- Anthropic released two new AI models, Claude Fable 5.1 and Claude Mythos 5.1, on September 1, 2026.
- Fable 5.1 works for everyone. Mythos 5.1 stays locked to vetted cybersecurity and life science experts.
- The new models cost less to run and make fewer safety mistakes than Fable 5, their predecessor.
Anthropic just gave its top AI models a fresh upgrade. The company launched Claude Fable 5.1 and Claude Mythos 5.1 on September 1. Anthropic calls them the world’s most advanced tools for coding and knowledge work. The move comes just three months after Anthropic first released Fable 5 and Mythos 5.
Both models share the exact same core system. The real difference lies in how each one handles safety. This split matters for anyone who relies on AI for daily tasks, from solo developers to large firms.
One Model, Two Sets of Rules
Fable 5.1 and Mythos 5.1 are built from identical code. Anthropic explains that the two only differ in their safety settings. Fable 5.1 keeps standard guardrails in place. It works well for regular business and coding tasks. Anyone with a paid plan or an API account can use it.
Mythos 5.1 drops some of those guardrails. It only serves vetted professionals through Anthropic’s trusted access programs. These programs mainly serve cybersecurity defenders and life science researchers.
Anthropic built Mythos this way because some advanced skills carry real risks. A tool that finds security holes could also teach someone to exploit them. So Anthropic keeps that power in fewer hands.
Right now, Mythos 5.1 mainly reaches select organizations in the US. Reports say Anthropic is working with the US government to widen access soon. The company wants more domestic and international partners to join over time.
Mythos 5.1 already powers Claude Security. This product scans company code for weak spots. It then suggests fixes for a human to review and approve, rather than making changes on its own.
Lower Prices, Fewer False Alarms
Money matters too. Anthropic says Fable 5.1 costs about 25 percent less than Fable 5 for normal workloads. Heavy, multi-step tasks can cost up to 45 percent less. Cache read prices dropped sharply, now sitting at 0.25 dollars per million tokens. Regular input and output prices stay the same, at 10 and 50 dollars per million tokens.
Fable 5.1 also makes fewer safety mistakes. Anthropic reports a 60 percent drop in false safety warnings for cybersecurity questions. The model can now spot software weaknesses for defense purposes. It still refuses to write actual exploit code, though. This keeps a clear line between finding flaws and abusing them.
Riskier jobs, like penetration testing, get sent straight to Anthropic’s Opus class models instead. This keeps the most dangerous tasks away from general users.
Data privacy got an upgrade as well. Anthropic built a new system called Enterprise Frontier Safeguards. It lets companies keep their data on their own servers, not on Anthropic’s. Anthropic still watches for misuse there. This setup should roll out later this year across Claude Code, Claude Enterprise, and several major cloud platforms.
Anthropic also ran its own internal tests on the new model. Fable 5.1 beat its predecessor on coding, reasoning, and general knowledge tasks. The tests included Terminal Bench, GDPval AA, and Humanity’s Last Exam.
Real companies are already testing Fable 5.1. Anthropic says the investment firm Millennium used the model to solve a rare system crash. Human engineers had struggled with that same bug for nearly five years. Other early partners include Jane Street, Cognition, MongoDB, and Datadog, all of whom reported strong early results.
New Wins in Science, and a Look Back
Mythos 5.1 stretched into pure science too. During tests, the model designed strong protein binders for medical research. It beat typical industry hit rates by a wide margin. Separately, Fable 5.1 built a detailed elevation map of Venus. It used decades old radar data from NASA’s Magellan mission, ahead of upcoming European and American space missions.
Anthropic also confirmed it follows the European Union’s AI Act rules. New outputs from Claude now carry invisible watermarks. A detection tool for regulators and researchers is in early testing too.
The watermarks are tied directly to the EU AI Act’s transparency rules, which took effect on August 2, 2026. Claude models released in the EU on or after that date carry embedded machine-readable marks from day one. The text watermark stays invisible to readers, survives copying and pasting, and does not change what a response means or how it reads. Files that Claude generates, such as PNG, JPG, and SVG images, also carry signed origin metadata.
This is not the first time these model names have made news. Anthropic paused the earlier Fable 5 and Mythos 5 models back in June 2026. US export control rules forced that decision at the time. Anthropic restored full access by the end of that same month, once the rules changed.
Fable 5.1 is available right now. Anyone can use it through Claude’s regular platforms. It is also open through the API, under the name claude fable 5 1. Mythos 5.1 remains reserved for vetted experts for the time being, though Anthropic says that circle should grow.
Independent reports back up several of these claims. Coverage from tech outlets notes that Fable 5.1 topped rival models from other AI labs. It did so on multiple public benchmarks. Analysts also point out that the timing lines up with fresh competition across the AI industry.
Rival companies are preparing their own new releases. For now, Anthropic frames this launch as proof. Cheaper, safer AI tools do not have to mean weaker ones.
