Anthropic releases Claude Mythos 5 and restricts the public variant Fable 5

„Ein strenger Sicherheitsfilter blockiere riskante Anfragen und verhindere, dass sich die KI mit heiklen Themen wie Cybersicherheit und Biotechnologie beschäftige“, the company stated. US government agencies and selected companies had previously received access to Mythos Preview, the predecessor version of the current model. These user groups can now use Mythos 5, the model version without restrictions.

Security concerns and political context

German domestic intelligence officials and security policymakers had subsequently warned that artificial intelligence could be used for devastating cyberattacks. A few months earlier, the predecessor model Mythos Preview had detected vulnerabilities in widely used software, fueling concerns that more powerful models could carry out such tasks autonomously in the future.

Anthropic released the model on a Tuesday, deviating from the usual Thursday schedule. „Die Leitplanken gegen potenziell gefährliche Nutzung seien ausgiebig getestet worden, hieß es weiter.“ The accompanying system card document for Mythos 5 and Fable 5 spans 319 pages and describes, among other things, the benchmarks tested.

Performance in tests

According to Anthropic's published table, the new model leads in 11 of 13 selected benchmarks and trails the preview variant narrowly in the remaining two. On vision tasks—that is, processing images—Anthropic describes Fable 5 as „Stand der Technik für Aufgaben, bei denen es auch um Sehen geht. Es kann präzise Zahlen aus detaillierten wissenschaftlichen Schautafeln extrahieren und komplexe bildabhängige Aufgaben ausführen, darunter den Nachbau des Quellcodes einer Web-App aus Screenshots“. Overall, the model is said to „alles übertreffen, was wir jemals allgemein verfügbar gemacht haben“.

Prices, access, and classifiers

For access, Anthropic is using tiered pathways. From 23 June, Fable 5 will be billed exclusively based on token consumption and will no longer be included in trial subscriptions. Pro, Max, Team, and Enterprise subscribers with seat assignments can test it for 14 days, but will consume double their usual quota during that period. Token prices are approximately twice as high as for Claude Opus 4.8: ten US dollars per million input tokens, 12.50 dollars for five-minute cache writes, 20 dollars for one-hour cache writes, one dollar per million cache read tokens, and 50 dollars per million output tokens.

Technically, Fable 5 uses smaller classifier models that monitor user inputs. If an input is classified as suspicious, the system switches to the older model Claude Opus 4.8 rather than rejecting the query outright. Opus 4.8 requires less computing power than Fable 5. Online, users had previously complained that they had been downgraded to Opus 4.8 even for innocuous questions such as interpreting a blood count.

Competitive pressure and accusations against competitors

Anthropic faces intensified competition and a defensive battle. In February, the company had accused the Chinese competitors Deepseek, Zhipu, and Moonshot of attempting to distill Claude through large-scale attacks. In the AI world, distillation means training other models on the outputs of existing models rather than working with raw data. In total, more than 24,000 fraudulent user accounts were deployed for 16 million such attempts.

„Diese Nachricht wurde am 10.06.2026 im Programm Deutschlandfunk gesendet“, the broadcaster reported. In the political context, Donald Trump had ordered that the restricted rollout of powerful AI models to the NSA and select partners be regulated as a voluntary measure.

Anthropic itself names one weakness: in at least one area, the new model hallucinates more than older versions—that is, it fabricates plausible-sounding but incorrect statements. Claude Opus continues to outperform the new models in this regard. According to reports, Anthropic is also preparing an initial public offering, which places the release in the context of a broader corporate strategy.