Anthropic ships a new Sonnet today, and Vercel's data from trillions of real developer tokens tells you exactly why that matters. The model race is real, but the usage data is starting to separate the story from the marketing.
01
Anthropic's Claude Sonnet 5 is out, and the pitch is agents first
Anthropic released Claude Sonnet 5 today, describing it as their "most agentic Sonnet yet" with improved coding and professional work capabilities. The framing is deliberate: this isn't just a smarter chat model, it's built to run tasks autonomously with less hand-holding. ---
Why it matters: Yesterday we covered Aaron Levie's argument that the enterprise AI battle is really about context, not models. Sonnet 5 is Anthropic's answer to the other half of that equation: you need a model that can actually execute inside complex workflows, not just answer questions. If you're evaluating AI coding tools for your team right now, this one is worth a fresh look.
Vercel's token spend data shows Anthropic winning with developers, and open-weight models rising fast
Vercel CEO Guillermo Rauch posted an animated visualization of AI token spend across the Vercel AI Gateway, which routes traffic from millions of developers processing trillions of tokens per month. The chart shows Anthropic holding a clear lead among developers building on Vercel's platform, with open-weight models (think Meta's Llama family, run on your own infrastructure) gaining ground steadily. ---
Why it matters: This is real usage data, not a benchmark. The rise of open-weight models in this chart is the signal worth watching: developers are routing meaningful production traffic to models they self-host, which means the "just use the API" model isn't inevitable. Every lab that prices aggressively to stop that bleed is reacting to exactly this trend.
An OpenAI Codex researcher is asking developers where the product still falls short
Thibault Sottiaux, who works on Codex at OpenAI, posted a direct question to developers: what's something surprising that Codex still can't do well, that should have been fixed by now? The post drew nearly 2,000 replies, suggesting the frustration backlog is real and well-documented. ---
Why it matters: When a product researcher asks this question publicly, it usually means two things: internal feedback loops aren't capturing the right signal, and something is about to change. If you use Codex and have a genuine complaint, the replies are worth skimming. More usefully, the thread is a live catalog of where AI coding tools still break down in practice, which is useful regardless of which tool you use.
**YC president Garry Tan on SF housing: build supply, stop subsidizing demand** — Not AI, but relevant context for anyone trying to hire engineers in San Francisco, where the housing crisis continues to shape who can afford to live near the labs building all of this. **Matt Turck on France's World Cup win** — Nothing here for the digest. Flagging for completeness: no AI content in this post.