TL;DR
Google Project Astra is still a research prototype in November 2026, so build its business uses on Gemini Live or the Gemini Live API. Only a limited number of trusted testers use Astra itself, while Google runs a waitlist instead of a product launch. Its video understanding, screen sharing and memory have moved into Gemini Live, which suits quick, low-risk trials. Teams that need their own tools, data and logs should build on the Live API with gemini-3.8-live, generally available on Google Cloud. IT support and training work now in Gemini Live, while field service, retail, inspection, intake and translation need a custom build. Accessibility services and glasses remain prototypes, so pilot only what ships and keep a human in the loop.
Key Takeaways
Google Project Astra is a Google DeepMind research prototype of a universal AI assistant, first shown in 2024 and so far used only by a limited number of trusted testers.
On May 20, 2025, Google said Astra’s video understanding, screen sharing and memory had also moved into Gemini Live.
Developers build Astra-style assistants on the Gemini Live API, using gemini-3.8-live, generally available on Gemini Enterprise Agent Platform, formerly Vertex AI.
Audio-video Live API sessions end at 2 minutes without context compression, so session design comes before the first pilot.
Of the eight business applications, two work now in Gemini Live, while five need a custom build and one stays a prototype.
Camera-on assistants need consent, retention limits, ephemeral tokens, scoped tool permissions and human oversight before launch.
Watch on YouTube
Google’s A2A Protocol: How AI Agents Collaborate Without Humans
A Kanerika video on Google’s Agent2Agent protocol and how AI agents pass tasks to each other, the pattern behind assistants that act as well as answer.
A Technician, a Phone Camera and a Fault Code
Picture a field technician kneeling beside a commercial boiler at the start of her shift. The control panel shows a fault code she has never seen, but the senior engineer is two hours away.
So she points her phone camera at the panel and asks what the code means. The assistant reads the label and finds the right page in the service manual. Then it talks her through the reset while her hands stay on the unit.
Google DeepMind first showed this kind of camera-and-voice help in 2024, under the name Project Astra. Two years on, the question for a business is simpler: which parts can you buy or build in November 2026? This guide answers that for eight business applications, and it also gives a readiness verdict for each.
What Is Google Project Astra?
Google Project Astra is a research prototype from Google DeepMind that explores what a universal AI assistant could look like. It sees through a phone camera, listens through the microphone, remembers context and then answers out loud.
DeepMind’s Project Astra page describes it as a research prototype used by a limited number of trusted testers. Demis Hassabis, who leads Google DeepMind, also set out its progress in a May 20, 2025 post on the Google blog .
Astra runs on Gemini, the model family covered in our Google Gemini AI guide. The difference matters for buyers, because Astra is the research testbed while Gemini Live and the Gemini API are the products you can use.
In the terms of AI agents vs AI assistants , Astra sits between the two, since it answers like an assistant yet can also call tools.
Research Prototype or Product? The Status in November 2026
As of November 3, 2026, Project Astra is still not a product you can download, license or deploy. DeepMind’s page lists it as a research prototype, open to trusted testers through a waitlist.
Instead, Google moves Astra capabilities into other products one at a time. Plan around what ships in Gemini Live and the Live API, and treat anything still labeled Astra as a preview.
Not to Be Confused With OpenAI’s GPT-6 Astra
OpenAI sells a model called gpt-6-astra through the OpenAI API . It is priced at $10 per million input tokens and $50 per million output tokens, with a 1.05M-token context window.
That model has no connection to Google’s project, so check which Astra a vendor, article or job post means before you compare anything.
Kanerika Service
Agentic AI Services
Kanerika designs and builds AI agents that see, listen, call enterprise tools and hand off to people, with governance in place from the first pilot.
Explore Agentic AI Services How Project Astra Works: Perception, Memory and Tool Use
Project Astra turns a live stream of sight and sound into a conversation that can also take action. Think of a coworker looking over your shoulder who sees what you see and replies before you finish typing.
DeepMind lists eight capabilities behind that experience, and each one maps to a design choice in an enterprise build.
Natural interaction: a real-time conversation through voice, camera and screen.Proactive responses: the assistant can speak up when it notices something relevant.Context-aware dialogue: it follows the speaker and ignores background noise.Pointing out objects: it shows the user which object or area it means.Tool use: it calls Google Search, Gmail, Calendar and Maps, and it can also control an interface.Personalized reasoning: it remembers preferences and then uses them in later answers.Content retrieval: it pulls facts from content the user shares, such as a PDF manual.Multimodal memory: it recalls what it saw and heard earlier, even after a switch to another device.
Under the hood, these are multimodal models at work, because they read images, audio and text in a single pass.
The Real-Time Perception Loop
The heart of Astra is a loop that repeats throughout the conversation. First, camera frames, microphone audio and screen content flow in. Then a Gemini model reasons over them with recent memory and answers with speech or an action.
Speed is the whole point of that loop. A typed chat can tolerate a pause, but a spoken exchange feels broken when the gap grows. So every stage is tuned for low latency.
The same pattern also powers multimodal AI agents in other settings. It relies on vision language models that turn pixels into words a model can reason over.
Memory and Personalized Reasoning
Memory turns a clever demo into a useful assistant. Astra can retain preferences within a conversation and, per DeepMind, carry memory across devices, so a user does not repeat context.
Content retrieval works the same way at the document level. When a user shares a PDF manual, the assistant answers from that file instead of from general knowledge. That matters when the right answer is one torque value on page 40.
Action Intelligence: Pointing, Tools and Interface Control
The third layer is action. The prototype can call Google Search, Gmail, Calendar and Maps. Google has also tested computer control with trusted testers, where the assistant operates an interface for the user.
In an enterprise build, tool use is where most of the value sits, and most of the risk too. For example, an assistant that books a technician visit needs the same permission checks as any system that writes data.
What Happened to Project Astra: What Shipped and What Is Still a Prototype
Project Astra’s capabilities have moved into Google products one at a time. The table below separates what you can use in November 2026 from what still stays in the lab, so you can check each line at its source.
Capability Where it lives Status Source
Video understanding through the camera Gemini Live Shipped Google blog, May 20, 2025 Screen sharing Gemini Live Shipped Google blog, May 20, 2025 Memory Gemini Live Shipped Google blog, May 20, 2025 Native audio, better memory, computer control Trusted testers In testing as of the May 2025 post Google blog, May 20, 2025 Real-time voice and vision for developers Live API, gemini-3.8-live Generally available on Gemini Enterprise Agent Platform Google Cloud documentation New Search experiences Google Search Named as a destination DeepMind Astra page Glasses Prototype glasses Prototype, no launch date DeepMind Astra page The full Astra assistant Trusted tester program Research prototype, waitlist DeepMind Astra page
Gemini Live: Camera and Screen Sharing
Gemini Live is where most people meet Astra’s ideas. A user opens Gemini on a phone, turns on the camera or shares the screen, and then talks about what it sees.
Google’s May 2025 post says video understanding, screen sharing and memory from Astra have moved into Gemini Live. For a business, that makes Gemini Live the fastest way to test camera-on help before anyone writes code.
The trade-off is control, because a consumer app has no custom tools, no link to your manuals and no logs you own.
Search, the Live API and Glasses
Hassabis named the next destinations in May 2025: Gemini Live, Search, the Live API and new form factors such as glasses. Of these, the Gemini Live API is the one developers can build on as of November 2026.
Glasses are the furthest out. DeepMind shows Astra running on prototype glasses. Separately, Google said on May 19, 2026 that Gemini audio glasses from Samsung, Gentle Monster and Warby Parker would launch later in the fall. Google has given no date for Astra’s camera features on glasses, so no business plan should depend on hands-free, camera-on glasses yet.
Search is the quieter channel. DeepMind lists new Search experiences as a destination, so some customers may meet Astra-style help inside Google before they meet yours.
How Project Astra Supports Blind and Low-Vision Users
Accessibility is where Astra’s camera-and-voice design could help the most people, so DeepMind treats it with care. A blind or low-vision user may need a room described, a label read or a doorway found.
DeepMind has partnered with Aira, a visual interpreting service, on a research prototype called Visual Interpreter. It runs as a trusted tester program with live professional human oversight, so trained people stay in the loop.
That design choice carries a lesson for every business use in this guide. When a wrong answer could hurt someone, the pattern is AI first with a qualified person watching each session.
Building Astra-Style Assistants With the Gemini Live API
An Astra-style assistant is a Gemini Live API application with your tools and data attached. The Live API handles the real-time part, since it runs a stateful WebSocket session that streams audio, images and text in and speech out.
Per Google’s Live API documentation , input is 16-bit PCM audio at 16kHz, JPEG images at up to 1 frame per second, and text. Output audio is 24kHz, and the session can run server-to-server or client-to-server.
Google also lists partner platforms for the transport layer, including LiveKit, Pipecat by Daily, Fishjam and Vision Agents. Teams that already run AI voice agents will find the shape familiar, with a camera added.
The build then follows seven steps, close to any guide on how to build AI agents :
Pick one task and one success metric, such as first-time fix rate for field repairs.
Choose the Live model, usually gemini-3.8-live, and decide which work goes to a slower back-end model.
Set up the session, including audio format, frame rate, voice and system instructions.
Connect tools through tool calls, such as a manual search, a ticket system or a parts lookup.
Add session management with context window compression and session resumption.
Secure the client with ephemeral tokens, then scope every tool permission.
Test on real footage and audio from the work site before you pilot with a person on call.
For the wider design, our guide to AI agent architecture also covers memory, tools and orchestration.
Which Gemini Models to Use
Model choice in a live assistant is a split decision. One model holds the real-time conversation, while heavier work such as a long report goes to another model through a tool call.
The table draws on Google’s gemini-3.8-live model page , its model list and the Gemini API pricing page , as checked on November 3, 2026.
Model Role in an Astra-style build Context Price per 1M tokens
gemini-3.8-live Default Live API model for low-latency voice agents Input 131,072 and output 65,536 tokens Free tier; paid input $0.75 text, $3.00 audio, $1.00 image or video; output $4.50 text, $12.00 audio gemini-3.8-live-extended-thinking High-reasoning Live model for harder spoken tasks See the model page See the pricing page Gemini 4 Argon Flagship reasoning outside the live session 1M tokens $2 / $10 through November 30, 2026, then $4 / $20; no free tier Gemini 3.8 Flash Lower-cost back-end text work See the model page $0.75 / $3.75 through December 31, 2026, then $1.50 / $7.50
Per minute, the 3.8 Live rates come to $0.005 for audio in, $0.002 for video in and $0.018 for audio out. For example, take a six-minute support call with the camera on, where each side speaks for three minutes.
Audio in is then 3 x $0.005, or $0.015, while video in is 6 x $0.002, or $0.012. Audio out is 3 x $0.018, or $0.054. So the total is about $0.08 per call at list rates, before text tokens or back-end calls. Measure real sessions before you set a budget.
Session Limits and How to Work Around Them
Session limits shape a live assistant more than any other setting. Per Google’s session management guide , a session without compression ends after 15 minutes of audio only, or 2 minutes of audio with video.
A single connection also lasts about 10 minutes. Because a two-minute ceiling would end most field repairs midway, treat these limits as a design input from the first sprint.
Turn on context window compression so the session can continue past the default limit.
Use session resumption, so a dropped or expired connection picks up where it left off.
Send camera frames when the view changes, and pause video when the user only needs to talk.
Move long tasks, such as writing a service report, to a back-end model through a tool call.
Taken together, these steps keep a long repair inside one conversation from the user’s point of view.
Running It on Google Cloud
Enterprises on Google Cloud can also run the same assistant through Gemini Enterprise Agent Platform, the new name for Vertex AI since April 22, 2026. Its Live API overview lists gemini-3.8-live as generally available, with native audio, voice activity detection, affective dialog, proactive audio and tool use.
Production apps should never ship an API key inside a phone or browser. Instead, Google’s guidance is to use ephemeral tokens for client-to-server connections, or to keep the model connection on your own server.
Eight Business Applications of Google Project Astra, Ranked by Readiness
The eight business applications of Google Project Astra below share one test. Does a worker need help with something they can see while their hands or attention are busy?
Each one is ranked by what you can deploy in November 2026. Our guide to AI assistants for business covers text-first assistants, while this list covers the camera-and-voice kind.
Application What it does Readiness in November 2026
1. Field service and remote expert guidance Reads equipment and manuals, talks a technician through a fix Custom Live API build 2. Frontline retail and product help Identifies products and answers shopper questions on the floor Custom Live API build 3. Visual inspection and quality checks Spots visible defects and logs them for review Custom Live API build, human sign-off 4. IT support and workplace productivity Walks staff through software issues by screen sharing Works now in Gemini Live 5. Training and onboarding Coaches new hires on tasks they film or share Works now in Gemini Live 6. Accessibility and assistive services Describes surroundings for blind and low-vision users Prototype only 7. Healthcare front desk and intake Guides patients through check-in and forms Custom Live API build, non-clinical only 8. Multilingual front desk and translation Speaks with visitors in their own language Custom Live API build
1. Field Service and Remote Expert Guidance
A technician points a phone at a pump, and the assistant matches the model plate to the right manual. Then it reads out the steps for a seal change, while a custom build adds a work-order lookup and a parts check.
Gemini Live can help informally, but hands-free use on glasses is still prototype only. For the factory side of this pattern, see our guide to agentic AI in manufacturing .
2. Frontline Retail and Product Help
A store associate holds up a shoe, and the assistant names the model, checks stock and suggests a matching size.
Gemini Live can name a product, but only a custom build can see your catalog, stock levels and prices. Have the assistant confirm the product ID before it quotes a price, because a wrong match costs a sale.
The vision side builds on what computer vision in retail already does for shelves and checkout.
Case Study
95% Accuracy in Counterfeit Detection with AI Vision
A global luxury goods retailer used FLIP computer vision , blockchain provenance tracking and Karl to verify products in real time, with 68% faster verification and 100% traceability.
Read the Case Study → 3. Visual Inspection and Quality Checks
An inspector walks a production line, and the assistant flags a visible scratch, names the station and logs a timestamped photo. Because visual hallucination is a real risk here, a person signs off every reject before it counts.
The Live API’s limit of 1 frame per second suits a walk-around check. A fast conveyor needs dedicated AI video analysis instead.
4. IT Support and Workplace Productivity Through Screen Sharing
An employee shares a screen in Gemini Live and asks why an expense form keeps rejecting a receipt. The assistant reads the error, points to the field at fault and then explains the fix.
This is the most ready use in the list, since screen sharing already ships in Gemini Live. Typical productivity tasks include:
Step-by-step help with software settings and error messages
Short notes that summarize what is on screen
Meeting notes with a list of action items
Project management with proactive deadline and task reminders
For a governed help desk with tickets and escalation, a custom build fits better, as our guide to AI agents for customer support explains.
5. Training and Onboarding
A new warehouse hire films a pallet wrap and asks whether it meets the standard. The assistant compares it with the training guide the hire shared, so a supervisor reviews fewer basic mistakes.
This works now in Gemini Live for informal coaching. A tracked program with completion records, however, needs a build that writes to your learning system.
6. Accessibility and Assistive Services
A visitor with low vision asks the assistant to read a building directory and guide them to the right elevator. DeepMind’s Visual Interpreter work with Aira shows the model, since it runs with live professional human oversight.
Treat this as prototype only as of November 2026. Any service you build should also keep a trained person in the loop.
7. Healthcare Front Desk and Intake
A patient at a clinic kiosk holds up an insurance card, and the assistant reads the details into the intake form. Then it asks the missing questions aloud, but it never gives a diagnosis.
Staff confirm every record before it reaches the chart. Health data also raises the bar on consent and retention, as our guide to AI agents for healthcare explains.
8. Multilingual Front Desk and Translation
A hotel guest speaks Portuguese, and the front desk assistant answers in Portuguese while the clerk sees an English summary. Live audio in and out keeps the exchange natural for both sides.
Test each language pair on real accents and background noise before launch, and keep a staff fallback for any pair that fails.
Kanerika Service
AI Application Development
Kanerika builds production AI applications, from live voice and vision assistants to the tool integrations, testing and monitoring behind them.
Explore AI Application Development Project Astra vs Gemini Live vs Other Assistant Models
Project Astra, Gemini Live and the large API models answer different questions, so the fair comparison is by status, access and price.
Option Status in November 2026 Access Developer route Price
Google Project Astra Research prototype Trusted testers, waitlist None directly; capabilities reach the Live API Not sold Gemini Live Shipped in the Gemini app Phones, through the app Live API Plan terms of the app Gemini Live API (gemini-3.8-live) Generally available on Google Cloud Google AI Studio and Gemini Enterprise Agent Platform Direct Free tier, then paid per token OpenAI gpt-6-astra API model, unrelated to Google OpenAI API Direct $10 / $50 per 1M tokens OpenAI GPT-6.1 Sol API model, published September 29, 2026 OpenAI API Direct $2 / $10 per 1M tokens Claude Sonnet 5.5 and Opus 5.5 API models Anthropic API Direct $2 / $10 and $4 / $20 per 1M tokens
For a camera-and-voice assistant built on Google, the decision is short. Use Gemini Live to test the idea with staff, then build on the Live API once you need your own tools, data and logs.
The other models matter when a Live session hands off heavy reasoning or writing, and each lists a context window of about 1M tokens. For a broader view, see our ChatGPT vs Gemini vs Claude comparison and how Gemini compares with Manus, Genspark and Gemma .
Security, Privacy and Governance for Camera-On AI Assistants
A camera-on assistant sees more than the task in front of it. A technician’s phone may also capture a colleague’s face, a customer’s address or a password taped to a monitor.
Privacy design is therefore part of the build from the first sprint, and our guide to AI privacy covers the wider principles. The NIST AI Risk Management Framework , released January 26, 2023, also gives a shared structure for the controls.
NIST then added its Generative AI Profile, NIST-AI-600-1, on July 26, 2024. The five controls below fit inside that structure, and each belongs in every camera-on pilot plan:
Consent and bystander capture: tell users and bystanders when the camera is on, and keep faces and screens outside the task out of frame.Retention: decide which audio, frames and transcripts you keep, for how long, and who can see them.Ephemeral tokens: never place a long-lived API key in a client app; instead, issue short-lived tokens per session.Tool permission scopes: give the assistant only the tools and records the task needs, then log every call.Human oversight: name who reviews flagged sessions and who takes over when the assistant is unsure.
Agents that act through tools also carry their own risks, covered in our agentic AI security guide. The five controls should sit inside a wider AI governance framework , with a named owner for each one.
Limitations and Production Risks
Every live assistant trades some reliability for speed, and the gaps show up quickly in the field. The six risks below are the ones to plan for, each with a practical mitigation.
Latency: long pauses break a spoken exchange, so keep the Live session on gemini-3.8-live and move slow work to background tool calls.Visual hallucination: the model can misread a label or invent a detail. Require a confirmation step before any answer triggers an action, and see our guide to LLM hallucination .Session limits: audio-video sessions end at 2 minutes without compression, so turn on compression and session resumption in the first prototype.Cost: audio output is the largest per-minute line, at $0.018. Cap session length and track cost per resolved task.Device battery and hardware: a streaming camera drains phone batteries. Test on the devices your staff carry, and consider edge AI when checks must run offline.User trust: one confident wrong answer can stall adoption. Show users what the assistant saw, then let them flag errors in one tap.
None of these risks is a reason to wait. Instead, each one is a reason to pilot a single task with clear limits, a named owner and an honest measure of results.
Checklist
AI Governance Checklist
Check consent, retention, access, logging and human review before a camera-on assistant reaches staff or customers.
Get the Checklist → How Kanerika Builds Real-Time Multimodal Assistants
Kanerika builds AI agents and visual AI systems for enterprises, and an Astra-style assistant combines the two. The work starts with the use case and the data, because a live assistant depends on the manuals, records and tools it can reach.
Our agentic AI services team follows five stages, and each one ends with a go or no-go check against the agreed metric:
Use-case and readiness scoring: rank candidate tasks against the readiness table above, then pick one with a clear metric.Data and consent design: define what the camera may capture, what is kept and for how long.Live API prototype: connect the assistant to enterprise systems through tool calls, such as a manual search or a ticket system.Evaluation on real footage and audio: measure latency, wrong-object rate and handoff rate on recordings from the actual work site.Governed rollout: launch to a small group with human escalation, then widen once the numbers hold.
Two Kanerika assets carry much of this work. FLIP , Kanerika’s AI platform, handles data preparation and vision model training. Karl , our AI data insights agent, also automates checks and answers questions over enterprise data.
For the application layer, our AI application development team builds the interface, the tool integrations and the monitoring.
Case Study: 95% Accuracy in Counterfeit Detection With AI Vision
A global luxury goods retailer needed to catch counterfeit products during returns and resale. Kanerika trained FLIP for computer-vision product authentication using image recognition, and also added blockchain provenance tracking.
Karl then automated real-time verification during returns and resale. The results were 95% accuracy in counterfeit detection, 68% faster product verification and 100% product traceability.
That project is not presented here as a Gemini or Astra build. Instead, it shows the visual AI pattern an Astra-style assistant reuses.
The pattern has four parts: a camera image, a trained check, a system lookup and a fast answer. Only the interface changes, from a scanning station to a phone camera and a spoken reply.
Talk to Kanerika
Planning a Camera-and-Voice Assistant?
Kanerika scores your candidate use cases, designs consent and retention, and builds a Live API prototype with your team, so a pilot starts with one task and one metric.
Book a Consultation → Where Project Astra Is Heading
Google has said where Astra’s capabilities go next, but it has not said when. Hassabis’s May 2025 post named Gemini Live, new Search experiences, the Live API and glasses. DeepMind’s page still lists Gemini Live, Search and glasses.
Model work keeps moving underneath. Google published the Gemini 4 Argon model card on October 13, 2026. It also announced Gemini 4 Carbon that month, although Google has not said which Gemini version runs inside Astra.
For planning, the safe reading is that Astra previews features that reach Gemini Live and the Live API first. In practice, buy what ships, build on what is documented and watch the rest.
Three signals are worth tracking: a date for Astra’s camera features on glasses, longer Live sessions without workarounds, and more of Astra’s tool use reaching the Live API. Each one would move an application in the readiness table up a level.
So, back to the title question. The eight business applications of Google Project Astra are real in 2026, and you reach them through Gemini Live and the Live API. Start with IT support and training, then build toward field service and inspection.
Frequently Asked Questions
What is Google Project Astra?
Google Project Astra is a research prototype from Google DeepMind for a universal AI assistant. It sees through a phone camera, listens through the microphone, remembers context and answers out loud. It can also call tools such as Search, Gmail, Calendar and Maps. As of November 2026, only a limited number of trusted testers use it.
Who developed Project Astra?
Google DeepMind developed Project Astra. DeepMind is the Google research lab led by Demis Hassabis, and it builds the Gemini models that Astra runs on. Google first showed Astra in 2024. Hassabis described its progress in a May 20, 2025 post, which explained how its features were moving into Gemini Live and other Google products.
Will Project Astra work on smart glasses?
Google DeepMind shows Project Astra running on prototype glasses, and it names glasses as one destination. Google said Gemini audio glasses would launch in fall 2026, but it has given no date for Astra’s camera features on glasses. Businesses should not plan around camera-on glasses yet. Phone assistants through Gemini Live or the Live API are the practical route.
What are the business uses of Project Astra?
The main business uses are field service guidance, retail product help, visual inspection, IT support through screen sharing, training, accessibility, healthcare intake and multilingual front desks. IT support and training work now in Gemini Live. Field service, retail, inspection, intake and translation need a custom Gemini Live API build. Accessibility services remain a prototype with human oversight.
What are the privacy concerns with Project Astra?
A camera-on assistant can capture bystanders, documents and screens beyond the task. Businesses should get consent, limit what audio and video they keep, and use ephemeral tokens instead of API keys in client apps. They should also scope tool permissions and keep a person reviewing flagged sessions. The NIST AI Risk Management Framework gives a useful structure.
Is Google Project Astra related to OpenAI's GPT-6 Astra?
No. Google Project Astra is a Google DeepMind research prototype for a camera-and-voice assistant. OpenAI’s gpt-6-astra is a separate model sold through the OpenAI API at $10 per million input tokens and $50 per million output tokens. The two share a name only, so check which one a vendor or article means before comparing them.
How does Project Astra help blind and low-vision users?
Google DeepMind partnered with Aira, a visual interpreting service, on a research prototype called Visual Interpreter. It uses Astra’s camera and voice abilities to describe surroundings, read text and help with everyday tasks. It runs as a trusted tester program with live professional human oversight, so trained interpreters stay involved while the AI does the describing.
Is Project Astra available to the public in 2026?
No. As of November 2026, Project Astra remains a research prototype open to a limited number of trusted testers through a waitlist. The public can use many of its ideas through Gemini Live, which offers camera and screen sharing on phones. Developers can build similar assistants with the Gemini Live API on Google AI Studio or Gemini Enterprise Agent Platform.
Is Google Project Astra free to use?
Project Astra is not sold, so there is no price for it. Trusted testers join through a waitlist run by Google DeepMind. Gemini Live, which carries several Astra features, is part of the Gemini app under Google’s plan terms. The Gemini Live API has a free tier, and paid use is billed per token or per minute.
How do I download or install Project Astra?
You cannot download or install Project Astra as an app. Google DeepMind offers access only to trusted testers, and you can join the waitlist on the DeepMind Project Astra page. For similar help, install the Gemini app and open Gemini Live. Developers can build their own Astra-style assistant with the Gemini Live API instead.
Is Project Astra the same as Gemini Live?
They are related products at different stages. Project Astra is the DeepMind research prototype where new assistant features are tested first. Gemini Live is the shipped product in the Gemini app. Google said in May 2025 that video understanding, screen sharing and memory from Astra had moved into Gemini Live, so Gemini Live is how most people try Astra ideas.
What AI model powers Project Astra?
Project Astra runs on Google’s Gemini models, and Google has not published which Gemini version the prototype uses. For developers building a similar assistant, the Gemini Live API uses gemini-3.8-live as its default model for low-latency voice agents. A higher-reasoning option, gemini-3.8-live-extended-thinking, is also listed on Google’s model page for harder spoken tasks.
Does Project Astra have a developer API?
Project Astra has no API of its own. Developers use the Gemini Live API, which Google named in May 2025 as one destination for Astra capabilities. It streams audio, images and text over a WebSocket and returns speech. Google documents it on Google AI Studio and Gemini Enterprise Agent Platform, formerly Vertex AI, where gemini-3.8-live is generally available.
How much does the Gemini Live API cost?
The Gemini Live API has a free tier. On the paid tier, gemini-3.8-live input costs $0.75 per million text tokens, $3.00 for audio and $1.00 for images or video. Output costs $4.50 for text and $12.00 for audio. That works out to about $0.005 per minute of audio in and $0.018 per minute of audio out.
How long can a Gemini Live API session run?
Without compression, a Gemini Live API session ends after 15 minutes of audio only, or 2 minutes of audio with video. A single connection lasts about 10 minutes. Google offers context window compression to extend a session and session resumption to reconnect after a drop. Most camera-based business uses need both features turned on from the start.
