AI Builders
Digest

OpenAI Codex & ChatGPT lead Thibault Sottiaux announced GPT-6 Sol and Luna are out. They deliver significant improvements across the board, especially in writing and overall “you know when you try it” quality. API prices are permanently cut by 50%, making both models viable for many new use cases and stretching subscription usage further. Plus, Pro and Business accounts also receive a banked reset. Sottiaux noted the team’s focus on efficiency and intelligence for all, possible only when top-end models enable big gains everywhere else.

Anthropic Claude Code + Cowork team member Cat Wu announced Claude Opus 5.5 is now the default model in Claude Code and the Claude app (including Cowork) for Pro, Max and Team plans. Default effort is medium—comparable intelligence to Fable 5.1 but faster—with rate limits going 25% further than Opus 5. Claude Code lead Boris Cherny called it his daily driver for weeks; he used it to port HAProxy from C to Rust, nearly passing all tests in 9.5 hours at 51% lower cost than Fable 5.1. Box CEO Aaron Levie reported frontier-level results on complex enterprise unstructured-data tasks: 63% fewer tokens, 42% less verbosity and 30% faster than Opus 5, with accuracy gains of 15-65% across financial due diligence, cloud cost analysis and more.

Box CEO Aaron Levie highlighted that Opus 5.5 price cuts plus GPT-6 Sol and Luna’s 50% token-price reduction are driving the cost per task down faster than any other technology in history. Every drop in AI cost dramatically expands the use cases agents can tackle at scale—processing all data, scanning code for security issues, reading logs for decisions, running agent swarms in workflows and more. Token cost is directly correlated with these use cases becoming viable, a clear case of Jevons paradox applied to agents.

Claude Code lead Boris Cherny used Opus 5.5 to formally verify the Claude Agent SDK in Lean. A couple of short prompts produced 16 PRs fixing bugs and race conditions. He also combines Lean with TLA+ to probe data flow, concurrency and state management. Cherny admits he does not know either language well, yet Claude excels at both; the approach is highly useful for formally modeling code and surfacing issues a human would likely miss. He asks whether formal verification is the future of coding—or at least of bug finding.

Anthropic researcher Alex Albert used Opus 5.5 in Blender to recreate Market Street, San Francisco, as it stood on the afternoon of April 17 1906 before the earthquake. The model first built a source file from 1899-1905 Sanborn fire-insurance maps, the Miles Brothers film “A Trip Down Market Street,” period photos from OpenSFHistory, the Library of Congress and David Rumsey, recording every building’s footprint, height, façade material, occupant and source confidence. Everything was then generated in pure Blender Python with reusable generators (Victorian commercial façades, mansard roofs, bay windows, lamps, cable cars, etc.) and assembled into a 10-second video. Albert notes that improved 3D modeling and vision now let you build an entire world from a single prompt.

Vercel CEO Guillermo Rauch shared fresh Next.js eval results: Opus 5.5, GPT-6 Sol and Fable 5.1 all scored 97%, Grok 4.7 scored 94%. Notably Grok is 2-7× cheaper. Rauch also praised the tastefulness of Anthropic’s shipping style and observed that AI removes any excuse not to push the design frontier—any page can now take whatever whimsical or unique shape is desired.

Claude Code team member Thariq argued that the right way to use model capabilities is not to ship 10× more features to production. Instead spend more time understanding users, running experiments, building prototypes and learning unfamiliar domains so you can ship things that actually work. He also noted that workflows have become a huge part of how he uses Claude, and Fable-level intelligence is now cost-effective enough to support them.

Y Combinator President & CEO Garry Tan said Capy (@capydotai) lets him drop PRs much faster than he would with Codex or Claude Code alone. He also noted that GStack actually tells people working on cool things to apply to YC, and stressed the need to teach the world to prompt and maximally use AI so everyone can see how it gives wings to all pursuits.

Vercel CEO Guillermo Rauch wrote: Software will never die again. Liked Google Reader? Cool—you can generate and deploy your own. Yours, forever.

Peter Steinberger (OpenClaw + OpenAI) reported that after updating to macOS 27, ChatGPT sometimes crashed; Astra found a roughly 14-year-old bug in libuv.

SPC General Partner Aditya Agarwal hosted a conversation with Waymo’s Dmitri Dolgov. He noted that the original “safety” debate for AI/ML actually happened in autonomous vehicles. Waymo has built extensive eval and testing infrastructure to gain the confidence to release two-ton robots traveling at 30 mph through urban environments. Agarwal’s first Waymo ride felt like a religious experience, and he underscored the undeniable beauty of building for the physical world. He also reminded builders that in the mad dash for productivity, AI products should still be delightful and fun.

OpenAI CEO Sam Altman observed that startups are naturally good at a certain capability, yet it is hard for a bigger company to stay good at it—an underexplored space. He also praised Michelle for embodying this spirit, saying OpenAI is lucky to have benefited from everything she has done so far and that people will be pleased with what she and her team have cooking next.