'project-based delegation', Management: 'in-house high-value operations and defensive posture'—this is the line where the rationality of using the new tier stands.
・Until June 22, it is included at no additional cost in Pro/Max/Team/Enterprise (seat-based) plans, and from the 23rd onwards, it is scheduled to transition to a usage credits system until capacity is stabilized. If you want to try it, do it this week.
It is not that 'the strongest model has arrived,' but rather that 'an era where design skills for utilizing different models are required has truly begun.' I believe this is the essence of this release.
I am also running comparative verification with Opus 4.8 on Claude Code (measuring completion rates and costs for long-duration agent tasks), and I will summarize the results on Zenn.
━━━━━━━━━━━━━━━
Reference Links
━━━━━━━━━━━━━━━
Claude Fable 5 and Claude Mythos 5 (Anthropic Official)
https://www.anthropic.com/news/claude-fable-5-mythos-5
Anthropic releases Mythos-like AI model to the public (CNBC)
https://www.cnbc.com/2026/06/09/anthropic-mythos-claude-fable-5.html
Claude Fable 5 is now available in Microsoft Foundry (Microsoft Azure Blog)
https://azure.microsoft.com/en-us/blog/claude-fable-5-is-now-available-in-microsoft-foundry-powering-the-next-era-of-autonomous-agents/
Anthropic's Claude Fable 5, Mythos 5: What you need to know (Constellation Research)
https://www.constellationr.com/insights/news/anthropics-claude-fable-5-mythos-5-what-you-need-know
Anthropic Launches Claude Fable 5 (MacRumors)
https://www.macrumors.com/2026/06/09/anthropic-fable-5/
※This article is based on information publicly available as of June 10, 2026. Pricing, terms of service, and benchmark values are subject to change.
━━━━━━━━━━━━━━━
About the Author
━━━━━━━━━━━━━━━
I post content themed around Cloud (AWS) × Generative AI × Career. Details of technical verification are updated on Zenn, and daily insights are updated on X. I would be happy if you followed me.
Zenn (Technical Verification & Architecture Commentary)
https://zenn.dev/ebe_ryuki
Qiita
https://qiita.com/ryukiebe0911
X (formerly Twitter)
https://x.com/EBE_Ryuki
LinkedIn
https://www.linkedin.com/in/ryuki-ebe-4783373b3
Instagram
https://www.instagram.com/ryuk.i0911