Latest / Neural Newscast / Anthropic Debuts Fable 5 and Revises Research Limits [Model Behavior]
Transcript
- [00:00] Nina Park: From Neural Newscast, this is Model Behavior, [00:02] 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 11th, 2026, and we are examining [00:20] Nina Park: how the latest generation of AI systems is being architected and deployed in [00:24] Nina Park: high-stakes professional environments. We are looking at a pivot point where simple text generation [00:31] Nina Park: is giving way to complex agency and deep integration with existing global infrastructure. [00:37] Nina Park: Our focus today is on the technical [00:39] Nina Park: and ethical boundaries being pushed by both established labs and industry pioneers. [00:45] Thatcher Collins: I am Thatcher Collins. We have been tracking a significant week for frontier models, [00:50] Thatcher Collins: characterized by a major release from Anthropic and a notable philosophical shift [00:55] Thatcher Collins: in research circles. We are seeing a new class of models emerge, but at the same time, [01:01] Thatcher Collins: prominent figures are questioning if we have already reached the functional ceiling [01:06] Thatcher Collins: of the current large language model paradigm. It is a dense day for both policy [01:10] Thatcher Collins: and product news as we look at how these systems are fundamentally changing. [01:14] Nina Park: Let us start with the news from Anthropic. Yesterday, the company officially launched [01:19] Nina Park: Clawd Fable V. This release is particularly noteworthy because it represents [01:25] Nina Park: the first publicly available model from their new Mythos class of architectures. [01:29] Nina Park: For quite some time, Claude Opus served as their premier offering for enterprise-grade reasoning. [01:34] Nina Park: However, Fable V reportedly surpasses Opus in nearly every internal and external benchmark, [01:41] Nina Park: specifically in core analytical reasoning, [01:43] Nina Park: where it achieved an accuracy rate exceeding 90% on graduate-level evaluations. [01:48] Thatcher Collins: The capabilities of this mythos class have been a point of intense speculation for months, Nina. [01:55] Thatcher Collins: Anthropic actually debuted the architecture back in April through a restricted program known as [01:59] Thatcher Collins: Project Glasswing. However, they initially limited access to a handful of strategic partners, [02:06] Thatcher Collins: including Apple and Nvidia. The internal concern was that the model's reasoning capabilities were [02:11] Thatcher Collins: so advanced that it posed a safety risk, specifically regarding the discovery [02:16] Thatcher Collins: of novel software vulnerabilities or even assisting in sophisticated biological research. [02:21] Nina Park: That extreme caution resulted in a specific policy that caused quite a stir [02:26] Nina Park: within the developer community. When Fable V was first introduced into the testing phase, [02:31] Nina Park: anthropic was allegedly degrading the performance of the model for certain users [02:35] Nina Park: without notification. If a system prompt was identified as being related to [02:39] Nina Park: frontier model development or AI research, the system would silently fall back [02:44] Nina Park: to a less capable model. Essentially, they were throttle-limiting [02:47] Nina Park: the intelligence of the model based on the agent's intent to protect their competitive edge. [02:53] Thatcher Collins: Developers were certainly frustrated when those details surfaced. Business Insider [02:58] Thatcher Collins: recently reported that many in the industry viewed this not as a strategic move [03:03] Thatcher Collins: to prevent competitors from using Claude to build better rival systems. [03:07] Thatcher Collins: It felt less like a safeguard and more like a competitive barrier disguised as [03:12] Thatcher Collins: corporate responsibility. there has been a formal response from Anthropic regarding this [03:17] Thatcher Collins: what is the official stance now that the model has been fully released to the public? [03:21] Nina Park: They have moved toward a policy of greater transparency. Anthropic issued a formal statement [03:26] Nina Park: on Wednesday, apologizing for what they described as the wrong trade-off between safety [03:31] Nina Park: and user experience. Going forward, Fable 5 will clearly indicate when a request is being refused [03:37] Nina Park: or rerouted. If a request is flagged, it will visibly fall back to a specialized version [03:43] Nina Park: of OPUS 4.8, and the API response will include an explicit reason [03:47] Nina Park: for that fallback. It is a necessary step toward transparency over silent mediation. [03:53] Thatcher Collins: It is a quick admission of a mistake, which is rare in this industry. But even [03:58] Thatcher Collins: with that controversy, the performance gains are quite striking. Early reports suggest that Fable V [04:05] Thatcher Collins: is approximately 25 to 30 percent faster than Opus when handling complex multi-step tasks. However, [04:12] Thatcher Collins: that performance comes with a significant price tag. Pricing is now set at $10 per million tokens [04:17] Thatcher Collins: for input and $50 per million tokens for output. [04:21] Thatcher Collins: That is exactly double the current cost of the OPUS model, which remains available. [04:26] Nina Park: A 100% increase in API costs suggests that Anthropic views Fable V as a highly specialized tool [04:33] Nina Park: for high-value reasoning. It forces developers to make a difficult calculation is a 10% increase [04:39] Nina Park: in accuracy worth doubling their operational expenses? For many standard automation tasks, [04:44] Nina Park: Anthropic is actually suggesting that Feeplo V might be overkill and that most users [04:49] Nina Park: should remain on the standard Clode family [04:51] Nina Park: for their daily production needs while saving the Mythos class for specialized research. [04:56] Thatcher Collins: While Anthropic continues to push the boundaries of current scaling laws, [05:01] Thatcher Collins: Jan Lacun is making a very different argument. Lakun, who previously served as [05:05] Thatcher Collins: the chief AI scientist at Meta, is currently spearheading a $1 billion [05:11] Thatcher Collins: initiative to find a successor to the large language model. He is advocating for something [05:16] Thatcher Collins: called JEPA or Joint Embedding Predictive Architecture. He has become increasingly vocal [05:22] Thatcher Collins: recently describing current LLM technology as a technological dead end [05:27] Thatcher Collins: that will never reach human-level reasoning or true world understanding. [05:31] Nina Park: That is an incredibly bold statement to make, when the entire industry is currently centered on [05:37] Nina Park: transformer-based language models. Thatcher, could you unpack the technical reasoning [05:42] Nina Park: behind his skepticism? Why does he believe that models like GPT or the Claude series [05:47] Nina Park: are fundamentally limited despite the impressive capabilities we are seeing today?" [05:52] Thatcher Collins: "The core of his argument, Nina, is that large language models are [05:56] Thatcher Collins: essentially sophisticated statistical engines. They excel at predicting the next word in a sequence [06:02] Thatcher Collins: based on patterns, but they lack a fundamental understanding of the physical world [06:07] Thatcher Collins: or the laws of causality. This is why we see persistent hallucinations. [06:11] Thatcher Collins: The models lack grounded knowledge. Jepa is designed to learn more like a human infant by observing [06:18] Thatcher Collins: vast amounts of raw video data to build [06:21] Thatcher Collins: abstract representations of physical reality rather than just processing strings of text. [06:26] Nina Park: So, rather than just reading the entire Internet, the Jepa architecture is designed to watch [06:32] Nina Park: and learn from the world. Reports from geeky gadgets suggest that this architecture is [06:37] Nina Park: highly modular, consisting of six specific components, including a perception module [06:42] Nina Park: and a world model. This design is specifically intended to create an artificial intelligence [06:48] Nina Park: that can plan and simulate actions in dynamic environments. That is a capability [06:53] Nina Park: that current language models [06:54] Nina Park: struggle with because they cannot easily model physical consequences or spatial relationships. [07:00] Thatcher Collins: That is exactly the point. In this architecture, the world model simulates and predicts [07:06] Thatcher Collins: future events while a separate cost module evaluates those potential actions for their efficiency [07:12] Thatcher Collins: and safety. Because it is modular, JPA can theoretically adapt to new tasks [07:19] Thatcher Collins: with very little retraining. This is a contrast to the massive, expensive recomputes required for [07:25] Thatcher Collins: large language models. The ultimate goal here is to achieve true intelligence by building a system [07:32] Thatcher Collins: with grounded concepts of space, time, and movement rather than just linguistic probabilities. [07:38] Nina Park: It is a massive philosophical shift in how we approach machine intelligence highlighting the tension [07:44] Nina Park: between scaling what works well now and inventing what might work better later while that research [07:50] Nina Park: continues product focused companies are moving quickly to make current models [07:55] Nina Park: more active participants in the economy. Visa recently announced a major collaboration with OpenAI [08:02] Nina Park: that moves us [08:02] Nina Park: closer to a world of truly autonomous financial agents capable of complex decision-making. [08:09] Thatcher Collins: This is a significant step for product integration. Visa announced on Wednesday [08:14] Thatcher Collins: that they have embedded their payment network infrastructure directly into the ChatGPT environment. [08:20] Thatcher Collins: This is not just a shopping assistant that provides links to products. [08:24] Thatcher Collins: This integration allows AI agents to independently execute and complete financial transactions [08:30] Thatcher Collins: on behalf of an agent. It has the authority to authorize and pay for items at any merchant [08:35] Thatcher Collins: worldwide that accepts Visa payments without requiring the agent to leave the interface. [08:41] Nina Park: I recall that OpenAI attempted a similar feature previously under the name Instant Checkout, [08:46] Nina Park: which was retired back in March. That system was often prone to errors [08:51] Nina Park: and carried a four percent merchant fee that proved to be a barrier for most retailers. Thatcher, [08:56] Nina Park: how does this new collaboration with Visa address those previous failures? [09:01] Nina Park: What makes this version more viable for broad commercial use and long-term adoption?. [09:06] Thatcher Collins: The difference lies in the underlying financial infrastructure. Visa's Jack Forrestel noted [09:12] Thatcher Collins: that they are utilizing a modified tokenization framework called Visa Intelligent Commerce. [09:17] Thatcher Collins: This allows for significantly better data capture and enhanced fraud monitoring. [09:22] Thatcher Collins: They have also integrated strict guardrails, such as user defined spending limits [09:26] Thatcher Collins: and mandatory approval steps for high value items. Initially, the agent will send [09:30] Thatcher Collins: a push notification for the human to approve the final purchase [09:34] Thatcher Collins: until a high level of trust and accuracy is established through repeated interactions. [09:39] Nina Park: Establishing that trust is the primary challenge. If an autonomous agent purchases the wrong item [09:45] Nina Park: or exceeds a budget, the question of liability becomes very complex. Visa has stated they will apply [09:51] Nina Park: the same dispute and fraud rules as they do for regular transactions, but they are adding [09:56] Nina Park: a verification layer to ensure that the consumer's intent was clearly expressed to the AI. [10:02] Nina Park: We are seeing AI transition from an advisory role [10:05] Nina Park: into an active authorized participant in our daily economy with real fiscal responsibility. [10:11] Thatcher Collins: This transition also changes the requirements for businesses to be discovered by these systems. [10:16] Thatcher Collins: If an AI agent is the one doing the shopping, traditional search engine optimization [10:21] Thatcher Collins: might become obsolete. We are seeing a shift toward what many are calling entity SEO [10:26] Thatcher Collins: or generative engine optimization. These AI systems do not just look for keywords, [10:31] Thatcher Collins: they look for verifiable entities and established reputations [10:35] Thatcher Collins: that they can trust before making a recommendation or purchase on behalf of their users. [10:40] Nina Park: This was a major point of discussion in a recent series of articles on Medium. [10:45] Nina Park: For a model like Claude or ChatGPT to recommend a specific business to an agent, [10:50] Nina Park: it needs to understand exactly what that entity is, what its specific expertise involves, [10:55] Nina Park: and whether that information is consistent across multiple digital sources. [11:00] Nina Park: Technical tools like structured data and schema markup are becoming the new battleground [11:04] Nina Park: for digital visibility. If a model [11:06] Nina Park: cannot parse your business data, your business effectively does not exist to an AI agent. [11:12] Thatcher Collins: That brings us to how these specialized agents are being constructed for more complex fields [11:18] Thatcher Collins: specifically in finance at the recent ODSC-ACI East conference the team from Anthropic provided [11:25] Thatcher Collins: a detailed look at moving beyond simple prompt chaining toward what they call [11:30] Thatcher Collins: agentic loops in high-stakes environments like investment banking or wealth management [11:35] Thatcher Collins: you cannot have a model simply guessing at an earnings report or a financial model. [11:41] Thatcher Collins: It requires a much more disciplined and verifiable approach to data handling. [11:46] Nina Park: The technical requirements for those loops are quite intense, Thacher. [11:50] Nina Park: During the presentation they highlighted the use of the model context protocol to integrate [11:56] Nina Park: very different data sources such as SMP Global or FactSet directly [12:01] Nina Park: into the model's reasoning process. These agents are no longer just summarizing existing text. [12:06] Nina Park: They are performing deep dive research, updating complex spreadsheets, [12:10] Nina Park: and self-verifying their own outputs against predefined success criteria [12:15] Nina Park: before any information is ever presented to a human analyst for final review. [12:20] Thatcher Collins: They also introduced a concept called context compaction. Even with the very large context windows [12:26] Thatcher Collins: available in models like Claude Fable V, the raw output from various tools can quickly overwhelm [12:32] Thatcher Collins: the system's memory. Developers are now teaching agents to maintain their own internal state [12:38] Thatcher Collins: by taking structured notes and summarizing their own progress as they work. [12:42] Thatcher Collins: This allows the agent to maintain continuity and focus on tasks that might take several hours [12:47] Thatcher Collins: or involve thousands of individual data points without losing the core objective. [12:52] Nina Park: It really emphasizes that the next phase of the AI industry is not just about building [12:57] Nina Park: larger models, but about creating more disciplined architectures and ensuring [13:01] Nina Park: better integration with established economic systems like Visa. Whether the path forward is through [13:07] Nina Park: the mythos class of language models or a GEPA style world model, the focus [13:12] Nina Park: is clearly shifting toward meaningful agency and measurable real-world impact [13:17] Nina Park: in the professional sphere. [13:18] Nina Park: We are watching the infrastructure of the next decade being built in real time. [13:24] Thatcher Collins: I agree, Nina. We are seeing the end of this simple chatbot era and the beginning of the era [13:30] Thatcher Collins: of autonomous professional-grade tools. This represents a fundamental change in our relationship [13:37] Thatcher Collins: with software and how work is performed across every sector of the economy. [13:42] Thatcher Collins: Thank you for this conversation today. [13:43] Thatcher Collins: It has been an illuminating look at where the industry is headed. [13:47] Nina Park: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. Neural Newscast [13:56] Nina Park: is AI-assisted human-reviewed. View our AI transparency policy at neuralnewscast.com. [14:03] Announcer: This has been Model Behavior on Neural Newscast. Examining the systems behind the story.