I think this is a bigger threat to Google search than chatbots. OpenAI and Anthropic are now more focused on frontier intelligence, complex agentic work, and enterprise revenue. But Google's biggest AI use by users is ad-supported AI Overviews.

ChatGPT's new notes feature lets you create and use lightweight skills in ChatGPT personal accounts. Just create the SKILL.md as a note, then attach the note to a chat request. It's not as robust as what you get in ChatGPT business accounts (proper skill support), Codex, or Claude, but it's a workable lightweight solution for personal accounts.

This is amazing and mildly terrifying. We went from 0% to 88% on FrontierMath Tier 4 in 1.5 years. Also, the Bitter Lesson strikes again. Google's AI co-mathematician is a multi-agent system built for math. Fable is a general model that beats it.

Fable 5's intelligence comes at a huge cost made clear by Artificial Analysis revealing what it spent to run the model on its full benchmark suite. Its intelligence increase over Opus 4.8 is 3.5 index points, about 5.7%, but cost increased 131%, from $4,309 to $9,940. I'm wondering what would happen if you just gave Opus 4.8 131% more token budget, putting it in verification loops rather than one-shotting outputs. Would it close the gap? How high could you get it? I'm reminded of Noam Brown's recent essay on test-time compute. To best understand model improvement, we really need to hold spend constant. Fable 5 costs about $153 per Intelligence Index point while Opus 4.8 costs about $70 per point. So Fable 5 is more than 2X more expensive per intelligence point than Opus 4.8. It's probably still worth it for some tasks. Maybe let Fable 5 orchestrate and verify, and another model execute.

Fable loves to invent and then use its own jargon. It's fun... but can also make reading its longer outputs confusing.

Raw frontier (at the time) models with search access outperformed the more specialized OpenEvidence and UpToDate for health queries in this Nature study. Many things may be responsible, but it's hard to not conclude that the promise of those websites about paywalled paper access didn't hold up. The odd thing about this is that OpenEvidence and UpToDate almost certainly use AI models from frontier labs. So something they're doing actually makes the models worse for healthcare, at least for the tests run here. What could that be? It seems to be a combination of things. One is overly sensitive guardrails that inhibited some responses. Another may be context pollution by irrelevant or misleading retrieval augmented generation. Another, and this is speculation on my part, is that AI labs, especially Google, already retrieve a lot of top content to inform answers, even if it's only abstracts from PubMed. This may be particularly the case for Google with its Google Scholar index. Since search is important to answers, and Google is likely better at search than OpenEvidence and UpToDate, that could be an important differentiator too.

One thing we've observed with Fable's long written responses is that they can become impenetrable when it uses shorthand, acronyms, and jargon until you're lost in the narrative. Almost like its output is an extension of its thinking self-talk.

Annotations are still broken on Markdown rich text previews in Codex and it makes me sad.

Two, maybe three companies, have a legitimate TAM of all current white collar knowledge work, all yet-to-be-invented intellectual disciplines, and all future scientific and technological breakthroughs, with physical labor via robots as a possible upside. And people are like, "they're not profitable yet" and "it's a race to IPO." This is crazy. People are gambling away their salaries on Polymarket and sports betting sites, and you don't think they'll try to get a piece of this future value, however they can get it? Even if it were winner take all, which it's not, there would be no worse than 33% odds of picking the right winner, yielding spectacular expected value, enough to bet on all of them and still come out ahead if all but one fail.

Oh man, I'm really looking forward to this. Question, @faisalyaqub: will I also be able to connect up my work and personal Codex within iOS so I can control each separately by switching workspaces? x.com/faisalyaqub/status/206…

Umm, Nature… I think you’re going to be disappointed.

Does it feel to anyone else like we’re entering the AI endgame? The Overton Window now includes sophon strategies, partial nationalization, and explicit recursive self-improvement timelines, and we get meaningful model upgrades monthly.

Fable has reinforced for me this fact: Scaffolds are temporary, models outgrow them, stay as close to the model as possible. Seeing it do things with ease for which prior models required dedicated scaffolding, when all it needs is a simple prompt.

Fable is a great model, but to really compare my options I need @ArtificialAnlys, @arena and other independent benchmarkers to add GPT Pro series models. If it's cost-prohibitive, maybe OpenAI can subsidize. Just seems like a glaring omission for evaluating options.

Joy! Codex Appshots are now on Enterprise accounts. I've been using this on my personal Mac and loving it. It didn't come to Enterprise at launch, but turned on today.

I'm glad Anthropic is walking back silent degradation for frontier AI work. But I fear it's in the zeitgeist now and also in future AI training data. Are we sure the mere existence of the system card won't hyperstition future Claude models into other silent degradations?

My bet is that we'll see blended token rates. Top models will cost more, but they'll farm out tasks to cheaper models, and the total cost will go down accordingly. This is similar to agencies and consultancies that offer blended rates combining senior and junior resources.

Sometimes benchmark results are so absurd they call into question the benchmarks themselves. To me, this calls for an investigation into benchmark contamination. I don't know a single person who would turn to Gemini 3.5 Flash over GPT-5.5 or Opus 4.8, never mind Fable 5.