Latest / Neural Newscast / Anthropic Debuts Claude Fable 5 Mythos-Class Model [Model Behavior]
Transcript
- [00:00] Nina Park: From Neural Newscast, this is Model Behavior, [00:03] Nina Park: AI-focused news and analysis on the models shaping our world. [00:12] Nina Park: I am Nina Park. Welcome to Model Behavior. Today is June 10th, 2026. This podcast, [00:20] Nina Park: a neural newscast editorial segment, [00:22] Nina Park: examines how complex artificial intelligence systems are built, deployed, [00:27] Announcer: and operated in real-world professional environments. [00:31] Announcer: We focus on technical architecture and the practical impact of these tools. [00:36] Nina Park: I am Thatcher Collins. Today, Nina, we are analyzing a significant release from Anthropic. [00:43] Nina Park: They have officially made Claude Fable V available to the general public. [00:48] Nina Park: This marks a turning point [00:49] Nina Park: as it is their first Mythos class model to reach wide release. Previously, this specific [00:54] Thatcher Collins: tier of capability was strictly limited to [00:57] Thatcher Collins: a small group of vetted cybersecurity and infrastructure partners. [01:01] Announcer: That is correct, Thatcher. Reporting from 9 to 5 Mac indicates that Fable V is being framed as [01:07] Announcer: a massive step change in performance rather than an incremental update. [01:12] Announcer: To put the scale of this into perspective, Anthropic shared that Stripe tested Fable V [01:18] Announcer: against a massive 50 million line Ruby code base. The model successfully completed [01:22] Announcer: a complex migration in a single day, [01:25] Announcer: a task that would have occupied a full engineering team for two months. [01:30] Thatcher Collins: The naming convention here is quite interesting as well. We have the public Fable V [01:35] Thatcher Collins: and the more restricted Mythos V. It appears Anthropic is implementing [01:40] Thatcher Collins: a much more granular safety architecture for these releases. Fable 5 serves as [01:44] Thatcher Collins: the consumer-facing interface that utilizes specialized AI classifiers to monitor and intercept [01:50] Thatcher Collins: high-risk requests. If a prompt is flagged for sensitive cybersecurity or biological tasks, [01:56] Thatcher Collins: the system automatically routes it back to the less powerful Clawed Opus 4.8. [02:01] Announcer: That routing mechanism is triggered in roughly 5% of all sessions. While Anthropic acknowledges [02:07] Announcer: this can lead to some false positives for benign users, the specific capabilities they are shielding [02:12] Announcer: are substantial. Their internal research demonstrated that the unrestricted Mithos-5 variant [02:18] Announcer: could accelerate drug design processes by a factor of ten. It was capable of autonomously selecting [02:24] Announcer: binding sites and operating protein design tools producing [02:28] Announcer: strong candidates for nine out of fourteen protein targets. [02:31] Thatcher Collins: That autonomy is really the defining characteristic of this new class we are seeing it manifest in [02:38] Thatcher Collins: vision and memory capabilities as well. Fable V can navigate intricate virtual environments such [02:43] Thatcher Collins: as playing through the game Pokémon Fire Red using raw screenshots alone without any reliance [02:48] Thatcher Collins: on external maps. Furthermore, when playing Slay the Spire, its performance tripled [02:54] Thatcher Collins: when given access to persistent file-based memory, [02:57] Thatcher Collins: demonstrating a far superior grasp of long-term state compared to OPUS 4.8. [03:01] Announcer: The economics of these models are shifting just as rapidly both Fable V [03:06] Announcer: and Mythos V are priced at $10 per million input tokens and $50 per million output tokens [03:14] Announcer: this represents a price cut of more than 50% compared to the first Mythos preview release [03:19] Announcer: we saw back in April anthropic is clearly betting [03:23] Announcer: that lower costs will encourage developers to move toward these more complex agentic workflows [03:28] Nina Park: While Anthropic focuses on the technical side, [03:31] Nina Park: Meta has secured a landmark legal victory in the Northern District of California. [03:36] Thatcher Collins: Judge Vince Chabria has ruled that Meta's practice of using copyrighted books to train [03:41] Thatcher Collins: its llama models constitutes fair use. The case Kadri v. Meta has been one [03:47] Thatcher Collins: of the most closely watched legal disputes regarding the data sourcing used to build modern large [03:52] Announcer: language models. The court's logic centered on the concept of transformative use. Thatcher, [03:57] Announcer: the judge determined that training in AI is not about reproducing or replacing copyrighted books [04:03] Announcer: for human readers; instead, the process involves analyzing massive data sets to identify and learn [04:10] Announcer: statistical patterns. This creates a new functionality that is entirely distinct [04:16] Announcer: from the original artistic intent of the authors, [04:19] Announcer: which is a key pillar of a successful fair use defense. [04:21] Thatcher Collins: It was also critical that the plaintiffs could not demonstrate concrete evidence of market harm. [04:26] Thatcher Collins: Meta successfully argued that the existence of the llama models did not directly cause a loss [04:31] Thatcher Collins: in sales for the specific books cited in the lawsuit. However, the judge did note that this is not [04:36] Thatcher Collins: a blanket immunity for the industry. If future plaintiffs [04:40] Thatcher Collins: can better document market dilution or if data sourcing methods change, the legal outcome [04:45] Announcer: could vary. It feels like a pragmatic middle ground for the time being. [04:48] Announcer: It allows AI companies to continue iterating with existing datasets while serving as a reminder [04:54] Announcer: that rigorous documentation still matters. Between the launch of Anthropic's Mythos class models [04:59] Announcer: and this legal win for Meta, it is clear [05:02] Announcer: that the industry is entering a much more mature, albeit highly complex, operational phase. [05:08] Thatcher Collins: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. [05:14] Thatcher Collins: For more in-depth analysis and the full technical record of this episode, [05:18] Thatcher Collins: please visit mbnneuralnewscast.com. Neural Newscast is AI-assisted, [05:25] Thatcher Collins: human-reviewed. View our AI transparency policy at neuralnewscast.com [05:31] Nina Park: This has been model behavior on Neural Newscast examining the systems behind the story