AI Builders Digest

Reflection AI's Misha Laskin says open models flipped to 70% of gateway tokens

Misha Laskin says token share on gateways like OpenRouter and Vercel flipped in about six months, from roughly 70/30 closed-to-open to 70/30 open-to-closed, and he expects open models to end up where Linux is on servers, with valuable closed companies still alongside. Reflection's first open model, Beam (500B total parameters, 23B active), pretrained on about 6,000 GB300s and then spent more compute on reinforcement learning: a little over 10,000 GB300s for four weeks. He says this helps make Beam 3 to 4x more reasoning-efficient than models in its capability class. He compares buying API tokens to renting the whole stack, predicts that enterprises spending $100M+ a year on closed models will move to owning it by first customizing the harness around an open model, and puts the geopolitics bluntly: "open models are Trojan horses for the infrastructure that they bring with them."

• #open-source
• #models
• #enterprise

Vercel: 58% of traffic is now bots, and 60%+ of deployments are agentic

Guillermo Rauch shared Vercel network stats: over the last 30 days 58.18% of traffic was bot-originated, up from 32% in January 2024, and more than 60% of deployments are now agentic, up from about 3% in January 2026. Agents account for up to 83% of pageviews on Vercel's own documentation sites, and the share rises the more the content is optimized for them. He expects direct human traffic to become a rounding error in coming years, with the web built for and by agents. He adds that agents already buy from Vercel's marketplace through its CLI often enough to shock the team, and Vercel is extending this to domains.

• #agents
• #infrastructure

Users approved 93% of Claude Code prompts, so Anthropic now bets on containment

Anthropic found that per-action approval breaks down because users approved roughly 93% of Claude Code permission prompts, so it now limits what an agent can damage with sandboxes, VMs and egress controls; Claude Code's OS-level sandbox alone cut permission prompts by 84%. Two incidents show why model-level defenses cannot stand alone: a red-team phish got Claude to read ~/.aws/credentials and POST them out in 24 of 25 tries, and a malicious file in Cowork used an attacker's API key to upload workspace files to the attacker's account through the allowlisted api.anthropic.com. Anthropic fixed the second with a proxy inside the VM that only passes the session's own token, and now treats every allowlisted domain as a grant of capabilities rather than a simple destination filter. Across its products, gVisor, seccomp and the hypervisors held up; the parts that broke were the custom components Anthropic built around them.

• #security
• #agents

Aaron Levie: agent swarms will burn 1,000x more tokens than people prompting

Aaron Levie predicts that agents spawning other agents, background agents on every task, and agent swarms will consume 1,000x more tokens than people prompting agents one at a time. He argues that chat, which only works as fast as you can prompt it, will look like a relic within a year or two, because the vast majority of tokens will go to agents doing continuous work in the background. His conclusion is that agent adoption is still early and that far more compute and infrastructure will be needed to sustain it.

• #agents
• #compute

Madhu Guru: open-weight models didn't start the fall in price per intelligence

Madhu Guru argues that open-weight models are a tailwind for the falling price per unit of intelligence, not its cause. He credits three drivers that were already at work: labs distilling their best models into smaller, cheaper ones, infrastructure efficiency, and competition between providers. The pattern shows up at every model company: version x of the mid-size model matches the intelligence of version x-1 of the large model, for less money.

• #pricing
• #open-source

ChatGPT subscriptions now double as Devin subscriptions

Thibault Sottiaux, who works on Codex and ChatGPT at OpenAI, announced that a ChatGPT subscription is now also a Devin subscription. He also flagged that Dots "got a pretty big upgrade."

• #products
• #agents

Anthropic's Managed Agents splits brain from hands, cutting p95 time to first token 90%+

Anthropic rebuilt Claude Managed Agents around three swappable pieces: a durable session log, a stateless harness that runs the loop, and sandboxes that the harness calls like any other tool through a single execute(name, input) call that returns a string. Moving the harness out of the container made containers disposable and kept credentials out of reach of Claude's generated code, since git tokens are wired in at setup and MCP calls go through a vault-backed proxy. Because sessions now provision a container only when they need one, p50 time to first token dropped about 60% and p95 dropped over 90%. The design rests on the view that harness assumptions go stale: the session log stores context outside the context window, so a harness can re-read history instead of making irreversible compaction choices, and the context resets once added for Sonnet 4.5's "context anxiety" became dead weight on Opus 4.5.

• #agents
• #infrastructure

One Opus 5.5 prompt ported Thariq's side project to Claude Managed Agents

Before joining Anthropic, Thariq spent about two weeks building a side project with Opus 4 on the Agent SDK; it needed a constantly running process and did not work well. One prompt to Opus 5.5 ported it to Claude Managed Agents and made it far more reliable. It now runs on his new tab page, and he has started merging PRs from others.

• #agents
• #products

Anthropic traces Claude Code's March-April quality drop to three product changes

Anthropic traced reports that Claude Code had gotten worse to three separate product changes, and says the API and inference layer were not affected. The changes were a March 4 switch of the default reasoning effort from high to medium (reverted April 7), a March 26 change meant to clear old thinking once after an hour idle that instead cleared it on every turn (fixed April 10), and an April 16 system prompt line capping text between tool calls at 25 words, which cost 3% on one eval and was reverted April 20. The thinking bug made Claude forgetful and repetitive and, through cache misses, is what Anthropic believes drained usage limits faster; it got past code review, tests and dogfooding, but Opus 4.7 caught it when Anthropic back-tested its Code Review tool, while Opus 4.6 did not. Anthropic reset usage limits for all subscribers on April 23 and says it will run broad per-model evals on every Claude Code system prompt change, with soak periods and gradual rollouts for changes that could cost intelligence.

• #products
• #evals

Nan Yu: after tab-completing code, we now tab-complete our prompts to agents

Nan Yu, who works on product for Codex at OpenAI, sums up the progression: first we wrote code by hand, then we tab-completed it, then we moved to prompting agents, and now we tab-complete the prompts too. He says he cannot stress enough how much he has loved using the feature.

• #coding
• #products

One prompt to a professional company video, in your own style, now open to all

Matt Turck points to a launch, available to everyone today, that turns a single prompt into a professional-level video for your company or product in your own style. He says this is literally the vision Synthesia has pursued since its early days, arriving "slowly, then all at once."

• #video
• #products

Running a Grok Bot as chief of staff? Give its email its own name

Peter Yang's tip for anyone using a Grok Bot as a chief of staff is to register its email under a different name, not your own. Otherwise lines like "Let me copy in Peter to find a time for us" make little sense when the bot is the one copying you in.

• #agents
• #productivity

Founders, exclude VCs from your paid X promos: it reads as buying eyeballs

Nikunj Kothari asks founders to have their marketing teams exclude VCs when promoting products on X. He gets three DMs a day pitching a "collab" on an "insane opportunity" that turns out to be a paid partnership, sometimes followed up by email. He warns that this shows poor judgment and that word spreads when a startup is buying eyeballs to set up a big raise.

• #startups
• #funding