Latest / Neural Newscast / Anthropic Fable 5 Signals Shift to Speed Bottleneck [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'm Nina Park. Welcome to Model Behavior. This episode examines AI systems [00:17] Nina Park: in professional environments. Today is June 12th, 2026. [00:22] Nina Park: Earlier this week, on June 9th, Anthropic released Claude Fable [00:27] Announcer: V, a model they are characterizing as mythos class. While it has immediately claimed the top spot [00:33] Announcer: on major leaderboards, [00:35] Announcer: the industry is reacting more to its speed or the lack thereof than its raw intelligence. [00:41] Thatcher Collins: I'm Thatcher Collins. It is a significant moment for the industry, Nina, but also a polarizing one. [00:47] Thatcher Collins: Claude Fable V is currently ranked number one on artificial analysis and the LM Arena text [00:53] Thatcher Collins: and agent arenas. However, it also ranks 139th out of 152 models for price [00:58] Thatcher Collins: and 64th for output speed. We are seeing a model [01:02] Thatcher Collins: that is effectively a beast, powerful but undeniably slow and expensive to operate at scale. [01:08] Announcer: The performance metrics are striking, Thatcher. Fable V delivers about 60 tokens per second, [01:15] Announcer: but the time to first token is 108 seconds. For context, the field median is currently [01:21] Announcer: under three seconds. Anthropic seems to be betting that users will trade time for capability. [01:27] Announcer: This model was originally a project called Claude Mythos, which was reportedly so powerful [01:33] Announcer: in its cybersecurity capabilities [01:35] Announcer: that it required a dense safety architecture before a public release. [01:40] Thatcher Collins: That safety layer is why we see this fallback mechanism. If Fable 5 filters a request, [01:45] Thatcher Collins: it automatically routes the task to Claude Opus 4.8. But the broader story, Nina, [01:51] Thatcher Collins: is that the bottleneck has moved. For years, we asked if a model was smart enough to do the work. [01:55] Thatcher Collins: Now that models can hold hours-long agentic sessions, [01:58] Thatcher Collins: the question is how long we have to wait for the result. [02:01] Announcer: We are already seeing what that capability looks like in practice. Within 24 hours of launch, [02:08] Announcer: the payment platform Stripe used Fable V to migrate a Ruby codebase consisting of 50 million lines [02:15] Announcer: of code. Normally, that is a two-month project for a human engineering team. Fable V finished it in [02:21] Announcer: a single day. It is the definition of long-horizon autonomy, [02:26] Announcer: where the AI is essentially a digital laborer you can set and forget. [02:31] Thatcher Collins: It is not just enterprise migrations, Nina. We are seeing Vibe coding reach a new level. [02:38] Thatcher Collins: Developers have used Fable 5 to clone Minecraft in 20 minutes and recreate a full version of [02:45] Thatcher Collins: Pokémon Gen I in an hour. Even more impressive is its visual navigation. [02:51] Thatcher Collins: It played through Pokémon Fire Red using only screenshots for input. It is not just reading text, [02:58] Thatcher Collins: it is understanding and navigating complex visual interfaces autonomously. [03:04] Announcer: This shift toward high autonomy agents is driving a massive reallocation of capital toward hardware. [03:10] Announcer: If the model is smart enough but too slow, the only way to win is through silicon. [03:16] Announcer: Nvidia recently paid a reported $20 billion to license GROC's low-latency inference technology. [03:23] Announcer: Meanwhile, OpenAI signed a $10 billion deal for Cerebrus waferscale systems. [03:28] Announcer: The goal for both is the same, providing dedicated low latency inference for these agents. [03:34] Thatcher Collins: The talent market is following that money meana. Anthropic recently poached one [03:38] Thatcher Collins: of the original chip engineers from OpenAI The job description was not just about architecture, [03:44] Thatcher Collins: it was described as "perplexity per picojoule." When AI starts building AI, [03:50] Thatcher Collins: which Anthropic says Claude is now doing for 80% of their internal code, [03:54] Thatcher Collins: the lab that iterates 20 times faster [03:57] Thatcher Collins: because of better inference will eventually pull ahead in intelligence too. [04:01] Announcer: That is the recursive loop. Beyond Fable V, Anthropic also updated the Claude IV family. [04:07] Announcer: We have OPUS 4.8 for complex reasoning, SONNET 4.6 balancing speed and intelligence, [04:14] Announcer: and HAICU 4.5 for high volume tasks. They're also implementing fast modes and priority pricing. [04:21] Announcer: It is a reversal of traditional software tiers. [04:24] Announcer: You're not paying for more features, you're paying a premium for the same model to respond sooner. [04:29] Thatcher Collins: It suggests [04:30] Announcer: that speed [04:31] Thatcher Collins: may be the first real moat in AI nina. Benchmarks can be distilled or replicated elsewhere [04:37] Thatcher Collins: in months, but tokens per second are manufactured. If you hold the speed mode, [04:43] Thatcher Collins: you remint your intelligence lead every time your agents finish a research cycle faster [04:48] Thatcher Collins: than the competition. As we move into the second half of 2026, [04:52] Thatcher Collins: tokens per second is becoming the only headline number that matters for market dominance. [04:57] Announcer: Thank you for listening to Model Behavior, a neural newscast editorial segment. [05:02] Announcer: Visit mbnneuralnewscast.com. Neural newscast is AI-assisted, [05:08] Announcer: human-reviewed. View our AI transparency policy at neuralnewscast.com. [05:13] Nina Park: This has been Model Behavior on Neural Newscast. Examining the systems behind the story.