Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Perplexity’s revenues have jumped 50 per cent in a month, as the start-up accelerates its shift into AI agents in an effort to keep pace with larger and better-funded rivals.
Estimated annual recurring revenue rose to more than $450mn in March, after the launch of a new agent tool and a shift to usage-based pricing, according to figures seen by the FT.
The move marks a shift of focus away from its chatbot-style search engine, once seen as the biggest challenge to Google in two decades.
Instead, it is pushing AI agents that can carry out tasks on users’ behalf, as companies across the sector experiment with pricing models to reflect the higher costs of running them.
The start-up has more than 100mn monthly active users from its search and agent tools, according to executives, including tens of thousands of enterprise clients. It makes money through consumer and enterprise subscriptions, with tiers ranging from $20 to $200 monthly.
With the launch of its AI agent Computer in February, Perplexity has also added a usage-based pricing model, under which top-tier subscribers receive a set number of credits before paying for additional use.
Unlike subscription revenue, usage pricing can be a more volatile metric for estimating annual growth and less comparable with previous months. Before the new pricing system, the start-up had grown its ARR from $16mn to $305mn over two years, according to the figures seen by the FT.
Even so, Perplexity’s growth trajectory is dwarfed by other leading AI start-ups. Coding company Cursor has grown to $2bn ARR from less than $100mn in 2024. Anthropic said its ARR hit $19bn at the end of February, compared to the $20bn OpenAI generated last year.
The company’s focus on AI agents comes after its efforts to launch an AI-powered search engine proved controversial. It has faced lawsuits alleging copyright infringement and plagiarism by publishers such as The New York Times and Britannica, alleging that it is “illegally copying” content. A recent privacy suit claims it shared user data with Google and Meta without consent. The company has denied wrongdoing on these claims.
Perplexity also put itself forward as a potential acquirer of the US arm of TikTok last year, a proposal that gained little traction as a group of existing investors and others with ties to the Trump administration eventually took control of the social media business.
Last year, Perplexity launched its AI web browser, Comet, one of the first of this kind to market. Comet can act as an agent on users’ behalf, following voice and text commands to perform tasks such as shopping, summarising social media feeds and sending emails. This year, Perplexity launched its agentic product Computer, as well as Model Council, which runs queries through different models and shows outputs side-by-side.
The company offers a range of models from OpenAI’s GPT and Anthropic’s Claude, to open-source offerings from Chinese companies such as DeepSeek’s R1 and Moonshot’s Kimi.
AI groups are now exploring pricing models to pay for heavier workloads driven by more agentic systems.
After Nvidia chief executive Jensen Huang encouraged the audience at its annual conference last month to subscribe to Perplexity’s computer tool and “pay the maximum amount”, Aravind Srinivas, Perplexity’s chief executive, replied: “There is no limit, you can spend as much as you want.”
Perplexity was last valued at $20bn in September, up from $500mn at the beginning of 2024.
Its investors include Nvidia, SoftBank’s Vision Fund 2, venture capitalists New Enterprise Associates and IVP — as well as Amazon founder Jeff Bezos and Meta’s former chief AI scientist Yann LeCun.
A Perplexity executive told the FT that revenue retention was “strong”, without providing a figure.
Lossmaking Perplexity pays model providers like OpenAI and Anthropic to use their models, as well as the inference costs of running queries.
However, one person close to Perplexity said the start-up’s advantage was being able to triage requests to the most efficient model for each purpose. For example, OpenAI’s Codex and Anthropic’s Claude Code for coding; OpenAI’s GPT-5 for writing or Anthropic’s Opus for reasoning, they said.