Anthropic Accidentally Revealed Their Most Powerful Model Ever
Episode
27 min
Read time
2 min
Topics
Fundraising & VC, Artificial Intelligence, Software Development
AI-Generated Summary
Key Takeaways
- ✓Vertical Model Performance: Intercom's Apex model, built on domain-specific post-training using billions of customer service interactions, achieves a 2.8% higher resolution rate, 65% fewer hallucinations, and lower cost than GPT-4.5 and Claude Opus 4.5. Companies with dense, labeled interaction data may hold untapped fine-tuning assets that outperform general-purpose frontier models in narrow tasks.
- ✓Post-Training as the New Moat: Cursor's Composer Two, built on open-weight Kimi K2.5 with reinforcement learning applied using proprietary coding interaction data, matched GPT-4.5 and beat Opus 4.6 on coding benchmarks at lower cost. This suggests that 75% of model performance gains can come from post-training rather than pretraining compute alone.
- ✓The Bitter Lesson Reframed: Computer scientist Rich Sutton's 1999 essay argues brute-force compute beats human-encoded knowledge every time. However, Sutton himself later clarified that systems trained on real-world experience, not human expert knowledge, represent the next phase, which is precisely what Apex and Composer Two demonstrate through interaction-derived training data.
- ✓Full-Stack AI as Competitive Necessity: Intercom's CPO argues that durable differentiation in AI products will migrate down the stack from application layer to model layer as app-layer features become easier to clone. Companies with sufficient labeled interaction data should evaluate whether proprietary post-training pipelines can reduce API dependency and improve task-specific performance simultaneously.
- ✓Claude Mythos Leak Details: An unsecured Anthropic database exposed a draft blog post describing Claude Mythos as a new tier above Opus, with dramatically higher scores in coding, academic reasoning, and cybersecurity benchmarks. Anthropic confirmed the model exists, flagged cybersecurity risks requiring extra caution, and noted it is computationally expensive, with no general release timeline announced.
What It Covers
Anthropic's accidental leak reveals Claude Mythos, a model surpassing their Opus tier, while Intercom and Cursor demonstrate that domain-specific post-training on proprietary interaction data can outperform frontier models, signaling a structural shift toward vertical AI specialization across enterprise software.
Key Questions Answered
- •Vertical Model Performance: Intercom's Apex model, built on domain-specific post-training using billions of customer service interactions, achieves a 2.8% higher resolution rate, 65% fewer hallucinations, and lower cost than GPT-4.5 and Claude Opus 4.5. Companies with dense, labeled interaction data may hold untapped fine-tuning assets that outperform general-purpose frontier models in narrow tasks.
- •Post-Training as the New Moat: Cursor's Composer Two, built on open-weight Kimi K2.5 with reinforcement learning applied using proprietary coding interaction data, matched GPT-4.5 and beat Opus 4.6 on coding benchmarks at lower cost. This suggests that 75% of model performance gains can come from post-training rather than pretraining compute alone.
- •The Bitter Lesson Reframed: Computer scientist Rich Sutton's 1999 essay argues brute-force compute beats human-encoded knowledge every time. However, Sutton himself later clarified that systems trained on real-world experience, not human expert knowledge, represent the next phase, which is precisely what Apex and Composer Two demonstrate through interaction-derived training data.
- •Full-Stack AI as Competitive Necessity: Intercom's CPO argues that durable differentiation in AI products will migrate down the stack from application layer to model layer as app-layer features become easier to clone. Companies with sufficient labeled interaction data should evaluate whether proprietary post-training pipelines can reduce API dependency and improve task-specific performance simultaneously.
- •Claude Mythos Leak Details: An unsecured Anthropic database exposed a draft blog post describing Claude Mythos as a new tier above Opus, with dramatically higher scores in coding, academic reasoning, and cybersecurity benchmarks. Anthropic confirmed the model exists, flagged cybersecurity risks requiring extra caution, and noted it is computationally expensive, with no general release timeline announced.
Notable Moment
Decagon revealed that over 80% of its model traffic now runs on internally trained models structured as a network of specialized components, each handling a distinct interaction layer, detection, orchestration, response generation, and evaluation, optimized independently rather than relying on a single frontier model API.
Episode Transcript
Today on the AI Daily Brief, are we entering the era of vertical AI models? Before that in the headlines, a big leak with Anthropic confirming the existence of Claude Mythos, what they call by far the most powerful AI model we've ever developed. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright, friends. Quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Blitsy, Assembly, and Robots and Pencils. To get an ad free version of the show, go to patreon.com/aidailybrief. And if you are interested in sponsoring the show, send us a note at sponsors@aidailybrief.ai. Late breaking one last night, a data leak revealed that Anthropic is testing a new model referred to as Claude Mythos. Anthropic has confirmed the existence of this model with a spokesperson saying that it was a step change, their words, in performance and, quote, the most capable we've built to date. They said the model is currently being trialed by early access customers. So here's what happened. On Thursday evening, a draft blog post describing the model was left in an unsecured publicly searchable database. The blog post says, we've finished training a new AI model, Claude Mythos. It's by far the most powerful AI model we've ever developed. Mythos, they write, is a new name for a new tier of model, larger and more intelligent than our Opus models, which were until now our most powerful. We chose the name to evoke the deep connective tissue that links together knowledge and ideas. Compared to our previous best model, Claude Opus 4.6, Mythos gets dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity among others. In preparing to release Cloud Mythos, however, they say, we want to act with extra caution and understand the risks it poses, even beyond what we learn in our own testing. In particular, we wanna understand the model's potential near term risks in the realm of cybersecurity and share the results to help cyber defenders prepare. Mythos is also a large compute intensive model. It's very expensive for us to serve and will be very expensive for our customers to use. We're working to make the model much more efficient before any general release. For those reasons, we're taking a slower, more gradual approach to releasing Mythos than we have with our other models. We're beginning with a small number of early access customers who will explore the model's cybersecurity applications and report back what they find. Now this blog post is very undercooked. It ends not too long after that. Now if you hear the term capybara thrown around, apparently the model was also referred to as that. I'm not sure if capybara was the codename and mythos is the intended launch name. But regardless, this draft blog post was in a cache of unsecured documents. In total, Fortune reports, there appear to be close to 3,000 …
Get the full transcript (5,411 words) + summary by email — free
One-time email with the complete transcript and AI summary of this episode. No account needed.
One email, no spam. We’ll also show you what SignalCast does.
You just read a 3-minute summary of a 24-minute episode.
Get The AI Breakdown summarized like this every Monday — plus up to 2 more podcasts, free.
Pick Your Podcasts — FreeKeep Reading
More from The AI Breakdown
41 Stats That Tell the Story of AI Right Now
Aug 8 · 22 min
Cognitive Revolution
AI:AM #3: Zvi on Fable, the Cases For & Against the Ban, + AI for Math, Logistics & More
Jun 21
More from The AI Breakdown
The Right Way to Worry About AI
Aug 7 · 28 min
Hard Fork
A.I. Safety Is So Back + Mythos Mayhem with Nikesh Arora + Hot Mess Express
May 15
Books, tools, and gear mentioned in this episode
SignalCast may earn commission on purchases via these links. As an Amazon Associate, SignalCast earns from qualifying purchases.
Books
- The Bitter LessonBy guest
by Rich Sutton
“Computer scientist Rich Sutton's 1999 essay argues brute-force compute beats human-encoded knowledge every time.”
Tools
- Claude OpusBy guest
by Anthropic
“Intercom's Apex model...achieves a 2.8% higher resolution rate, 65% fewer hallucinations, and lower cost than GPT-4.5 and Claude Opus 4.5.”
“Cursor's Composer Two, built on open-weight Kimi K2.5 with reinforcement learning applied using proprietary coding interaction data.”
- Claude MythosBy guest
by Anthropic
“Anthropic's accidental leak reveals Claude Mythos, a model surpassing their Opus tier...An unsecured Anthropic database exposed a draft blog post describing Claude Mythos as a new tier above Opus.”
- Intercom ApexBy guest
by Intercom
“Intercom's Apex model, built on domain-specific post-training using billions of customer service interactions, achieves a 2.8% higher resolution rate, 65% fewer hallucinations, and lower cost than GPT-4.5 and Claude Opus 4.5.”
- Cursor Composer TwoBy guest
by Cursor
“Cursor's Composer Two, built on open-weight Kimi K2.5 with reinforcement learning applied using proprietary coding interaction data, matched GPT-4.5 and beat Opus 4.6 on coding benchmarks at lower cost.”
- GPT-4.5By guest
by OpenAI
“Intercom's Apex model...achieves a 2.8% higher resolution rate, 65% fewer hallucinations, and lower cost than GPT-4.5 and Claude Opus 4.5.”
More from The AI Breakdown
We summarize every new episode. Want them in your inbox?
41 Stats That Tell the Story of AI Right Now
The Right Way to Worry About AI
Google’s AI Leadership Shakeup: Disaster or Exactly What It Needs?
Why the Data Center Fight Has Little to Do With AI
Why AI Washing Won’t Work Much Longer
Similar Episodes
Related episodes from other podcasts
Cognitive Revolution
Jun 21
AI:AM #3: Zvi on Fable, the Cases For & Against the Ban, + AI for Math, Logistics & More
Hard Fork
May 15
A.I. Safety Is So Back + Mythos Mayhem with Nikesh Arora + Hot Mess Express
Deep Questions with Cal Newport
Jun 17
Was the Mythos Ban Justified? (Good Idea. Bad Execution.) | AI Reality Check
How I AI
Jun 9
Claude Fable 5 review: what the new Mythos model gets right (and very wrong)
Accidental Tech Podcast
Apr 30
689: The Positive Effect of Enthusiasm
Explore Related Topics
This podcast is featured in Best AI Podcasts (2026) — ranked and reviewed with AI summaries.
Read this week's AI & Machine Learning Podcast Insights — cross-podcast analysis updated weekly.
You're clearly into The AI Breakdown.
Every Monday, we deliver AI summaries of the latest episodes from The AI Breakdown and 192+ other podcasts. Free for one show.
Start My Monday DigestNo credit card · Unsubscribe anytime