Welcome to AI Decoded, Fast Company‘s weekly newsletter that breaks down the most important news in the world of AI. I’m Mark Sullivan, a senior writer at Fast Company, covering emerging tech, AI, and tech policy.
Sign up to receive this newsletter every week via email here. And if you have comments on this issue and/or ideas for future ones, drop me a line at sullivan@fastcompany.com, and follow me on X @thesullivan.
Perplexity’s valuation is soaring. Its traffic tells a murkier story
In a feature about the AI answer engine Perplexity in 2024, I suggested that the startup wasn’t likely to last for the long haul against better-financed players like OpenAI and Anthropic, and that it was most likely an acquisition target. But here we are, well into 2026, and Perplexity is still alive and showing potential.
Nvidia is reportedly in talks to invest in Perplexity in an equity round that would value the AI startup at more than $30 billion. That would mark a 50% increase from its estimated $20 billion valuation just a year ago. The Information reported that the company’s annualized revenue is now more than $750 million, up from under $250 million at the start of 2026. At $750 million in annualized revenue, a $30 billion valuation is roughly 40 times sales.
In February, Perplexity launched its Computer product, which deploys AI agents to do work on behalf of users. It also switched to a usage-based pricing plan in which users get a certain amount of credits for agentic work every month. The Financial Times reported that Perplexity’s revenue increased 50% in a month after the launch of Computer. By March, Perplexity’s annual recurring revenue had risen to $450 million, the report said, with some of that revenue coming from “tens of thousands” of enterprise customers.
The picture of how many people are actually going to Perplexity for answers is somewhat muddled. Numbers from Similarweb suggest that traffic to the Perplexity.ai website peaked at 219 million visits per month in October 2025 and has been retreating ever since. App usage tells a different story. Similarweb numbers show that monthly active users of the Perplexity app on iOS and Android grew from 26.2 million in August 2025 to 37.9 million by February 2026, a roughly 45% increase in six months.
Confidence seems to be high. In early June, Perplexity’s CEO, Aravind Srinivas, told CNBC that his company plans to go public in 2028, regardless of how well the much-anticipated Anthropic and OpenAI offerings are received. “Agnostic of these two companies, we were planning for something in 2028 so that still remains the case,” Srinivas said.
The AI space is awash with venture capital. Bloomberg reported last month that private capital pushed AI deal volumes to record highs in the first half of 2026, and that there’s no slowdown in sight. Private investors appear comfortable with Perplexity’s 40x valuation-to-sales multiple. Whether public market investors will be is a big question mark.
Meanwhile, some observers smell a whiff of bullshit in the air. The financing for many high-profile AI startups is circular in nature, meaning the money is coming from investors in the same ecosystem whose fortunes depend in some way on the success of the startup. Nvidia, for example, is essentially providing Perplexity with the money it needs to buy access to more Nvidia chips. (Nvidia has a similar deal with OpenAI and others.) The supplier underwrites the customer so the customer can buy more supply. But if the customer fails to generate enough profit from that supply, both parties suffer, and the whole scheme may collapse.
OpenAI says Astra is the first model that can hack on its own
OpenAI said Tuesday that its upcoming Astra model can break into computer systems on its own, without a person guiding it step by step. It is the first time OpenAI has said that about one of its models, confirming a classification the company said it could not rule out three weeks ago.
During evaluations, OpenAI says, Astra found two previously unknown security flaws and chained them together to gain access to a software platform. The company is notifying the software’s maintainers so the flaws can be patched. Axios reported that the VP of research, Amelia Glaese, said Astra can find unknown vulnerabilities and figure out how to exploit them across many well-defended systems without human guidance at each step.
Anthropic’s new model remapped Venus and designed proteins
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on Tuesday. Their benchmark scores show significant gains, but the bigger story is how much better the models are at science.
Fable 5.1, for example, trained a neural network on 30-year-old radar data to create a new elevation map of Venus with detail down to 2 or 3 kilometers (1.2-1.9 miles). Mythos 5.1 designed proteins that bind to target molecules, a basic step in developing many modern drugs. On three targets, its designs bound 10 times more tightly than the best human entries in competitions run by the biotech company Adaptyv Bio. Across 12 targets, about half of its designs worked, compared with a typical success rate of 10% to 15% in the field.
Google is shipping Flash models faster than most teams ship bug fixes
Google is preparing to release Gemini 3.8 Flash, known internally as Skimaki, as soon as Wednesday, The Wall Street Journal reported. The new model is said to be much better at generating computer code than past Gemini models. Google has been gunning hard to improve its models’ coding capabilities—not only because it’s fallen behind Anthropic’s Claude tool and OpenAI’s Codex, but because advanced coding tools are crucial for researching and building progressively better frontier AI models.
ChatGPT is now inside Epic’s medical records
OpenAI has integrated ChatGPT Health with Epic’s electronic health record system, which holds data for more than 325 million patients, TechCrunch reported. Clinicians can now pull appointment notes, lab results, medication lists, and specialist documentation into ChatGPT and ask questions across a patient’s medical history. In some deployments, they can use ChatGPT without leaving the patient’s chart.
When OpenAI launched ChatGPT Health in January, it could access patient records only through a third-party integration built by B.well. The Epic integration is read only, meaning ChatGPT can access information from the record but cannot write anything back to it.
New York City is banning AI for 600,000 students
New York City public schools will bar students starting with 2-K, a program for 2-year-olds, up through eighth grade from using generative artificial intelligence for the 2026-27 school year. The rules cover chatbots and AI tutors and bar schools from issuing individual devices before third grade. For third through fifth grades, individual screen time is capped at 30 minutes, and for middle school, at 45 minutes. High school students get restricted access, with possible use for literacy and career readiness and five pilot programs. ABC News reported that the district is also eliminating student-facing AI software and banning companion chatbots.
More AI coverage from Fast Company:
- We finally know more about OpenAI’s rogue-agent incident. It’s worse than we thought
- How the Kansas Department of Labor has implemented AI to make it easier to file for unemployment
- AI wants to become your family’s chief of staff. Should you let it?
- Shopify is giving its engineers free rein on AI. Here’s why
Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.