AivexaNewsSearch
AI news for builders and product teamsChecked every hour

[AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026

Collected Nov 17, 2026

Perplexity crossed a $1B annual revenue run rate, according to CEO Aravind Srinivas, who disclosed the figure to Semafor at a live event. The trajectory is steep: The Information reported the company was under $250M at the start of the year and above $750M by August, with its Computer product credited with driving part of the growth. The company is framing the milestone as a roughly fourfold jump within the year.

The revenue base matters more than the headline number for anyone tracking valuations in the agent space. Using the multiples recently applied to comparable agent companies, the report estimates Perplexity would now be worth roughly $30B to $50B, contingent on unit economics holding up. That would be a meaningful step up from the $20B round the company raised in September 2025. Treat the valuation range as an external estimate built on peer comparables rather than a priced round; private valuations are set by what investors actually pay, so the gap between the implied figure and the last round is a signal of momentum, not a confirmed mark.

Several recent product launches sit behind the growth story. Computer, the company's agentic product, gained Automations in September that let it resume recurring tasks from where it previously stopped, and Teammates in November that take on ongoing jobs. Computer can also run fully locally on supported Macs. Alongside that, a Monitor API lets developers watch a topic or page and receive a webhook when something relevant changes, which is a useful primitive for anyone wiring agents into monitoring or alerting pipelines. The company has also set up Frontier Research, a new lab with access to more than 10,000 GPUs that plans to publish papers, data, models and code.

The broader model landscape in the same period was active. Alibaba moved its Qwen4-Max preview out of preview into a shipping model. At the highest reasoning setting it scored 43 on the Intelligence Index, up from 34, while average cost per task fell from $10.02 to $3.27. Output pricing dropped from $6 to $4.50 per million tokens, and the model produced roughly 39% fewer output tokens, so the cost decline compounds: fewer tokens at a lower rate. The total parameter count stayed at 2.4T with active parameters per token reduced by 25% to 71B, the model accepts images and text with a 1M-token context window, and smaller open-weight Qwen4 variants are still to come. Independent comparisons are mixed: Kimi K3.1 Pro scored 67% on Terminal-Bench 4.0 against Qwen's 48%, and Roboflow called the vision results underwhelming while noting it was roughly 4.5x faster and 68% cheaper per image than Qwen3.8-Max at high reasoning effort. On a legal benchmark it got 95.7% of checklist items right but passed only 21.3% of tasks, a reminder that partial credit and full task completion diverge sharply.

On safety and evaluation, OpenAI gave Apollo, METR, Transluce and Goodfire internal access to its models and systems. The scopes differ: Apollo places six evaluators onsite for four months; METR reviews OpenAI's largest RL run with access to transcripts, monitor records and checkpoint snapshots; Transluce runs a six-week audit of post-training alignment evals with a fuller follow-up in early 2027; Goodfire embeds interpretability researchers to monitor training runs. METR says its scope excludes the misalignment incidents OpenAI has disclosed and is narrower than what METR proposed, and critics say the arrangements fall short of a stated promise of employee-like access. Reported gaps include no Slack access, OpenAI selecting which staff are interviewed, and no access to weights or training-based elicitation.

Inference infrastructure saw a wave of results. Together reported 28% higher throughput on Llama 3.2 1B on B200 after moving from hand-tuned megakernels to agent-searched fusion of smaller groups of GPU operations. DeepSeek V4.1 Flash on Vera Rubin, served through vLLM with Dynamo, showed 63.4x per-chip throughput over GB300 NVL72 at 400 tokens/s/user and 93.3x at 600 on SemiAnalysis's AgentX benchmark. MIT's Evolving Infra replaces fixed prefill/decode disaggregation with online adaptation and reports up to 4.1x lower latency and 14x lower P95 time-to-first-token against NVIDIA Dynamo on up to 256 GB300s. In MLPerf Training v6.1, a GB300 NVL72 trained GPT-OSS-20B in 19.7 minutes, 2.86x faster than H100. Ollama also flagged a practical trap: a single system-prompt timestamp or a change in reasoning effort can wipe the prefix cache.

On the science side, Google said Gemini 4 Carbon powered AlphaProof Code agents that verified Signal's cryptographic code, spotted wet-lab errors and found 40x more related protein sequences, and it expanded free Carbon access to 10,000 scientists through Antigravity, where the context window was doubled for Argon, Carbon and Hydrogen, cutting compaction 3.5x per user turn. OpenAI is opening free frontier access to up to 200,000 scientists with applications from Nov 23.

Why it matters: Perplexity's number is a data point on whether agentic consumer products can convert usage into durable revenue at scale, which affects how investors price the whole category. For developers, the launch mix points to where the competitive surface is moving: webhook-driven monitoring, locally running agents and recurring-task automation rather than pure question answering. Inference cost curves reported across Qwen4-Max, Together and the Vera Rubin results suggest per-task economics keep improving, which is itself an inference from vendor-reported benchmarks rather than an independent finding.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

November 16, 2026 — Computer gets Teammates, Perplexity ships Monitor API, and a new research lab launches.