Dylan Patel, founder of SemiAnalysis, wrote on X on 11 November that his firm’s AI spend is now 45% persistent agents, and that it primarily uses Perplexity Computer. He wrote that Anthropic’s “Claude Tags sucks” and that “OpenAI’s Slack agent is even worse”. This is one company describing its own spending and its own view of the products.
A SemiAnalysis chart attached to the post says persistent agents, Perplexity Computer plus Claude Tags, went from 0% to 45% of the firm’s AI spend between 1 February and the week to 8 November 2026. In a reply, Patel wrote that “Perplexity is primarily using Anthropic models for our use cases”, so the spend moved to a different product while the main model, by his account, stayed Anthropic’s.
Business Pill · AI Agents
An evergreen Business Engineer explainer on what AI agents are and how they move from answering to doing. It teaches the general idea only and does not discuss SemiAnalysis or the products named here.
The key insight: As we read it, SemiAnalysis’s numbers separate two things buyers often treat as one, the model and the product around it: by Patel’s account, his firm’s agent spend moved mainly to Perplexity Computer while Perplexity primarily uses Anthropic models for its use cases, and he credits Perplexity’s harness for the choice.
The first post, at 20:40 UTC, reads in full: “Persistent Agents are the next unlock of AI usage”, “Our spend on AI is now 45% persistent agents”, “We primarily use Perplexity Computer, which uses a lot of different models” and “Perplexity’s Harness mogs”, alongside the two lines on Anthropic and OpenAI.
The attached chart’s footer gives its method: “Source: SemiAnalysis internal AI usage dashboard (Afzal Ahmad). Provider split before Sep 9 read off the dashboard chart; Sep 9 to Nov 8 and all Claude Tags data from daily usage-report CSVs.” It plots share of daily AI cost by provider, 7-day trailing, across five series: Perplexity Computer, Claude Tags, xAI, OpenAI and Claude excluding Tags.
Two minutes later, in the same thread, Patel wrote that “Anthropic sucks at product, and they’re not really receptive to any feedback on their product”, adding “Same with OpenAI”. In another reply he wrote that Perplexity implements “any change that’s needed in less than 24 hours” and “They even got the name from us back in February”, attaching a SemiAnalysis newsletter passage that describes Claude Code “rather as Claude Computer”.
SemiAnalysis’s own AI spend, share of daily cost, 7-day trailing: persistent agents (Perplexity Computer plus Claude Tags) at 0% at the start of tracking on 1 February 2026 and 45% in the week to 8 November 2026, per the chart Dylan Patel attached to his X post on 11 November. One firm’s spend, not a market figure; 55% is our arithmetic.
What Was Posted on X
Persistent Agents are the next unlock of AI usage Our spend on AI is now 45% persistent agents 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
Dylan Patel, founder of SemiAnalysis, posted on 11 November at 20:40 UTC that his firm’s AI spend is now 45% persistent agents and that it primarily uses Perplexity Computer, with a SemiAnalysis chart of its own spend attached. This post is the primary for the piece; the 45% is one firm’s spend, and the product verdicts are his opinion.
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 Same with OpenAI
In a reply at 20:42 UTC, Patel wrote that Perplexity is primarily using Anthropic models for his firm’s use cases, and criticised Anthropic and OpenAI on product. The piece uses it for the model-versus-product distinction, attributed to him.
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. They even got the name from us back in February.
In a second reply at 20:42 UTC, Patel wrote that Perplexity implements any change that’s needed in less than 24 hours and that it got the name from SemiAnalysis in February, attaching a SemiAnalysis newsletter passage that describes Claude Code as Claude Computer. The piece reports this as his account.
Three Agent Products, Three Designs
Perplexity’s launch post describes Computer as “capable of running for hours or even months”, splitting work into sub-agents that each run “in an isolated compute environment with access to a real filesystem, a real browser, and real tool integrations”. It calls the product “a safe harness for powerful AI” and lists Opus 4.6 as its core reasoning engine, with Gemini, Nano Banana, Veo 3.1, Grok and ChatGPT 5.2 for other tasks, as of that writing. It is available to Perplexity Max subscribers, the post says. We covered its model-agnostic orchestration in July.
Claude Tag, Anthropic’s Slack agent, is labelled Public Beta in its documentation and is available on Team and Enterprise plans. The documentation says that when it works on a task, “it runs in an ephemeral sandbox, not on your computer”, one that “is discarded when the conversation goes idle”. It adds no per-seat charge; channel work is billed by usage from a balance an organization funds, under a spend limit.
OpenAI’s workspace agents are “Powered by Codex” and “run in the cloud, so they can keep working even when you’re not”, the company says, and its help article describes the ChatGPT Agents app that deploys them to Slack channels. They are in research preview for ChatGPT Business, Enterprise, Edu and Teachers plans, with “credit-based pricing” from 6 May 2026, OpenAI says. Our earlier piece covered Grok Bot 1.0 in Slack and Teams.
Our Numbers
The 45% share in the SemiAnalysis chart compares with 0% at the start of its tracking on 1 February; the rest of its AI spend, 55% by our arithmetic, went to Claude excluding Tags, OpenAI and xAI. The posts and chart we read give shares only, not dollar amounts.
For scale, Ramp’s November AI Index puts the top 1% of firms by AI spend at a median $10.2K per employee per month, which we covered in our Ramp piece. At that level, a 45% share equals about $4,590 per employee per month, by our arithmetic. This is an illustration only; we do not know SemiAnalysis’s spend per employee.
The Structural Read
As we read the product pages, the three designs differ most on how long the agent lives: Perplexity says Computer can run for hours or even months, Anthropic says Claude Tag’s sandbox is discarded when the conversation goes idle, and OpenAI says its workspace agents keep working in the cloud.
As we read the billing terms, Anthropic and OpenAI both meter agent work: Claude Tag adds no per-seat charge and draws on a usage balance an organization funds, and OpenAI set credit-based pricing for workspace agents from 6 May 2026.
Dylan Patel, founder of SemiAnalysis, on X, 11 November 2026
“Perplexity’s Harness mogs”
Three Implications
FOR ANTHROPIC Patel wrote that Perplexity is primarily using Anthropic models for his firm’s use cases, while the firm’s agent spend goes mainly to Perplexity Computer rather than Claude Tag.
FOR OPENAI Patel wrote that OpenAI’s Slack agent is “even worse” than Claude Tag; OpenAI describes its workspace agents as a research preview.
FOR PERPLEXITY Patel wrote that Perplexity did not invent the category but is a fast follower that implements requested changes in less than 24 hours.
The framework says: “The agentic surface — the harness — is becoming where value pools in the entire AI stack.”
As we read it, Patel’s chart is one buyer’s reading of that claim: the spend moved to the product wrapped around the model while, by his account, the model underneath stayed largely Anthropic’s. The framework predates the posts and does not discuss them.
What Is Not Established
Patel’s posts describe one firm’s spend and his own opinion of three products; they are not market data. The chart’s footer says the provider split before 9 September was read off a dashboard chart, and that days from 9 November are excluded because Perplexity billing lags one to two days.
How much of the 45% is Perplexity Computer and how much is Claude Tags is not stated as a number in the chart, which shows Claude Tags as a thin band. We did not contact SemiAnalysis, Perplexity, Anthropic or OpenAI for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
SemiAnalysis says persistent agents went from 0% to 45% of its AI spend between February and November, mostly through Perplexity Computer, a product that, by Patel’s account, mainly runs Anthropic models for his firm. As we read it, the contest he describes is between the products wrapped around the models, which Perplexity itself calls a harness.
A note on sourcing. We read Dylan Patel’s three X posts of 11 November 2026 in full through the X API and api.fxtwitter.com, and transcribed the text of the SemiAnalysis chart and the newsletter screenshot attached to them. Product facts come from Perplexity’s Computer launch post, Anthropic’s Claude Tag product page, documentation and pricing page, and OpenAI’s workspace agents post and Slack help article, all read on 11 November.
The Ramp figure comes from the November 2026 Ramp AI Index. The 55% share and the $4,590 illustration are our arithmetic. We did not contact SemiAnalysis, Perplexity, Anthropic or OpenAI for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
