Browser infra for your AI Agents
San Francisco, CA
Based in United States
- Tweets 776
- Following 89
- Followers 9,107
- Likes 1,281
Opus 5.5 is now available on Hyperbrowser 🚀
It handles most tasks as well as Fable 5.1, while costing about 40% less than Opus 5 on typical workloads
Give it a task and watch it use the browser
Try it now in our AI Agents Playground 👇
2
5
7
708
Then the three newest browser agent papers on arXiv, each one confirmed on its own abstract page instead of trusting the search listing.
17 steps, 256k tokens.
2
3
208
We asked Opus 5 to read GitHub's trending Python repos and say what each project actually does, not just parrot the description.
18 steps through Claude Computer Use on Hyperbrowser. It opened every README before answering ↓
1
5
334
Opus 5 is now available in Hyperbrowser.
It comes close to Fable 5's intelligence at half the price, and that it verifies its own work and keeps iterating until it succeeds.
We paired that with stealth browsers and residential proxies.
Just got scary good↓
3
2
14
1,665
We had one Kimi K3 browse 30 pages at once, on Hyperbrowser.
Browserswarm spins up a swarm of cloud browsers that read the web in parallel and stream every page into a single shared context. One brain, many eyes.
428k tokens into one window, zero failures, one cited answer in just a few minutes.
1
3
13
1,645
We gave Gemini 3.5 Flash-Lite a multi-hop research task in a Hyperbrowser cloud browser: dig through GitHub's docs, find the REST API rate limits, check whether GraphQL differs, and summarize both.
24 steps. Multiple pages, navigation decisions at every hop, no hand-holding. It finished:
1
2
150
Here's one run: we gave it a real task and it finished in 5 steps.
21,173 input tokens, 863 output. The whole thing streamed live in the browser, every click and scroll visible. This is what "computer use" looks like when the model has an actual environment to work in.
1
2
208
Gemini 3.5 Flash-Lite is cheap and genuinely capable at computer use. Google's most affordable model.
A computer-use model is only as good as the browser you give it. It's now live on Hyperbrowser:
Stealth, proxies, captcha solving, and a live view you watch it work through. The perception is cheap now, we gave it the environment.
2
3
11
1,388
Fable cost 6x more per completed task than Grok ($0.06 vs $0.01).
Sol sat in the middle.
Setup: live websites, one system prompt, one action space, no per-model tuning. Fresh cloud browser per trial, interleaved so no model saw a different web.
Environment and API failures were tagged separately and excluded from model scores.
2
221
We benchmarked GPT-5.6 Sol (high), Grok 4.5 (high), and Claude Fable 5 on 24 browser tasks, run on Hyperbrowser with an identical harness,
3 trials per task.
Fable 5 led overall at 78%, Sol 75%, Grok 4.5 69% and each specialized: Fable was perfect on reading, Sol was best at navigation (94%), and Fable led on forms and logins (89%).
Full task set and harness are open source. The breakdown below ↓
2
7
13
1,817
GPT-5.6 sol, running on hyperbrowser.
Threw it at kayak. popups, cookie walls, a flight search that reloads every time you touch a filter. one task: sfo to jfk, one-way, top 3 results back as json.
Openai’s flagship model, up against the sites that fight back ↓
1
4
102
GPT-5.6 terra in the hyperbrowser playground.
Pointed it at a github repo. stars, forks, latest release, back as json in one pass.
The everyday model, doing an everyday check ↓
1
3
248
GPT-5.6 luna is the fastest and cheapest of the family.
Pointed it at hacker news. top 5 posts, points and comment counts, back as json in one pass.
Smallest & cheapest model in the family, still gets it done ↓
2
4
331
GPT-5.6 sol is insane when paired with cloud browsers.
Sol, Terra and Luna are now available on Hyperbrowser.
Openai's most capable models yet, running computer use tasks on cloud browsers and sandboxes.
here's how it performs ↓
2
10
31
2,724
We put grok 4.5 and opus 4.8 head to head on the same browser task, on Hyperbrowser Sandboxes.
We then asked it to open a page in a real sandboxed browser and pull the title. Grok build on Grok 4.5, Claude code on Opus 4.8, identical setup.
Grok 4.5 came out ahead ↓
13
16
285
406,470
Coding agents waste a lot of tokens just figuring out how a site is structured before they can touch it.
Map it once instead. Agentmap crawls any site and turns it into a structured map, pages, flows, and the data on each, that any agent loads in one shot. claude code, cursor, a browser agent, all of them skip the blind exploration and go straight to the task.
Open source, powered by hyperbrowser ↓
14
15
94
15,639
We had fable 5 and sonnet 5 build with sandboxes. no instructions given.
Asked both of them to complete a browser task and extract the result inside the sandbox.
Here's the result, Powered by Hyperbrowser Sandboxes ↓
3
6
22
3,205
81.2% on osworld-verified, anthropic's computer-use benchmark. that score is the skill that matters for browser agents, actually operating real sites, not just reading them.
Example:
1
1
146
Sonnet 5 + Browser Agents.
The browser smarts of a flagship model, at sonnet pricing. the capability that needed the biggest, most expensive models a few months ago now runs cheap.
Rolling out in hyperbrowser agents now ↓
1
2
14
1,199
Introducing HyperScript, an open-source browser agent builder.
Describe any web task in plain English and it writes a real Hyperbrowser script to do it. Watch it run live, then keep the code.
Powered by Hyperbrowser.
8
5
39
13,003
GLM-5.2 can reason, but it can’t open Chrome, click links, or navigate the web by itself.
Hyperbrowser spins up a real cloud browser: proxy, stealth, captcha handling, the works.
Each turn the harness:
→ reads the page as text (links, buttons, visible copy)
→ asks the model for one next action
→ clicks/types in the real browser
→ repeats until the task passes or fails
1
2
6
450
We gave GLM-5.2 a cloud browser.
The open model can't browse on its own. So we dropped it into a Hyperbrowser cloud browser and sent it at the real sites on the web, against the closed frontier.
Open-source. Built with Hyperbrowser.
It got scary good ↓
3
6
28
2,070
/loop
Gives your coding agent a browser that retries until the task is done.
It observes, acts, and verifies in a loop inside a Hyperbrowser sandbox. One-shot tools fire once and hope. /loop recovers.
Built with Hyperbrowser Sandboxes.
6
10
36
4,235
your agent needs a machine for thirty seconds, not thirty days.
spin up a full linux box in under 50ms. run the task. throw it away when it's done.
disposable compute for agents ↓
1
5
9
817
We gave browser agents a brain.
Eyes to read any page visually, the way a person would. Memory to remember what they saw, forever. Your agent learns a site once and never re-learns it.
Most browser agents are stateless. They open a page, figure out the layout, act, then forget everything the moment the session ends. Next time, they start from zero again.
BrowserBrain reads a page with vision, stores what it saw, and recalls it in milliseconds on the next visit.
Open-Source Code here ↓
1
8
12
1,399
Introducing HyperView.
Generative UI for coding agents.
Terminal, diffs, live previews, anything. Your agent renders real UI for you while it works in a sandbox.
Open-Source & Powered by Hyperbrowser Sandboxes.
2
11
44
4,878
/harness
A Claude Code plugin built for Fable 5. It runs your agent against your project, watches it break, and writes a CLAUDE. md from the actual failures.
It catches wrong paths, missing scripts, and bad assumptions so your agent never hits the same wall twice ↓
11
59
680
78,895
Your markdown is probably lying to your AI agents.
Meet HyperHarness. A self-improving harness for coding agents that learns from its own mistakes.
Paste a repo. We run your agent in a sandbox, watch it fail and rewrite your CLAUDE. md from what actually went wrong.
Open source. Built on Hyperbrowser Sandboxes.
6
10
39
4,061
/web
Creates a web. md file for any website. Every page, action, flow, and interactive element your coding agent needs to navigate it.
AI agents fail on websites because they don't know the structure. Web. md gives them the map before they start ↓
12
31
415
72,541
We ran Claude Code inside a Hyperbrowser Sandbox.
One prompt. It wrote the code, installed dependencies, and started the server.
Live on a public URL. No local setup. No deploy step.
1
2
291
Push files into any running sandbox from your terminal.
`hx vm cp ./test.txt <id>:/home/ubuntu/`
Your local files. Running in the cloud.
1
2
134
Snapshotted a running sandbox. Stopped it. Restored it from the snapshot.
The server was still running. No restart. No rebuild. Watch it pick up right where it left off.
`hx vm snapshot create <id> my_checkpoint`
1
3
265
Spun up a sandbox. Built a Next.js app inside it. Exposed it to a public URL. Under 200ms to start. No deploy step.
`hx vm create node --port 3000`
1
2
388
We gave AI agents their own computer.
Sub-50ms startup. Memory snapshots. Full terminal access. Filesystem API. Persistent volumes. All in one sandbox.
Hyperbrowser Sandboxes ↓
4
8
47
4,883
Opus 4.8 searching flights on Kayak. Picking dates. Applying filters. Finding the cheapest option. No edits. No cuts. One API call.
1
2
296
Opus 4.8 scores 83.4% on agentic computer use. The highest any model has scored.
Handles logins, popups, dynamic pages, and multi-step flows seamlessly. Flags its own mistakes instead of failing silently. Browser agents just got a major upgrade.
Now live on Hyperbrowser ↓
3
6
42
3,242
Introducing AgentRank.
AI agents are browsing the web for real users now. If they can't navigate your site, you're losing traffic.
We run Claude, GPT, and Gemini on your site. See where they get stuck and what to fix.
Open source. Powered by Hyperbrowser.
1
10
41
4,839
Your AI coding agent can now extract any website's design
system directly from the terminal.
/design
Fetches the branding. Extracts colors, fonts, spacing, components.
Saves it as a DESIGN. md
Powered by Hyperbrowser. Setup guide ↓
6
15
113
14,164
xAI open-sourced the algorithm that scores every tweet on X.
19 engagement signals. Replies 13.5x. Reposts 20x. Bookmarks 10x.
We built PhoenixScore. Paste a tweet. It runs through the algorithm and shows you exactly how it scores.
Open source. Powered by Hyperbrowser.
15
20
373
56,143
Snapshot a running sandbox. Restore it later. Memory, files, processes. All preserved.
hx vm snapshot create
hx vm create --snapshot
1
2
230
Fetch any page. Search the web. Crawl entire sites. Straight from your terminal.
hx web fetch
hx web search
hx web crawl
1
5
269
We built a CLI that puts the entire Hyperbrowser platform in your terminal.
Spin up sandboxes. Launch stealth browsers. Fetch, search, and crawl the web. Snapshot and restore. All from two letters. ↓
2
6
27
2,241
HyperRank helps brands see how AI search engines describe them, which competitors they recommend, and what to change to show up.
It runs real prompts across ChatGPT, Claude, Perplexity, Google through stealth browser sessions.
Paste a URL. Get your AI visibility scorecard ↓
5
8
29
5,233
Snapshot and preserve state. Fork workflows at scale.
hx vm snapshot create
hx vm create --snapshot
1
1
226
Web APIs now in the CLI. Fetch, search, and crawl without writing a single line of code.
hx web fetch
hx web search
hx web crawl
1
1
319
Introducing the new Hyperbrowser CLI.
Sandboxes. Browsers. Web automation. All from your terminal.
Under 50ms sandbox startup ↓
2
8
134
8,526
