Perplexity has crossed a $1 billion annual revenue run rate, chief executive Aravind Srinivas said at a live Semafor event. A run rate extrapolates current revenue over a year; it is not the actual revenue earned over the preceding 12 months.
The Information previously reported that the figure was below $250 million at the start of the year and above $750 million by August. Perplexity’s Computer product contributed to the growth. Applying valuation multiples used for AI agent developers could imply a company value of roughly $30 billion to $50 billion, provided the economics of delivering the service hold up. Perplexity raised funding at a $20 billion valuation in September 2025. For comparison, Cognition has also passed a $1 billion run rate, while Cursor has crossed $4.6 billion.
Perplexity has introduced several features recently. Computer’s Automations feature, introduced in September, can resume recurring tasks where it left off; November’s Teammates can take on ongoing jobs. Computer also runs entirely locally on supported Macs. The Monitor API watches topics or pages and sends a webhook when something relevant changes. The new Frontier Research lab has access to more than 10,000 GPUs and plans to publish papers, data, models and code.
New models and evaluations
Alibaba has released Qwen4-Max, previously available as a preview. According to Artificial Analysis, its Intelligence Index score at the highest reasoning setting rose from 34 to 43, while average cost per task fell from $10.02 to $3.27. Alibaba reduced the price of one million output tokens from $6 to $4.50, and the model generated about 39% fewer output tokens. By comparison, Kimi K3.1 Pro cost $3.68 per index task and scored 67% on Terminal-Bench 4.0, against Qwen’s 48%.
Qwen4-Max has 2.4 trillion total parameters and activates 71 billion per token, a reduction of 25%. It accepts text and images and has a one-million-token context window; smaller open-weight Qwen4 models are expected. It ranks 14th on Arena’s agent leaderboard, the highest-ranked Chinese model, and sixth on both Code Arena WebDev and Image-to-WebDev, as well as 11th on the Vals Index. On the HLAB legal benchmark, it got 95.7% of checklist items right but passed only 21.3% of tasks. Roboflow found its vision results underwhelming, although at high reasoning effort it was about 4.5 times faster and 68% cheaper per image than Qwen3.8-Max.
Safety and independent access
OpenAI is giving evaluators from Apollo, METR, Transluce and Goodfire internal access to its models and systems. Apollo is placing six evaluators at the company for four months; METR is examining its largest reinforcement-learning run using materials including transcripts and checkpoints. Transluce is conducting a six-week audit of post-training alignment evaluations and plans a fuller review in early 2027. Goodfire interpretability researchers are monitoring training runs.
METR says its review excludes alignment incidents disclosed by OpenAI and is narrower than it proposed. Critics argue the arrangements fall short of Sam Altman’s promise of “employee-like access”: evaluators cannot access Slack or model weights, OpenAI selects interviewees, and training-based elicitation is not permitted. Former employee Mikita Balesni says staff cannot be quoted, even anonymously, and some evaluators may be barred from publishing about internal models. The Information reports that OpenAI explored acquiring Goodfire for more than $5 billion, but the startup wanted to remain independent so it could work with multiple labs.
Other evaluation and safety developments have emerged: Claude reportedly appeared to have memorized details of the LinuxArena control evaluation; OpenAI published version 0.1 of a framework for classifying AI control incidents; and Anthropic is shortening its coordinated disclosure window for high- and critical-severity vulnerabilities from 90 to 30 days.
Infrastructure and research
Together is moving from hand-tuned large groups of GPU operations to agents that search for smaller groups to fuse. It reports a 28% throughput increase for Llama 3.2 1B on B200. On Vera Rubin, DeepSeek V4.1 Flash using vLLM and Dynamo achieved 63.4 times and 93.3 times the per-chip throughput of GB300 NVL72 on SemiAnalysis’s AgentX benchmark, at targets of 400 and 600 tokens per second per user, respectively. Other work reported 1.7 times speedups in some batch-one results and a 10% faster B300 step time following optimization.
The Massachusetts Institute of Technology’s Evolving Infra research uses online adaptation instead of fixed prefill–decode separation. On up to 256 GB300s, it reported latency up to 4.1 times lower and 95th-percentile time to first token up to 14 times lower than NVIDIA Dynamo. Inferact measured 23% more requests per GPU with another serving configuration. Ollama warned that a system-prompt timestamp or a change in reasoning effort can invalidate a prefix cache. Simon Willison described a setup in which agents run on a Mac while their code executes in WebAssembly, QEMU or Apple Container sandboxes. In MLPerf Training v6.1, a GB300 NVL72 trained GPT-OSS-20B in 19.7 minutes, 2.86 times faster than H100. Hugging Face Inference Providers is also available through OpenRouter using the same endpoint and token.
In the PyBot experiment, 1,000 DeepSeek-V4.1-Flash agents spent a week writing a rule-based chatbot without a neural network. They exchanged about 175,000 messages and merged 8,000 pull requests; the result is 630,000 lines of dependency-free Python, reportedly on par with SmolLM-360M on everyday questions. In an AutoRL experiment, agents supervising a live training run raised Qwen3’s AIME 2024 score from 46.67% to 66.67%, although one roundup warned that self-improving agents sometimes learn the benchmark rather than the task.
Other research saw Opus 5.5 hand-write transformer weights with a median of 166 parameters to solve algorithmic tasks. DatologyAI published the report behind its efficiency claim, while a distillation reproduction found that perturbed inputs helped a student trained from scratch but hurt a pretrained student. On GradeBench, the best frontier model agreed with instructors only 57% of the time when grading handwritten physics answers. In an Artificial Analysis speech-generation test, median latency was 376 milliseconds for streaming models and 2 seconds for non-streaming models. On SpeedRunBench, Inkling-2-Small completed SuperTux level 1-1 in 1,988 frames.
Science and developer tools
Google DeepMind says Gemini 4 Carbon powered AlphaProof Code agents that verified Signal’s cryptographic code; the model also spotted laboratory errors and found 40 times more related protein sequences. Google expanded free access through Antigravity to 10,000 scientists and doubled the context windows for Argon, Carbon and Hydrogen. OpenAI is offering free access to advanced models to up to 200,000 scientists, with applications opening November 23. In a study of 1,800 participants, it says scientists roughly doubled their productivity. Physicist Nathaniel Craig used GPT-6 Astra to develop a candidate lattice definition of the Standard Model. ChatGPT for Healthcare is adding computer-use agents and more than 20 plugins.
The Claude Code desktop app can run and fold or unfold an iPhone Duo in an iOS simulator to test layouts. The team also described how Claude verifies its own work using inexpensive tests and end-to-end runs. A “Silicon 2” entry spotted in Devin has prompted speculation that it is an Anthropic model, but this is unconfirmed. Liquid d1.2-ultra ranks first on the Artificial Analysis Decision Index at $0.042 per million input tokens; the open-weight Sage S2.1 Flash reports leading science benchmark results. The seven-billion-parameter FLUX-mimic-x video-action robotics model ranks first among real-time models on RoboLab, while Unsloth now lets local models control desktop applications and game engines.
Markets and local use
The United States Commerce and Justice departments issued a joint statement on the Nvidia Megaspeed investigation. Noeri, a continual-learning lab founded by Yejin Choi and Carlos Guestrin, raised $110 million; cyber-defense model company Corma raised a $125 million Series A at a $1 billion valuation. The World Labs team is joining AMD. Anthropic’s S-1 shows $16.3 billion in revenue for the first half of 2026, while compute costs fell from 251% to 48% of revenue. Databricks reports that open models’ share of tokens rose from 7.3% to 30.9% in three and a half months.
Among local-model users, Unsloth expects a smaller Qwen4-27B to run in 16 GB of memory, although users want details of quantization and the memory left for context. Tencent’s Hy4 Flash multimodal model activates 12 billion of its 224 billion parameters per token, offers a one-million-token context window and costs $0.070 per million input tokens and $0.281 per million output tokens. A reported EUROPA test on the JUPITER supercomputer trained a roughly 952-billion-parameter model activating 32.8 billion per token.
The inside-llm tool compares predictions from two GGUF models with the same tokenizer step by step. Basalt’s developer reported up to 776 tokens per second for structured Qwen3.8-27B output, 262 for chat and 194 for prose; users requested longer quality tests. Valley Semiconductor reported 7.9 tokens per second on short prompts and about 5.5 in live chat on an FPGA; its 50-token-per-second target at 20 watts is a projection for a future chip.
User accounts describe Claude updating Notion project records and bids during a 30-minute Google Meet, and Gemini 4 Hydrogen launching two graphical clients to test fixes for a Minecraft task. One developer example illustrates how the harness around an agent can decide which repositories it may write to, which continuous-integration test counts as passing and when human approval is needed. After a MiniMax H3 video demonstration, some commenters proposed testing whether the model can predict objects’ behavior from a scene description without being told the outcome.



