Anthropic In the early morning of June 10 (Taiwan time)officially announced Claude Fable 5The first publicly available Mythos-class model, alongside the release of Claude for Project Glasswing partners Mythos 5Fable 5 has set the highest coding capability record among all publicly available models with an 80.3% score on SWE-Bench Pro, priced at $10/$50 per 1M tokens.Only half of Mythos PreviewThis is the biggest product update since Anthropic released the Mythos Preview in April. At the time, citing “capabilities had reached a dangerous threshold,” Anthropic limited access to only a handful of cybersecurity partners. Two months later, they’ve returned with a new safety classifier and said: “It’s now safe enough to let more people use it.”
Fable 5 vs Mythos 5: Same Model, Two Security Modes
Fable 5 and Mythos 5 are the same underlying model, differing in whether and how strictly the safety guardrails are applied.
- Claude Fable 5Open to the public, equipped with three types of safety classifiers (cybersecurity, biological/chemical, and distillation attacks). When the classifier detects high-risk requests, it automatically downgrades to Opus 4.8 responses rather than outright rejection. Anthropic stated that over 95% of conversations are completely unaffected.
- Claude Mythos 5For use by Project Glasswing partners only. Removes network security guardrails. The trusted access program will be gradually expanded, including a bio/chemical guardrail removal version for biomedical researchers.
Both models share the same pricing: $10 per million input tokens and $50 per million output tokens — half the price of Mythos Preview’s $20/$100.
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.
Its capabilities exceed those of any model we’ve ever made generally available. pic.twitter.com/2AvmEjHIX8
— Claude (@claudeai) June 9, 2026
Stripe Completes Two Months of Work in One Day
Fable 5 achieved 80.3% on SWE-Bench Pro, outperforming GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%).
But a more practical validation comes from Stripe. Stripe reported that in a 50 million line Ruby codebase, Fable 5 completed a full repository migration in just one day—a task that would have taken two engineers over two months to do manually. Fable 5 is also the most token-efficient among all models: on Cognition’s FrontierCode benchmark, it achieved the highest score with medium effort settings.
GitHub’s Head of Product Mario Rodriguez said after testing: “Fable 5’s automation and reliability in handling complex, long-duration coding tasks surpassed previous benchmarks.” Cursor CEO Michael Truell said: “Fable 5 is the SOTA model on CursorBench, unlocking a class of long-horizon problems that previous models couldn’t tackle.”
Visual Capabilities: Rebuilding Apps from Screenshots, Solo Pokémon Runs
Fable 5 also achieves a new SOTA on visual tasks. It can extract precise numbers from scientific charts in screenshots and reconstruct the source code of a web app solely from screenshots. Even more impressively, Fable 5 defeated Pokémon FireRed using a minimal pure-vision harness—a feat that previous Claude models struggled with even when given tooling.
Holy SHittttt Claude Fable 5 just finished Pokémon FireRed with vision alone 🤯
raw screenshots only
no map / no nav / no hidden game stateolder Claude needed a helper harness
This timelapse goes hardddddd…. https://t.co/DLrtcLNwL6 pic.twitter.com/rzOUP2O3yx
— Chetaslua (@chetaslua) June 9, 2026
Life Sciences: 10x Faster Drug Design, Autonomous Genomics Research
Mythos 5 has demonstrated genuine research capabilities in the life sciences. Anthropic’s protein design team used Mythos 5 to accelerate the drug design process by approximately 10x. In one case, Mythos 5, paired with protein design and bioinformatics tools, without human assistance, “matched or even surpassed skilled human operators.” 9 out of 14 protein targets yielded strong candidate molecules.
Even more impressive is the genomics research. In over a week of autonomous work, Mythos 5 compiled single-cell data spanning 138 species and millions of cells, designing and training a custom machine learning model to identify cell functions. With only high-level human input, this model surpassed recently published results in Science—while being 100 times smaller. Anthropic plans to publish these findings in the coming months.
Mythos 5 is also the first model capable of consistently generating “novel and compelling” scientific hypotheses. In blind comparisons, Anthropic’s scientists preferred Mythos’s molecular biology hypotheses over Opus-class models approximately 80% of the time. One hypothesis about a novel mechanism in E. coli proteins has been confirmed by an independent research laboratory.
Security Architecture: Classifier Detection + Graceful Degradation
The safety design of Fable 5 is the core of this announcement. Anthropic designed three types of classifiers:
- Network Security ClassifierCovering all stages of the attack chain including exploitation, reconnaissance, and lateral movement. Internal testing shows the classifier was able to completely prevent Fable from making progress on these tasks.
- Biological/Chemical ClassifierThe scope has expanded from narrow biological weapon-related queries to most biology and chemistry requests. Anthropic acknowledged this is a conservative setting and expects to gradually narrow the scope going forward.
- Distillation attack classifierPreventing large-scale extraction of Claude’s capabilities to train competing models
When the classifier triggers, the system doesn’t directly refuse but automatically switches to Opus 4.8 to respond and informs the user. In over 1,000 hours of external red teaming, no universal jailbreak method was found, though the UK’s AISI (AI Safety Institute) has made progress in preliminary testing.
Releasing a model this capable comes with risks. Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage.
Queries on a narrow range of topics will instead receive a response from our next-most-capable model, Opus 4.8. pic.twitter.com/vJ71vCdkjc
— Claude (@claudeai) June 9, 2026
Anthropic also announced that Fable 5 and Mythos 5 will implement a 30-day data retention policy, where all traffic data will be retained for 30 days for security monitoring, not used for model training, and all human access will be logged. This policy will also apply to all future models at the Mythos tier or above.
Mythos 5 Impact AreasIPO Targeting One Trillion Dollars
This announcement is not an isolated event. Anthropic is now valued at $965 billion (approximately NT$31.4 trillion), surpassing OpenAI’s $852 billion. Annualized revenue has reached $47 billion (approximately NT$1.53 trillion). Anthropic secretly filed an IPO application with the SEC on June 8, the same day as OpenAI. This puts the two companies in direct competition for the title of “the first AI company to go public,” while SpaceX is set to go public this Friday, with Anthropic and OpenAI’s IPOs expected in the second half of the year.
Fable 5 is now available through the Claude API (claude-fable-5) and Claude Pro/Team plans. Pro/Team users can use Fable 5 for free under their subscription until June 22, after which API billing will resume.
Source: KOCPC Chinese