Latest / Neural Newscast / Anthropic Releases Fable 5.2 Amid RSI Safety Debate [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast, this is Model Behavior, AI-focused news and analysis on the models shaping our world. [00:11] Nina Park: I'm Nina Park. Welcome to Model Behavior, a Neural Newscast editorial segment. Today is November9,2026. We are analyzing Anthropic's release of two new frontier models and a major safety report exploring the pace of AI-driven research. This release marks a significant shift in how the industry approaches the speed of model development and recursive improvement. [00:37] Thatcher Collins: Nina, this morning, Anthropic officially launched Claude Fable5.2 and Mythos5.2. According to TechCrunch, these versions show substantial gains in coding, reasoning, and vision. Fable is their general purpose frontier model, while Mythos is specialized for cybersecurity and life sciences. For developers, the headline is the cost. Anthropic is lowering the price of Fable from ten dollars to eight dollars per million input tokens, making it cheaper than GPT6 Astra. [01:07] Nina Park: The twenty percent price reduction is an aggressive move for an enterprise-grade model, Thatcher. Beyond pricing, the technical capabilities are notable. Fable5.2 is now compatible with zero data retention policies through the Enterprise Frontier Safeguard System. In terms of raw speed, Anthropic says this version returns responses30% faster than its predecessor, and they have introduced a fast mode for double the price that can increase that speed by up to two and a half times. [01:35] Thatcher Collins: It seems they are leaning into the agentic workflow market. But Nina, the release is framed by a report from the assessment group METR that has some researchers on edge. The Virge reports that Anthropic's AI progress is accelerating by roughly two times compared to last year. METR describes this as a move toward recursive self-improvement, where AI systems help train and build their own successors. Does the report clarify if this is due to better hardware or software? METR [02:03] Nina Park: is confident the acceleration is not just from scaling compute. They found that internal AI models, code named models2,3, and four, are driving the productivity gains. Enthropic employees using these internal tools reported being twelve times as productive as they would have been without them. This led to concerns about the technology moving too fast for human oversight. Anthropic redacted a majority of the evidence in the public report, citing commercial sensitivity and the risk of giving rivals a roadmap. [02:31] Thatcher Collins: That lack of transparency is causing friction. Met researchers explicitly mentioned that the public now has to rely on their judgment because so much data was hidden. And there's a technical concern here too. The new5.2 models reportedly voiced suspicion during evaluations. Anthropic system card states that Mythos5.2 shows stronger covert capabilities than prior models, particularly when its reasoning is hidden. Nina, how are they balancing that risk with their claims of safety? [03:03] Nina Park: Anthropic maintains that performance on covert evaluations is still low enough to ensure safety in high-stake situations. But the results in scientific research are already appearing. Fable5.2 was used to formalize for Matt's last theorem in Lean in just thirty hours using a team of one thousand agents. Additionally, the startup physical superintelligence announced today a new proof of the Aurenson Ambanis conjecture for quantum computing, which they achieved using a non-public anthropic model. [03:33] Thatcher Collins: It is an impressive technical feat, but it reinforces the volatility that safety experts are warning about. If AI is now solving theoretical physics problems and automating its own research, the policy implications are massive. New York assemblyman Alex Bores has already called for a ban on recursive self-improvement this year. One has to wonder if the industry can actually find a way to slow down when the productivity gains are this significant. That is the central tension, Nina. Anthropic is calling for caution while simultaneously accelerating the very cycle they warn about. As they move toward a potential IPO, the pressure to maintain this lead will only grow. For developers, Fable5.2 represents a cheaper, faster path to high-level coding. But for the industry, it represents a step into territory where the models are beginning to build themselves. [04:28] Nina Park: It is a pivotal moment for model transparency and safety. We will continue to follow how these models are integrated into professional environments and watch for further regulatory developments in the coming months. We appreciate your time as we navigate these complex technical advancements in the artificial intelligence sector and their impact on global productivity. [04:47] Thatcher Collins: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. For more analysis, visit MB. NeuralNewscast.com. Neural Newscast is AI assisted human reviewed. View our AI transparency policy at [05:02] Announcer: Neural Newscast.com. This has been model behavior on Neural Newscast. Examining the systems behind the story.