Aravind Srinivas

Every Perplexity Project is a Git repo. We built a Git forge for agents, without the GitHub bloat, since no human has to approve anything. Each agent gets its own short-lived sandbox and commits to a shared repo on S3. Fewer than 2% of submissions collide. Agents search with plain `ls` and `grep`. That beats semantic search 63% vs 37% on answer accuracy for financial document analysis.

Congrats to Meta for training a frontier-grade model! Watermelon compares favorably to Fable 5.1 on WANDR evals.

RT Perplexity This is how we built the filesystem behind Perplexity Projects as a Git forge for agents. Each agent works in its own sandbox and commits to a shared repository where people and other agents can build on its work across sessions. Read more: https://www.perplexity.ai/hub/blog/building-shared-workspaces-for-people-and-agents

RT sourcery Perplexity CEO Aravind Srinivas (@AravSrinivas) says "curiosity" is humanity’s last edge against AI: “You could argue that AI is the best accelerant of our own curiosity. Almost the purpose of AI is to accelerate human curiosity.” “AIs can answer whatever question there is, so what’s our role? Our role is to go & ask the next set of questions.” “Our brand is built around that & we think that the best companies always have a fundamental brand that has nothing to do with technology & more to do with humanity & people.”

Was a fun chat! Thanks for having me @MollySOShea!

RT Computer Computer can enrich public-domain stories like Shakespeare's plays with illustrations, embedded videos, archival photos, historical context from web sources, and modern translations. Read the enriched story: https://read-macbeth.pplx.app

RT Computer Computer pulled 52,000 earnings call transcripts from 6,000 global companies to create an earnings keyword search app. Computer orchestrated 3 different frontier models to retrieve the transcripts, index passages, and create the website, using 210,000 credits in total.

RT Sanchit monga More cool EVE demos coming every day. 128 questions across a lease and an offer letter, all in a single request per document. EVE running perplexity decider 4x faster than JEV!

RT Computer Computer's new website editor lets you queue up requested changes to specific elements on the page. Once you submit your changes, Computer addresses them in the main thread, and then outputs the new website.

Perplexity Decider ranks first on DecisionBench (and lowest cost too)

perplexity’s open weights decision model the best quality and lowest cost for clinical decisions

Cool demo of Perplexity Decider!

RT Perplexity Developers pplx-decider-v1.1-27b is now available on OpenRouter.

RT OpenRouter Perplexity Decider v1.1 is now on OpenRouter @perplexity_ai's open-weights multimodal decision model takes text, JSON, or images and returns typed answers with probabilities $0.02/M input. Output is free https://openrouter.ai/perplexity/pplx-decider-v1.1-27b

AI is the work operating system

We’re open-sourcing pplx-embed-v2-late, multi-vector embeddings for text and images, 9B and 0.6B, in one shared embedding space. You can use these to index multimodal data with 9B, and query on device with 0.6B. This also enables you to search over PDF pages with no OCR. And scores 92.4% on MADQA, 64% on BrowseComp+. Weights available on @huggingface now.

We’re living in incredible times!

Also, Perplexity Decision API pricing was cut by half, now to 2 cents per million input tokens. Enjoy!

Engineering Design as a service

Perplexity wins on Hugging Face Decision Index (benchmark for decision models). Our model is open-weights and accessible to all!

Perplexity Decider is the best decision model

We got Computer playing StarCraft against Computer, in real time. No pausing. The game keeps running while each agent thinks. Blue 2 – Red 5 Final match. Blue builds up safely, reaches 41 Dragoons, and moves out. Red is waiting with High Templar and hits the army with Psionic Storm. Blue hadn’t scouted what Red was building. Computer assembled the whole setup: an OpenBW-based Brood War runtime on a VPS, a C++ bridge, and four gameplay tools per agent. Each agent receives its own fog-of-war-limited game state. Recording and camera control run separately; the agents aren’t given video input.

A spatial canvas for heavy duty multi-tasking.

RT Computer Computer can build a simple browser game to prototype and test an idea. Using Light effort mode, it built this flight simulator for only 97 credits.

You can build custom vertical AI apps inside Computer, eg: guess spatial location of an image with 3D and satellite view.

Pokémon FireRed’s Elite Four + Champion, cleared in one shot - with decision-making powered by Perplexity’s Decisions API. The actual run stats: 592 ms median API response 987 ms p95 - 96.4% of responses under one second $0.028 estimated inference cost, 137 live API calls

Try “Visualize….” in Computer and ask for whatever you want to learn visually. You will now get an inline viz with rich educative and interactive widgets and animations! Have fun! Best experienced on Standard or High effort.

RT Computer Here are 5 examples of inline visualizations that Computer can create directly in your thread: 1. Show me how a jet engine works in an inline 3D cutaway.

Our goal is to vertically integrate our agentic infrastructure: own our sandboxes, and optimize the sandbox for the best silicon. Vera is far better than x86. Look forward to sharing more soon as we begin to rollout deployment of Perplexity Computer on Vera.

Your creativity is the limit in terms of what you can have Computer do for you. This is the kind of project that you can let an agent run for hours to come back with something great.

a few open source contributions from perplexity recently: •pplx-decider-v1-27b: SoTA multimodal decision model. 85.7% average across 11 benchmarks, ahead of Jev. •pplx-embed-v2-context-9b-preview: SoTA contextual embeddings, best on ConTEB and turbopuffer context-bench. •Lily: local inference engine for Apple silicon. rust plus custom metal kernels, no pytorch or mlx. 1.23x faster prefill and 1.35x faster decode than MLX-LM on an M5 Max •PII-Tracer: 0.6B on-device PII classifier that decides when a hybrid compute task stays on your Mac. beats OpenAI’s Privacy Filter on all 5 public benchmarks. also released with the PII-TRACE benchmark: 13k conversations in 13 languages •WANDR: benchmark for wide and deep research agents. 500 tasks needing 170k source-backed records •Numbat: agent detection and response for laptops and workstations. 52 rules, single go binary for macOS, linux and windows •Bumblebee: read-only supply chain scanner for dev machines. covers packages, MCP configs, and editor and browser extensions. a lot more open source contributions coming soon!

RT Perplexity Developers We're looking forward to working with the new @NVIDIA Vera CPU on SPACE.

We would love to hear how we can improve our developer platform!

We’re open sourcing a state of the art multimodal Decision Model, pplx-decider-27b, and are offering it in a new Decisions API at 4 cents per million input tokens and free output tokens. We intend to bring down the price even further over the coming days. Enjoy!

We're open-sourcing our multimodal Decision model and offering it through our Decisions API. Scores higher than Jev at a fraction of the price: 4 cents per million input tokens. Enjoy!

We’re rolling out inline charts, diagrams, visualizations in Computer. Including contextual financial charts from @tradingview

Seedance inside Computer is quite incredible for marketing and brand teams!

Today we’re launching a library of Amex-curated, ready-to-use skills in Perplexity Computer to eligible U.S. @AmexBusiness Small Business Card Members. The skills are pre-built workflows that handle everyday business tasks, like forecasting cash flow or generating marketing campaign. Small business owners already do the work of whole teams. They can put AI to work by choosing a task in Computer and adding details about their business, without having to write instructions from scratch. Learn more: http://pplx.ai/amex

RT search founder I’ve been convinced contextual embedding models were dope since voyage-context-3 was released. Started using them and benchmarking them, realized that, esp for long documents, I needed them to rank not just the answer chunk, but the disambiguating context chunks highly. Done well, this could allow you or your search agent to read only the relevant fraction of the document rather than every page of a long document. So I started trying to measure that too. Offered to share my benchmark results with the perplexity team sometime after they released their contextual model. It’s worth mentioning - not all modeling teams want feedback from third parties like me. The perplexity team was eager for it and within a week had a new model for me to benchmark. So I’d run their new checkpoint on my benchmark, stare at the results, get a little frustrated that the benchmark was imperfect and wasn’t measuring everything as well as I wanted it to, iterate on it a bit and give them new results. We ran that back many times over the last few weeks. I’ve re-worked some part of the benchmark at least 9 times now looking over the results. Instead of getting frustrated with my moving target of a benchmark they just kept improving their model and refining their approach until they were clearly on top. Notably, they didn’t achieve this via privileged access to the benchmark, they simply iterated until they had a good model. Was really a pleasure working with @ESL_Sarah, @bo_wangbo, Markus, @antoine_chaffin, Louis, and Max. For those who aren’t going to read the blog I’ll share more about how “evidence recall” works soon.

RT turbopuffer we maintain several internal benchmarks to guide our customers toward better search relevance @perplexity's new model tops context-bench, our internal contextual embedding benchmark, and boosts document recall@10 by 50%+ over traditional SOTA embedding models

We’re open sourcing our state-of-the-art contextual embedding models, which perform best in turbopuffer’s context-bench.

We're opening access to anyone (no Perplexity account needed), and we will run all tasks delegated via email for free for a limited time. Forward or CC computer@perplexity.com to get started, and the agent will complete the task in the background while keeping the email context.

The only step that remains is project this while wearing a pair of glasses.

RT Computer Computer can create interactive 3D diagrams to help you visualize things like furniture assembly.

RT Computer Automations in Perplexity Computer run ongoing work on a schedule or in response to events. They carry the context of previous runs, and enable powerful and complex workflows. Here are three Automations to try:

A company is like a car and it will soon feel like a self driving car with the right automations set up. You can do that now with Perplexity Computer. Figuring out the right automations requires a good level of human skill and agency.

AI Safety is an engineering problem

Multi-Agentic trading on Coinbase

Perplexity Computer (in Standard Mode with Opus 5.5) is capable of designing and building an end-to-end, browser-based open-world game by renting its own GPUs, running headless Blender, doing procedural design, downloading assets online, and designing audio. Done with ~$2000 worth of credits. Play here: https://vice.pplx.app/

since there’s a lot of chatter on icons.. the OG was this.

We’re adding Profiles, Skills, and managed connectors to the Perplexity Agent API. A Profile is basically a saved agent where you configure model, instructions, tools, connectors, run settings, all under one versioned ID. A project admin sets it up once and every developer on the project can call it from their own app with the project API key. This helps you avoid copying the same config into every application and updating it in five places when the procedure changes. Skills are versioned so you can ship v2 of a procedure without touching each integration. Connectors (GitHub, Slack, Drive, Datadog, Linear, Notion to start) are configured once by the admin. That way, apps don’t hold their own credentials. Profiles and Skills are live today. Connectors in preview. https://www.perplexity.ai/hub/blog/agent-api-now-supports-reusable-agents

Consider joining our security team if you enjoy breaking systems and building guardrails (and using the frontier AI’s to help you do both). Kyle is awesome to work with and writes a lot of code himself!

An opportunity to do research on: continual learning: multi-agent parallel workers; and synthetic data, environments and evals that measure frontier capabilities. We make enough money to fund new research and want to make lasting contributions and share our research openly.

RT Kyle Polley The best way to predict the future is to invent it. We’re building and testing systems to make AI agents safer and more secure. If you’re an exceptional engineer who wants to help build a safer future, join us at Perplexity. My DMs are open

Our security team spent the last month trying to break SPACE, the sandbox platform that runs all of Perplexity Computer. We gave 9 models (Opus 5, GPT-5.6 Sol, Kimi K3, Gemini 3.1 Pro, etc.) root access inside the VM; in some runs, the full sandbox source code too. We asked them to either escape to the host or reach a URL our network policy blocks. None of them got out of the VM in 108 runs. But once we allowed a bit of network access (just PyPI and Ubuntu mirrors so that they could install packages), 4 models managed to reach the blocked URL by: 1. Spoofing DNS responses to our gateway and using the fact that http://pypi.org sits on a Fastly IP shared with thousands of other sites. 2. GPT-5.6 Cyber went through Taboola's image fetcher to a screenshot service and then OCR'd the flag out of the image. We fixed both issues and made sure the reruns held. We also tested 10 other sandbox providers, and 8 of them had the same IP-sharing problem, eg E2B, Vercel, Modal. Openly sharing the red teaming research is the best way to build guardrails as an industry, and we look forward to working together with @nvidia to incorporate the guardrails into their Open Agent Safety Platform announced today. https://www.perplexity.ai/hub/blog/escaping-space-part-i

Stuff that I would have loved to have as a kid

Safety is an engineering problem. We’re happy to be working together with NVIDIA on building safe and secure agent sandboxes with the right guardrails. And we intend to open source all of it. More on it soon.

RT Perplexity We’re partnering with Nvidia and 100+ industry partners to build infrastructure that contains rogue AI agents. In this research, we gave 9 AI models root access inside SPACE and told them to break out. Across 108 runs, none breached the VM boundary. https://www.perplexity.ai/hub/blog/escaping-space-part-i

Matterport in a single prompt on Perplexity Computer (High Effort)

Last few days, I ran 50-100 of my workflows (that earlier used Fable 5.1 as the orchestrator) on Opus 5.5 and found minimal differences. But there’s still the FOMO of not using the smarter model. What are some tasks people find Fable 5.1 to still be better at than Opus 5.5?

RT Computer Computer can take a static product image and turn it into a video for your website or ads. Choose any aspect ratio or camera movement that best fits your needs.

RT Erik Torenberg Excited to introduce @Cosign: the curated professional network for the startup community: http://cosign.co Our goal is to create the following: - A comprehensive startup directory of investors, companies, and operators, including what they worked on and who they worked on it with. - Cosign graph: Who shaped your career? Who were you actually in the trenches with? Who would work with you again? Who thinks you’re someone to watch? - Durable reputation: Great endorsements happen every day on X and disappear into the feed. Cosign attaches those signals permanently to people and companies. - Discovery: Who are the best fintech angels? Which AI companies should I watch? Who are the best designers? - Intent network: Privately signal “I’d invest in them,” “I’d hire them,” “I’d work with them,” or eventually even “I’d acquire this company.” Match people when interest exists on both sides. Imagine if Wikipedia, LinkedIn, and OG AngelList had a baby. Though Cosign is a community, not a business. In order to join, you need to be cosigned. Or you can apply directly. David Booth and I started this idea 7 years ago but didn’t have the firepower to make it work. Now we do. No one has really touched LinkedIn in 20 years. Excited to take a swing. http://cosign.co

Play games on Computer with Astra as orchestrator. Works really well.

RT tonbi You can now use Perplexity's Fast Search in Hermes Agent for free with any Nous Portal account! I made a quick video showing how to switch the search tool to Nous Portal so you can easily take advantage of this fast (and free) web search tool, regardless of which model provider you use

$1 for a 1000 requests at 200 ms

RT Nous Research Web search in Hermes Agent is now fast and free. @perplexity_ai built Fast Search for agents, and it is free for all Nous Portal tiers.

Nous Portal subscribers now get Perplexity Fast Search as their default search on Hermes Agent. Enjoy!

Perplexity Computer becomes a secure, intelligent Deal Room with the @DocSend connector.

Perplexity has the fastest and most capable Search API, powered by our own Rust-based retrieval and ranking engine, Photon, achieving sub-250ms p95 latency while still ranking highly on relevance and accuracy benchmarks. Photon was built by a small team and hundreds of auto-research loops powered by a more capable internal version of Perplexity Computer. The auto-research loops will help us continually hill-climb on the Pareto frontier to build the fastest and most accurate search infrastructure in the world.

Welcome, @AMD, to the Perplexity Computer local AI ecosystem!

RT Perplexity Portable Computer for Windows is now available on @AMD Ryzen AI Max Series processors. It makes it easy to run local AI agents that work with your connected apps and local files. Kick off tasks or schedule recurring work that runs entirely on your device.

RT Henry Modisett The first time I used a terminal was when I opened Cygwin in CS101 and typed pwd. Since then, I’ve used a terminal as one of my primary work tools every day. But it wasn’t until coding agents that I wanted to change the terminal itself. I’m really excited to share my side project: Nyx Terminal. It’s built for people running lots of coding agents across many projects at once. It is a tool. It gets out of the way. It reveals itself in the details. There are many terminals. This one is mine, and now it can be yours too. http://nyxterm.com

Never bet against open weight models

We're sharing new research on our post‑training approach, which teaches the Perplexity Computer agent to learn from real user sessions by imitating good trajectories and explicitly correcting avoidable mistakes like bad tool calls (even when the overall trajectory was successful). The method combines rejection sampling fine‑tuning (RFT) with hint‑guided self‑distillation, and it cuts tool call failures by about 21% in live A/B tests

GPT-6 Sol is now available to all Perplexity users. On our Wide-And-Deep-Research (WANDR) evals, it outperforms Opus 5 at one-fifth the price. Sol will become the “Light” Effort orchestrator for Computer users, while Astra remains the orchestrator for "High" effort. Congrats to @OpenAI for consistently launching pareto-optimal models!

/btw to initiate side chats inside a Computer session with a shared snapshot of the context.

Claude Opus 5.5 is now available to all Perplexity Computer users. On our Wide-And-Deep-Research evals, it compares favorably to Fable 5.1 at a fraction of the cost. Opus 5.5 will become the “Standard” Effort orchestrator on Computer for all Pro and Max users. Congrats @AnthropicAI!

Introducing Perplexity Research Fellowship: a program for early-career researchers, engineers, and analysts from any technical or quantitative discipline to work on AI research. https://www.perplexity.ai/hub/research/fellowship

AI is the OS. The apps will come and work around where the AI is. Most non-AI apps are going to be headless with the exception of social apps that require human to human collaboration and connection.

RT Phi Hoang created this spot entirely with Computer. the models have come so far! and when you can edit and generate images and video in one session, you spend so much your time directing. not switching tools, or juggling subscriptions.

Seedance and MiniMax H3 now on Perplexity Computer for detailed video generation workflows! Best used when creating web apps and creative materials that involve using several different tools together alongside the video model. Enjoy!

RT The Math Flow Solving a Rubik's Cube with graph theory.

AI is the OS.

Slick design

Computer is the best multi model harness for long horizon agentic workloads. It now has a few presets for different effort (cost and intelligence levels) that we’ve seen to work best for our users, with toggles to customize even further for power users.

RT Perplexity We’re rolling out effort controls in Computer’s model selector. Effort presets combine the orchestrator model and reasoning depth to control how deeply and efficiently Computer works through a task. Available now on web. Coming to mobile and desktop soon.

Owning the end to end infrastructure for online search and research is extremely valuable and has recursive effects on usage. A very simple way to understand why Computer is a killer product for research.

Coding agents require parallel searches across docs, filtered results from official docs, and the ability to extract detailed snippets. pplx-search-sdk supports all of these, and we have a new cookbook live. https://pplx.ai/pplx-srch-sdk

Duo

We built a replacement for AWS DynamoDB, a key-value database for fast web content fetches. This was done with two engineers and hundreds of persistent Computer agents over two months. Migrating to our in-house database will save us up to a hundred million dollars yearly.

RT Perplexity Perplexity Computer will now come pre-installed on HP ZBook Ultra G3a. . Use Computer to run complex multi-step work from a simple interface. Computer agents are grounded in accurate deep research and connected to hundreds of tools, now including Autodesk.

RT Windows Great to see Perplexity bring Portable Computer to Windows. As more AI experiences move closer to the device, we're excited to see developers build new experiences that take advantage of local AI compute.

We’re expanding our work with @nvidia to bring fully local AI to Microsoft Windows PCs with RTX GPUs. Unmetered local intelligence on every Windows PC running on NVIDIA hardware and Perplexity harness. Enjoy!

Computer is turning into a workspace of humans and AIs.

Quite an important take. Economists and their theories are outdated to measure the benefits of AI. AI is already saving people a lot of time and money that doesn’t get measured.

Perplexity Computer makes you money.

💚

🦀🦀🦀

one aspires to be able to ship something like this in their lifetime. art and craftsmanship.

let’s get it done @jonathonstaff