Anthropic priced its shares at $83, above the $70 to $78 range it showed investors, and they were set to trade on the Nasdaq from Thursday. Underwriters raise a price when buyers outnumber the stock on offer, and a price above the range points that way.
NBC News calls it the largest IPO ever. Epoch AI, working from the company’s own filing, puts one quarter of its compute bill near $10 billion, about half of it paid to SpaceX.
Google has three models in play: Neon, a cheaper tier, Carbon, which one company’s test found painfully slow, and Genie 4, which lets you rewrite a world as you walk through it.
And agents keep taking real jobs. OpenAI says its own finish 46% of research tasks alone, and Meta says its agents do most of its ads-model rollouts. Both are company numbers, so read them that way.
Anthropic Goes Public: An $83 Price and a $10 Billion Bill
The IPO priced above range, and the numbers behind it show how much Anthropic spends to stay ahead. One forecaster now argues the pace could get much faster.
- Anthropic prices its IPO at $83, above range, ahead of first trade. Anthropic sold shares at $83 against a marketed $70 to $78 range, and NBC News calls it the largest IPO ever and puts the company’s value at $1.5 trillion. Trading is due to start at 10 a.m. ET on Nasdaq under ANT.
- Anthropic’s compute bill is growing 4x a year, Epoch AI estimates. Epoch AI puts Anthropic’s third-quarter compute spend at $9.863 billion, with OpenAI’s bill growing at about 3x a year. Roughly $5 billion of Anthropic’s quarter went to SpaceX, per figures from its own S-1 filing.
- Zvi Mowshowitz says a 2027 superintelligence is plausible, so plan for it. In a Substack essay, Mowshowitz reads METR’s report that Anthropic now moves close to twice as fast and puts a Superhuman AI Researcher by 2027 below 50 percent, though not far below. The forecasts are his judgment, not METR’s.
Google’s Models: A Cheaper Tier, a Slow One and a World Builder
Google shipped a mid-priced model, drew a harsh outside test of its top one, and opened a tool that edits a virtual world while you explore it.
- Google launches Gemini 4 Neon, priced at $3 input and $15 output. Google says Neon scores 61 on the Artificial Analysis Intelligence Index, third overall, and costs 50 to 80 percent less per task than top rivals. The post does not name those rivals, so the comparison cannot be reproduced.
- Kernel test: Gemini 4 Carbon took nearly an hour on bugs rivals fixed in minutes. In Kernel’s private benchmark, Carbon needed roughly 7 times as many steps as GPT-6 Astra, and about half went to work the bug reports never asked for. It is one company’s closed test, not a verdict.
- Google’s Genie 4 lets you rewrite a world while you walk through it. Google AI Ultra subscribers can now type new prompts mid-session and hear native audio, while Pro members get a waitlist. Google itself lists camera lag and scene consistency as unsolved.
Agents on the Job: Research, Ads, Releases and the Pager
Four companies say agents now carry real work, from lab research to ad ranking to 3 a.m. alerts. Each set of numbers comes from the company selling or running them.
- OpenAI says its agents now finish 46% of two-to-four-hour research tasks alone. OpenAI’s own data shows unaided success on those tasks up from 23% in March, and a median researcher now uses about $2,400 of agent work a day. OpenAI warns researchers may be handing agents different jobs than before.
- Meta says AI agents now run most of its ads-model upgrade work. Meta estimates agents handle over 60% of rolling proven improvements across its ads ranking models, with $45M+ in compute savings projected by the end of 2026. Every number in the post is Meta’s own.
- Google Labs’ Stitch Loop reads your live metrics and opens fix PRs. The agent watches a shipped product around the clock and opens GitHub pull requests with screenshots, but a human approves every merge. Access starts with a waitlist.
- ResolveAI says its SRE-1 model and Primary On-call can take over the pager. ResolveAI says its 27B SRE-1 model lifted one-investigation pass rate from 6.9% to 38.8% on identical incidents, and Primary On-call answers pages itself. Every figure comes from the company or its customers.
Running Agents: Sandboxes, Caches and Switching Costs
Keeping a coding agent safe and affordable comes down to details most users never see: what it may touch, when its memory expires, and when swapping models costs more than it saves.
- OpenAI details how Codex runs commands inside limits users set. Michael Bolin’s post describes permission profiles built on each operating system’s own tools, with the network off in the built-in workspace profile. OpenAI admits API keys in config files may stay readable and names an attack the sandbox does not stop.
- Edgee says a coding agent’s model switch is a ski-rental problem. Edgee’s gateway moves a running session to a cheaper model only once the forgone savings reach 1.5 times the cost of rewriting the cache. It publishes no headline saving, only per-customer estimates from its own replays.
- One developer finds Antigravity’s prompt cache dies after 5 idle minutes. A blogger measured the cache dropping between 4 and 5 of idle time, with each session opening on 13,800 to 16,800 tokens of overhead. It is one person’s tests on one Mac, and nobody has reproduced them.
Today’s Quick Hits
- OpenAI says its models trimmed Jalapeño chip blocks by 10% to 13%. OpenAI estimates model-proposed circuit changes shrank two blocks of its Jalapeño inference chip by 13% and 10% in silicon area, and says AMD saw gains of up to 14%. Every figure is the lab’s own.
- Perplexity launches Context API to feed agents whole web pages. The new endpoint reads up to 10 URLs per call and returns Markdown, links and query-matched passages, billed at $1 per 1,000 URLs that succeed, with developers choosing how fresh each page copy must be.
- Prime Intellect sells companies the training loop frontier labs use. Model Factory puts reinforcement learning, fine-tuning, pre-training and inference on one control plane, with health checks and checkpoint recovery, so a company can train and keep improving its own models, per Prime Intellect.
- Dean Ball names the AI future nobody says aloud: humans stop steering. In a Harvard lecture adapted for his newsletter, Ball argues a quiet consensus expects human owners to hand authority to AI-run firms step by step, and says it can be resisted. It is his view, not OpenAI’s.