Best AI for Web Browsing and Computer Use - July 2026

Seven models now exceed the 72.4% human expert baseline on OSWorld-Verified. Four months ago, only one did. Claude Fable 5 sits at the top with 85.0%, but it spent 19 days offline under a US government export-control directive before access was restored today. For teams that need stable production pipelines, Claude Opus 4.8 at 83.4% is the safer anchor. And if your cost model matters - it almost always does - Claude Sonnet 5 clears the human baseline at 81.2% for $2 per million input tokens under introductory pricing.

Web Browsing vs Computer Use

These terms still get conflated constantly, and the distinction matters when choosing a model.

Web browsing agents operate inside a browser. They read DOM elements, click links, fill forms, and navigate pages. The primary benchmark is WebArena, which runs 812 tasks across shopping, forum, GitLab, CMS, and map environments. Note that WebArena scores reflect complete agent systems - model plus scaffolding, tools, and retry policy - not raw model performance.

Computer use agents control an entire desktop via screenshots, mouse movements, and keyboard input. They can interact with any application. OSWorld-Verified is the standard benchmark: 369 tasks across real operating systems, scored independently where possible. The human expert baseline is 72.4%.

Some products blend both. OpenAI's Operator (now ChatGPT Agent) and Google's Project Mariner operate inside Chrome but also coordinate with desktop apps. The benchmark categories still measure distinct skills, though.

Rankings: Computer Use (Desktop)

Human expert baseline: 72.4%

Source: BenchLM.ai OSWorld-Verified Leaderboard, as of July 1, 2026. All scores are self-reported by providers unless independently verified.

Desktop computer use agents take screenshots and control mouse/keyboard to complete real tasks across operating systems. Source: unsplash.com

Rankings: Web Browsing

WebArena scores reflect full agent systems (model plus tools and scaffolding), not isolated model performance. Treat them as benchmarks of the complete stack.

Source: Steel.dev WebArena Leaderboard, last updated May 27, 2026.

Browser Use (open-source library) reports 80.0% on WebVoyager using Claude Fable 5 as the backbone - but WebVoyager is a different benchmark than WebArena, and the scores aren't directly comparable.

Detailed Analysis

Claude Fable 5 - The New Leader, With a Caveat

Claude Fable 5 scores 85.0% on OSWorld-Verified, a 10-point jump from GPT-5.4's 75.0% that was the category leader as recently as March 2026. On Browser Use's WebVoyager benchmark, it posts 80.0% - 12 points above the next frontier model.

The caveat is significant: Anthropic disabled Fable 5 and its companion Claude Mythos 5 on June 12 following a US government export-control directive, citing a potential jailbreaking method for sensitive use cases. Access was restored globally today, July 1, after the Department of Commerce lifted the controls. For teams that can't tolerate unplanned model outages, this event is a real production risk to weigh. Benchmark scores are unaffected - they were measured before the suspension - but reliability history matters for enterprise planning.

Pricing sits at $10/$50 per million input/output tokens, matching Mythos 5. That's twice Opus 4.8's input cost for roughly 1.6 points of OSWorld improvement.

Claude Opus 4.8 - The Production-Grade Choice

Claude Opus 4.8 scores 83.4% on OSWorld-Verified, 10.6 points above Claude's prior leader Claude Sonnet 4.6 from the March 2026 ranking. It launched May 28 with no access disruptions and sits at $5/$25 per million tokens.

For production computer use deployments, Opus 4.8 is the current practical ceiling. It beats GPT-5.5 by 4.7 points at the same input cost - $5 per million tokens - while GPT-5.5 charges $30 per million on output versus Opus 4.8's $25. The 1.6-point gap to Fable 5 is likely undetectable in most real workflows.

Anthropic's Computer Use Leaderboard shows the progression clearly: 14.9% (Sonnet 3.5) to 72.5% (Sonnet 4.6) to 83.4% (Opus 4.8) across roughly 20 months.

Holo3-35B-A3B - The Open-Source Surprise

H Company released Holo3 on March 31, 2026, and nobody saw it coming. The Holo3-35B-A3B variant runs on just 3 billion active parameters from its 35 billion total in a sparse MoE architecture, built on Qwen3.5 with additional computer-use training. It scores 82.6% on OSWorld-Verified - above GPT-5.5 and Gemini 3.5 Flash - and ships under Apache 2.0 with weights on HuggingFace.

The catch is real-world reliability. One independent review found that web form filling and structured spreadsheet tasks worked consistently, but multi-application workflows and anything triggering unexpected dialog boxes broke frequently. The 82.6% score means roughly 1 in 6 tasks fails. For research prototypes and internal tools, it's compelling. For customer-facing automation, the failure modes need closer evaluation.

H Company subsequently shipped Holo3.1 with function-calling protocol support and smaller model sizes for local deployment. The 122B closed variant (Holo3-122B-A10B) scores 78.8% via H Company's API.

Claude Sonnet 5 - Best Value Right Now

Claude Sonnet 5 launched June 30 with introductory pricing of $2 per million input tokens and $10 per million output tokens - active through August 31, 2026, after which it reverts to $3/$15. Its 81.2% OSWorld-Verified score is 8.8 points above the human baseline.

The tokenizer changed in Sonnet 5, increasing token counts by 1.0-1.35x depending on content type. Anthropic set the intro pricing so total cost stays roughly neutral versus Sonnet 4.6 for typical workloads. For new projects that aren't bound to the Sonnet 4.6 API, Sonnet 5 is the obvious starting point.

Gemini 3.5 Flash - Budget Native Computer Use

Gemini 3.5 Flash added native computer use in June 2026, pairing screen control with built-in Google Search and Maps integration. Source: unsplash.com

Google baked computer use natively into Gemini 3.5 Flash on June 24, 2026. The integration is architectural: one model handles browser, mobile, and desktop control with no model-hopping, and it pairs computer use with built-in Google Search grounding and Maps access in the same call.

OSWorld-Verified: 78.4%, 0.3 points behind GPT-5.5. Pricing: $1.50/$9 per million tokens - 33% of GPT-5.5's input cost. For budget-sensitive deployments that don't need the top few percentage points of accuracy, this is the most defensible cost trade.

The context window is 1,048,576 input tokens, a 8x expansion over Gemini 2.5's 128K. That matters for computer use agents that need to maintain long interaction histories.

Methodology

Rankings use OSWorld-Verified as the primary benchmark for desktop computer use. The benchmark runs 369 tasks across real operating systems using natural-language instructions and screenshot-based observation. Human expert baseline is 72.4%.

A persistent limitation: most OSWorld-Verified submissions are self-reported by model providers. Independent verification by the xlang.ai team lags behind provider claims by weeks or months. When two providers are within 2-3 points of each other, treat the difference as within measurement noise until independently copied.

WebArena scores reflect complete agent systems, not base model capabilities. Prompting strategy, browser action space, step budget, and retry policy can shift WebArena scores by 10-15 percentage points independently of the underlying model.

ScreenSpot-Pro - which tests GUI grounding accuracy across 1,581 annotated screenshots from 23 professional applications - shows that even top models score under 70% on precise click targeting. Task completion benchmarks can mask this underlying weakness; models often succeed through retries rather than precision.

See the web agent benchmarks leaderboard for full WebArena standings and agentic AI benchmarks for GAIA and related evaluations.

Historical Progression

• April 2024 - OSWorld benchmark launches. GPT-4 with computer use scores around 12%. The 72.4% human baseline feels unreachable.

• October 2024 - Anthropic releases Computer Use API with Claude 3.5 Sonnet at 14.9% on OSWorld. First major commercial computer use product.

• January 2025 - OpenAI launches Operator for ChatGPT Pro subscribers. First mass-market computer use product.

• Late 2025 - Claude Sonnet 4.5 reaches 61.4%. Claude Sonnet 4.6 hits 72.5% - the first model to match the human baseline.

• February-March 2026 - Claude Sonnet 4.6 confirmed at 72.5% on OSWorld-Verified. GPT-5.4 claims 75.0%, the first model to beat the human baseline.

• March 31, 2026 - H Company releases Holo3 open-source with 35B and 122B variants. Benchmarks put both above GPT-5.4.

• April 23, 2026 - OpenAI releases GPT-5.5, its first full base-model retrain since GPT-4.5. OSWorld-Verified: 78.7%.

• May 28, 2026 - Anthropic releases Claude Opus 4.8 at 83.4% on OSWorld-Verified. Holo3-35B-A3B is the only open-source model within 1 point.

• June 9, 2026 - Claude Fable 5 and Claude Mythos 5 launch at 85.0% on OSWorld-Verified. Three days of peak adoption.

• June 12, 2026 - US government export-control directive disables Fable 5 and Mythos 5 for all customers worldwide. Teams scramble back to Opus 4.8.

• June 24, 2026 - Gemini 3.5 Flash ships with native computer use integration: browser, mobile, and desktop in one model at $1.50/$9 per million tokens.

• June 30, 2026 - Claude Sonnet 5 releases at $2/M input (intro). US Department of Commerce lifts Fable 5/Mythos 5 export controls.

• July 1, 2026 - Fable 5 and Mythos 5 restored globally. Seven models now beat the 72.4% human expert baseline.

Practical Guidance

For desktop automation (native apps, spreadsheets, legacy software): Claude Opus 4.8 is the stable production choice at $5/$25 per million tokens and 83.4% OSWorld. Fable 5 scores higher but just returned from a 19-day outage - if uptime guarantees matter, Opus 4.8 is the safer default until Fable 5 accumulates reliability history post-restoration.

For budget-sensitive desktop automation: Gemini 3.5 Flash at $1.50/$9 posts 78.4% on OSWorld - 6 points below Opus 4.8 at 30% of the cost. For most non-critical automation tasks, that tradeoff is worth checking before committing to Anthropic's pricing.

For new agentic projects: Claude Sonnet 5 at $2/$10 (intro pricing through August 31) offers 81.2% OSWorld performance above the human baseline. Start here unless you have data showing a specific requirement for Opus-level accuracy.

For web browsing and form automation: The best AI browser automation tools guide covers the framework layer in detail. Browser Use (93K+ GitHub stars, MIT license) with Claude Fable 5 backbone now posts 80.0% on WebVoyager. For TypeScript environments, Playwright MCP remains free and faster via structured accessibility snapshots rather than screenshots.

For self-hosted deployments: Holo3-35B-A3B is the current open-weight answer at 82.6% under Apache 2.0. The Holo3.1 update added function-calling support. Qwen3.7 Plus at 73.3% is the second option if Holo3's memory requirements are too high for your hardware.

FAQ

Which AI model leads computer use benchmarks right now?

Claude Fable 5 leads OSWorld-Verified at 85.0% as of July 1, 2026. Claude Opus 4.8 is the runner-up at 83.4% and the more stable production option given Fable 5's recent access suspension.

What is the best open-source model for computer use?

Holo3-35B-A3B by H Company scores 82.6% on OSWorld-Verified under Apache 2.0. It uses only 3B active parameters in a sparse MoE design and is available on HuggingFace. Qwen3.7 Plus is the second-best open option at 73.3%.

What's the cheapest model that can beat the human baseline?

Qwen3.7 Plus scores 73.3% at $0.32 per million input tokens. Gemini 3.5 Flash scores 78.4% at $1.50 per million tokens and includes native computer use without extra tooling setup.

Is Claude or GPT better for computer use in 2026?

Claude leads at every price point. Claude Sonnet 5 (81.2%) beats GPT-5.5 (78.7%) at lower cost. Claude Opus 4.8 (83.4%) ties GPT-5.5's $5/M input price but scores 4.7 points higher. Claude Fable 5 (85.0%) leads the full leaderboard.

How often do computer use rankings change?

Fast. The top model has changed every 30-60 days since early 2026. Check the lastVerified date on this page and the computer use leaderboard for current standings.

Can AI computer use agents replace human workers?

Not yet, broadly. Even at 85%, roughly 1 in 6 tasks fails. The more realistic current use is semi-supervised automation - AI handles routine steps, humans review exceptions. ScreenSpot-Pro scores under 70% across frontier models expose click-precision gaps that compound in multi-step workflows.

Sources:

• BenchLM.ai OSWorld-Verified Leaderboard
• OSWorld Benchmark
• Steel.dev WebArena Leaderboard
• Anthropic statement on Fable 5/Mythos 5 suspension and restoration
• Claude Opus 4.8 Benchmarks - Vellum
• Claude Sonnet 5 Launch - TechCrunch
• H Company Holo3
• Holo3-35B-A3B HuggingFace
• Google Gemini 3.5 Flash Computer Use
• GPT-5.5 Guide - O-mega.ai
• Browser Use Benchmark
• ScreenSpot-Pro GUI Grounding Benchmark
• Kimi K2.6 Benchmarks - Artificial Analysis
• Gemini 3.5 Flash Benchmarks - LLM Stats