Aside AI Browser Launch: Features, Benchmarks, Pricing, and Privacy

TL;DR: Aside launched publicly on June 23, 2026 as an AI browser built around real browser work rather than a side-panel chatbot. It combines vertical tabs, browser automation, local memory, a password manager for agent access, and bring-your-own Claude or ChatGPT subscriptions. The strongest claim is benchmark performance. The public benchmark repository reports 297/300 on Online-Mind2Web and detailed Odysseys and BU Bench V1 runs. The privacy story is good but not absolute: Aside is local-first, while account, billing, analytics, sync, recovery, support, and hosted model features can still use server-side services.

Research note: This article was researched on June 25, 2026 using Aside's launch post, official site, browser-agent page, pricing page, privacy policy, public benchmark repository, Hyojun Kim's X profile, Aside's X profile, and Y Combinator company profile. It is a source-based guide, not a claimed hands-on review or security audit. Browser-agent products change quickly, so check the linked sources before making a paid or production decision.

Jump To

Key Takeaways

  • Aside is an AI browser from Aside Computer Inc., a Y Combinator Fall 2025 company based in San Francisco.
  • Founder Jun Kim announced the launch on X on June 23, 2026, describing Aside as an AI browser built to do real work.
  • The product pitch is browser-native automation: it works through websites where you are already logged in instead of waiting for formal integrations.
  • Aside says its agent uses local browsing history and memory, guarded password access, human approval for sensitive actions, and local task storage.
  • The public benchmark repository reports a 99.0 percent Online-Mind2Web pass rate for a gpt-5.5 and openai-codex run, plus Odysseys and BU Bench V1 results.
  • The launch post mentions outperforming Claude Fable. The official site chart I found names Claude Opus 4.8, and the public benchmark README does not show a Claude Fable row.
  • Pricing starts with a free plan at 500 credits per month, with Pro at $20 per month and Max at $200 per month.

What Is Aside AI Browser?

Aside is a browser built around agent tasks. The practical difference from a normal AI sidebar is access. A chatbot can answer questions about a workflow; Aside is trying to operate the workflow inside the browser tabs, sessions, files, and logged-in web apps you already use. Y Combinator describes the company as an AI browser that acts on your behalf across apps such as Gmail, Notion, Slack, Figma, banking apps, and internal tools without requiring integrations.

The launch timing matters. Jun Kim, who posts as @hyojun_at, announced Aside on June 23, 2026. The Aside profile describes the product as an operating system for AI where the browser is the operating layer. That is a big claim, but it fits the trend: agents are moving from chat windows into browsers because the browser already has the user's accounts, history, cookies, files, and permissions.

For SoftPortal readers, the useful question is narrower: can Aside reduce the gap between asking an AI to explain a task and having an AI complete the task in real websites? That depends on task type, account access, safety prompts, model choice, and how much you trust a browser agent with session state and credentials.

Aside AI Browser Features Explained

Browser-native agent work. Aside's homepage says it works across logged-in websites and can handle messages, payments, internal tools, comments, replies, follow-ups, documents, and spreadsheets. That means the browser itself is the integration surface. No plugin can make a flaky website reliable, though. Live pages change, anti-automation behavior changes, and account security prompts can interrupt a run.

Vertical tabs and browser polish. The launch post calls out vertical tabs and Liquid Glass. I would treat that as usability context, not the core story. Agent work still needs a browser that feels fast enough for daily use, so tabs, profile switching, search, downloads, notifications, and permission prompts matter. If the browser feels worse than Chrome, Arc, Dia, Brave, or Safari, the agent layer has to work harder to justify the switch.

Memory from browser context. Aside says it turns browsing history into memory so users do not need to repeat context. The example on the homepage is asking the agent to find a candidate opened yesterday and prepare interview notes. This is exactly where browser agents can beat a generic chatbot: the relevant clue may be in a tab, a history entry, a logged-in app, or a local file rather than in a prompt.

Password manager for agents. Aside positions its password manager as agent-aware. Its site says credentials can be autofilled into websites without exposing raw secrets to the AI, that sensitive actions such as payments, posts, and messages wait for human confirmation, and that credential access is logged. Those controls are not optional decoration. Without scoped access, approval prompts, and audit history, browser agents become too risky for serious work.

Claude, ChatGPT, and API keys. Aside says users can bring a ChatGPT or Claude subscription, or their own API key. That is useful because it separates the browser workflow from a single model vendor. It also means privacy and cost depend on the provider selected for a task. A local browser action may stay local, while a hosted model request can send selected context to the model provider.

Benchmarks, Claude Fable, and What the Numbers Mean

The launch post says Aside is state of the art on agentic browsing benchmarks and outperforms Claude Fable. The official browser-agent page says Aside ranked number one on Online-Mind2Web, BU Bench V1, and Odysseys, surpassing OpenAI, Anthropic, and Browser Use. The public benchmark link goes to at-inc/aside-benchmarks, which is more useful than the marketing line because it lists tasks, failures, configurations, graders, and result files.

Benchmark Reported Aside result What to check before trusting it
Online-Mind2Web 297 passed, 2 failed, 1 impossible out of 300 tasks. Pass rate 99.0 percent, or 99.3 percent excluding impossible tasks. The repository says tasks run against live websites, so inventory, flows, authentication, and anti-automation can change.
Odysseys 151 perfect tasks out of 200, 1,050 rubric items passed out of 1,182, and 86.5 percent task-average rubric score. This benchmark uses partial-progress rubrics. Perfect-task rate and rubric score tell different stories.
BU Bench V1 The gpt55-20260524 run passed 93 of 100 tasks, with 1 impossible task. The README lists a separate Kimi run at 88 of 100. Compare model, provider, settings, and category failures.

There is one wording issue I would not ignore. The X announcement mentions Claude Fable. The official site chart I scraped lists Claude Opus 4.8 at 84.0 percent, not Claude Fable, and the public README sections I reviewed do not show a Claude Fable comparison row. That does not make the launch claim false by itself. It means the cited public evidence I found is not enough to independently verify the exact Claude Fable comparison.

Benchmarks are still helpful. Online-Mind2Web's 300 tasks across 136 popular websites are closer to real browsing than a frozen toy environment. Odysseys is useful because it measures long-horizon work with rubrics, not only binary success. BU Bench V1 mixes tasks from WebBenchREAD, Online-Mind2Web 2, InteractionTests, GAIA, and BrowseComp. But none of these numbers tell you how Aside behaves in your payroll app, CRM, admin panel, bank flow, or messy internal dashboard.

A fair reading: Aside has published more benchmark detail than many AI browser launches, and that is a good sign. A cautious reading: the strongest numbers depend on model selection, settings, automated graders, live website state, and task selection. Treat the benchmark repository as a starting point for evaluation, not as proof that every browser task will work.

Aside Privacy and Data Handling

The homepage says tasks, memory, and data stay on your device and that nothing is shared with LLM providers. The privacy policy adds the missing detail: Aside is local-first, but account, billing, analytics, support, feedback, email, referral, password manager sync, browser sync, recovery backup, and hosted AI model features use server-side services.

For AI tasks, Aside's privacy policy says model-visible context can include prompts, instructions, tool results, selected browser snapshots, selected screenshots, selected files, and assistant responses. Task transcripts and artifacts are stored locally. Hosted AI providers receive the context needed to run the task, and a third-party model or API key can bring that provider's terms into the picture.

That is not a dealbreaker. It is the reality of a browser agent that can work across real websites. But it changes the question from "is this local?" to "which parts are local, which parts sync, which parts are sent to a model, and which websites receive data during the run?" I would answer those questions before connecting accounts that contain customer data, financial records, health data, unreleased code, or legal documents.

  1. Start with a throwaway browser profile or low-risk accounts.
  2. Turn on only the model provider you intend to test.
  3. Keep memory and sync settings visible during the first runs.
  4. Use agent password access on one account before trusting it with several.
  5. Review task transcripts, screenshots, artifacts, and credential access logs after each sensitive run.
  6. Do not approve payments, posts, messages, or account changes until the agent's plan is understandable.

The best privacy feature here may be human confirmation at the edge. If a browser agent can click buttons inside logged-in accounts, sensitive actions need a final checkpoint. Aside says payments, posts, and messages wait for confirmation. That should be tested before relying on it, but it is the right design direction.

Aside Pricing and Plans

Plan Price listed on June 25, 2026 Included notes from Aside
Free $0 forever 500 credits per month, up to 3 routines, built-in password manager, personalized memory.
Pro $20 per month Everything in Free, 3x more credits, Ultrabrowse for deep research, unlimited routines, cloud handoff, 14-day free trial.
Max $200 per month Everything in Pro, 10x more credits than Pro, early access program, 14-day free trial.
Enterprise Contact sales Team seats, shared profiles, vault access, app or cron starts, team logins, rollout help, and production support.

The pricing page does not make credits self-explanatory in the scraped text, so do not compare plans only by price. For agent browsers, the unit that matters is how much a real task consumes. Deep research, long runs, retries, screenshots, model calls, and cloud handoff can make one task much more expensive than a simple page summary.

Bring-your-own subscriptions also need careful math. If Aside uses your Claude or ChatGPT subscription for parts of the workflow, the practical cost includes Aside's plan, model provider limits, API key billing if used, and the time spent reviewing actions. A $20 browser plan can be cheap if it finishes boring browser work. It is expensive if it mostly asks for confirmation and stalls.

Who Should Try Aside?

  • Good fit: users who already live in web apps and want agents to operate Gmail, Notion, dashboards, spreadsheets, forms, and internal tools.
  • Good fit: teams evaluating browser-native automation before building custom integrations.
  • Good fit: AI power users who already pay for Claude or ChatGPT and want that model access inside a browser workflow.
  • Test carefully: users who want the agent to touch passwords, payment flows, posting tools, support queues, hiring data, or customer accounts.
  • May not fit: users who want a simple chatbot, a normal Chrome extension, or a browser that never sends selected context to hosted model providers.
  • May not fit: Linux users if the current download path does not support their platform. One launch-post reply specifically asked about Ubuntu support.

The most interesting part of Aside is not the glassy interface or the launch wording. It is the product bet: the browser is already where work happens, so the agent should live there too. I think that bet is directionally right. I also think the permissions model has to be boring, visible, and conservative. Browser agents do not fail like chatbots. They fail by clicking, sending, buying, deleting, or exposing context in places where a normal assistant would only write text.

Frequently Asked Questions

What is Aside AI browser?
Aside is an AI browser from Aside Computer Inc. that runs browser-agent tasks across logged-in websites, browser history, tabs, files, memory, and account context. It is backed by Y Combinator's Fall 2025 batch.

Does Aside run everything locally?
No. Aside is local-first and stores task transcripts and artifacts locally, but its privacy policy says account, billing, analytics, support, sync, recovery, and hosted AI model features can use server-side services. Hosted model providers receive model-visible context needed for selected tasks.

Can Aside use Claude or ChatGPT?
Yes. Aside's site says users can bring Claude or ChatGPT subscriptions or their own API key. The selected provider affects cost, limits, and data handling.

Is the Claude Fable benchmark claim verified?
The launch post says Aside outperforms Claude Fable. The public pages I reviewed show detailed Online-Mind2Web, Odysseys, and BU Bench V1 results, and the official site chart names Claude Opus 4.8. I did not find a Claude Fable row in the public benchmark README.

Is Aside free?
Aside lists a Free plan with 500 credits per month. Paid plans are Pro at $20 per month and Max at $200 per month, with Enterprise pricing handled by sales.

Sources