SemiAnalysis CEO Dylan Patel says close to half of his firm's AI bill now goes to persistent agents. In an X post on Nov 11, 2026, he wrote: "Persistent Agents are the next unlock of AI usage" and "Our spend on AI is now 45% persistent agents."
The same post names the tools behind that share and criticizes products from Anthropic and OpenAI. Here is what a persistent agent is, what his chart shows, and what he says about why the spend moved.
What is a persistent agent?
A chat assistant works turn by turn: you type, it answers, and it waits for you. A persistent agent is built to keep working after you hand it a task. It can run in the background, come back to a job on a schedule, and report what it did.
Patel's chart puts two products in this group. Here is how each vendor describes its own product.
- Perplexity Computer. Perplexity's product page calls it "a general-purpose digital worker that operates the same interfaces you do." It says Computer can run workflows "for hours or even months," and that you can "set recurring tasks that Computer executes autonomously and persistently over time." The page says it is available to Pro and Max subscribers. A Perplexity help article on Personal Computer, Perplexity's desktop app built around Computer, says Computer "creates teams of agents across 20+ frontier models."
- Claude Tag. Anthropic's launch post from Jun 23, 2026 says Claude Tag joins Slack as a team member. Anyone in a channel can tag @Claude and hand it work. Anthropic says it "can also schedule tasks for itself, pursuing a project autonomously over hours or days." At launch it was in beta for Claude Enterprise and Team customers.
Patel's post writes the Anthropic product as "Claude Tags." Anthropic's own name for it is Claude Tag.
For more background on the idea, see our explainers on what computer use is and what agent skills are.
What the chart shows
Patel attached a chart to his post. Its title reads "Persistent agents (Perplexity Computer + Claude Tags) went from 0% to 45% of our AI spend." The chart plots each provider's share of SemiAnalysis's daily AI cost, as a 7-day trailing average, from Feb 1 to Nov 8, 2026. A callout on the chart says persistent agents were "45% of spend in the week to Nov 8."
The chart has five groups:
- Perplexity Computer
- Claude Tags
- xAI
- OpenAI (API + ChatGPT/Codex)
- Claude (Enterprise + API, excl. Tags)
Two details shape the 45% figure. First, it combines Perplexity Computer and Claude Tag. It is not a Perplexity-only number. Second, the chart's source note says the data comes from the "SemiAnalysis internal AI usage dashboard." The note says the provider split before Sep 9 was read off the dashboard chart, and that days from Nov 9 on are excluded because Perplexity billing lags 1 to 2 days.
What fills the other 55%? When a reader asked that question, Patel replied: "Claude code." Asked in another reply whether Perplexity has usage, Patel wrote: "We spend $7m run rate." That reply does not say whether the figure covers all SemiAnalysis AI spend or one vendor.
What Patel says about the products
Patel's main post gives his view of the tools in his own words:
- "Unfortunately for Anthropic, Claude Tags sucks"
- "OpenAI's Slack agent is even worse"
- "We primarily use Perplexity Computer, which uses a lot of different models"
- "Perplexity's Harness mogs"
In a follow-up post, he wrote: "The funny thing is that Perplexity is primarily using Anthropic models for our use cases, but Anthropic sucks at product, and they're not really receptive to any feedback on their product." He added: "Same with OpenAI."
His stated reason for picking Perplexity is speed of change. In a second follow-up, he wrote: "What I love about Perplexity is that they're very fast on any feature request. Even though they did not invent the category, they are fast followers and implement any change that's needed in less than 24 hours." He also wrote: "They even got the name from us back in February." The screenshot attached to that post shows a SemiAnalysis article, Claude Code is the Inflection Point, dated Feb 5, 2026. That article says Claude Code should be thought of "rather as Claude Computer." The naming claim is Patel's. Perplexity has not commented on it in any source we found.
A reply from Anthropic's side
Noah Zweben, who describes himself as the "Real life human PM for Claude Tag," replied that Patel should "feel free to lmk what's not working or DM me." He added: "We'd love to improve the product and we've shipped a ton of improvements over the last month." Patel answered: "Join the slack will rant at u."
We found no public reply from OpenAI in the thread at the time of writing.
Reading the numbers
The chart shows how one research company split its AI budget in early November 2026. It does not show how other companies spend.
As of the morning of Nov 12, 2026 (PT), X showed the main post with roughly 126,000 views and about 175 replies. Those counts change.
For another recent look at how agents are changing AI use, see our post on a16z's claim that agents are the majority of web traffic.
Sources
- Dylan Patel on X: persistent agents are 45% of SemiAnalysis AI spend (Nov 11, 2026)
- Dylan Patel on X: Perplexity uses Anthropic models; criticism of Anthropic and OpenAI
- Dylan Patel on X: Perplexity's speed on feature requests
- Dylan Patel on X: "Claude code"
- Dylan Patel on X: "We spend $7m run rate"
- Noah Zweben on X: reply from the Claude Tag PM
- Dylan Patel on X: reply to Noah Zweben
- Perplexity: Computer product page
- Perplexity Help Center: What is Personal Computer?
- Anthropic: Introducing Claude Tag (Jun 23, 2026)
- SemiAnalysis: Claude Code is the Inflection Point (Feb 5, 2026)
