SAN FRANCISCO – Anthropic has launched Claude Fable 5 and Claude Mythos 5, introducing a new “Mythos-class” of models designed for autonomous, long-horizon tasks in software engineering, finance, and scientific research.
The release marks a strategic shift toward agentic AI, moving beyond conversational interfaces to models capable of executing complex, multi-step workflows with minimal human intervention. For policymakers and corporate boards watching how AI systems make and execute decisions, the launch underscores how quickly frontier models are moving from assistive tools to semi-autonomous operators inside critical workflows.
By pricing the new models at a fraction of previous versions, Anthropic is positioning itself to capture a larger share of the enterprise API market, directly challenging the pricing structures of competitors like OpenAI and Google.
Enterprise Scaling and Pricing Strategy
The Fable 5 and Mythos 5 models are priced to accelerate adoption across the corporate sector, significantly reducing the cost of high-tier reasoning and long-context analysis for large internal datasets.
| Token Type | Price per Million Tokens |
|---|---|
| Input Tokens | $10 |
| Output Tokens | $50 |
This pricing represents a reduction of more than 50% compared to the Claude Mythos Preview and is calibrated for large enterprises that can consume billions of tokens per month.
The rollout for subscription-based users follows a strict timeline to manage capacity and give Anthropic time to monitor real-world safety behavior at scale:
- Through June 22, 2026: Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no additional cost.
- June 23, 2026: Fable 5 will be removed from standard subscription inclusions and will require usage credits.
- Post-June 23: The model will be restored to standard plans as compute capacity allows, with enterprise access prioritized for high-volume and regulated customers.
Executives and procurement teams weighing deployment will see the new pricing land in a landscape already under scrutiny from competition and digital markets authorities, which are examining whether concentration in cloud and AI infrastructure could distort pricing power and access over time.
Agentic Capabilities in Corporate Workflows
The Mythos-class models demonstrate a significant increase in autonomy, particularly in software engineering and financial analysis, where systems must plan, execute, and verify long sequences of actions.
Stripe reported that Fable 5 completed a codebase-wide migration in a 50-million-line Ruby codebase in a single day, a task that previously required a full engineering team over two months. While such internal testimonials are difficult for outsiders to independently verify, they highlight how quickly AI tools are moving into work once considered core to in-house engineering staff.
In the financial sector, the model achieved the highest score on Hebbia’s Finance Benchmark for senior-level reasoning. IMC noted the model’s proficiency in trading-analysis evaluations, specifically in root-cause and expected-value analysis, a domain where errors can quickly translate into material risk and regulatory attention.
“Claude Fable 5 is a real step forward for the developers GitHub serves. In our early testing, it took on complex, long-horizon coding tasks with a level of autonomy and reliability that exceeded previous benchmarks,” said a representative from GitHub.
The new systems are designed to operate not just as chatbots but as orchestration layers that can call tools, manage memory across long projects, and hand off work between human and machine, raising fresh questions about accountability when AI-driven workflows touch critical infrastructure, finance, or safety-related code.
Government Collaboration and Cybersecurity
Claude Mythos 5 is deployed through Project Glasswing, a collaboration with the U.S. government focused on applying frontier AI models to cyber defense and critical infrastructure resilience. This version of the model has specific cybersecurity safeguards lifted or relaxed to assist cyberdefenders and infrastructure providers, while remaining gated behind a controlled access program.
The model is described as having the strongest cybersecurity capabilities of any model globally, including offensive-security style reasoning that would be too risky to expose directly to the general public. Anthropic intends to expand access to Mythos 5 through a systematic trusted access program for cybersecurity organizations, aligning with emerging expectations from frameworks such as the White House’s Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, which calls for tighter oversight of powerful dual-use systems.
For governments, this model of “high-capability but restricted” deployment offers a template for how states might leverage advanced AI for national cyber defense while trying to reduce the risk that the same tools leak into criminal or state-backed offensive use.
Governance and Safety Architecture
To mitigate the risks of “uplift” for malicious actors – scenarios where a model substantially raises the capabilities of a would-be attacker – Anthropic has implemented a system of safety classifiers. These separate AI systems detect potential misuse in cybersecurity, biology, chemistry, and model distillation before the underlying model sees the full request.
When a classifier triggers, the request is automatically rerouted to Claude Opus 4.8, a more heavily constrained model. According to internal data, these fallbacks occur in less than 5% of sessions, suggesting that most traffic proceeds without intervention but that the company expects a non-trivial tail of high-risk queries.
Anthropic has also instituted a new data retention policy for all Mythos-class traffic. All data on first- and third-party surfaces will be retained for 30 days to defend against novel jailbreak attacks and reduce false positives, a move likely to draw close attention from privacy officers and data protection regulators in jurisdictions with strict data minimization rules. Anthropic says this data is not used for training purposes.
The company frames these architectural choices as a step toward shared governance of powerful models, in which independent auditors, regulators, and enterprise customers can eventually inspect logs, safety performance, and red-teaming results rather than relying solely on vendor assurances.
Application in Life Sciences and Genomics
Anthropic is extending its collaboration with the U.S. Department of Commerce and other agencies to apply Mythos-class capabilities to biomedical research – a domain where most major governments are simultaneously encouraging AI-enabled innovation and tightening controls on dual-use biological research under export, biosafety, and public health regimes.
Internal experts reported that Mythos 5 accelerated aspects of the drug design process by approximately ten times. The model was able to choose binding sites and run protein design tools without human assistance, matching the performance of skilled human operators in several studies, according to Anthropic. For drug developers and regulators, such capabilities raise the prospect of faster preclinical pipelines but also intensify debates over how to vet AI-generated candidates for safety and how to document model involvement in regulatory filings.
In genomics, Mythos 5 conducted over a week of autonomous research, assembling single-cell data for millions of cells across 138 animal species. The resulting custom machine learning model reportedly outperformed a recent study published in the journal Science, suggesting that advanced agentic systems may soon play a direct role in designing and validating scientific models rather than merely summarizing literature.
Claude Fable 5 is currently available via the Claude API and standard subscription plans, while Mythos 5 remains restricted to Glasswing partners and select biology researchers. As regulators in the United States, Europe, and elsewhere move from voluntary commitments toward binding AI rules, deployments like Mythos 5 – high-stakes, high-autonomy, and tightly gated – are likely to become early test cases for how frontier models are supervised in practice.
