Recommendation for Computer use
Computer & Browser Use
Our top recommendation for Computer & Browser Use, based on the public evidence we track, is Z.ai: GLM 5.2.[1][2][3][4] Use for code-heavy agent tasks where you can feed text context from other tools, as it handled IDA reverse engineering workflows effectively when given pre-processed information. Watch out: Avoid for any task requiring direct image input; the model receives empty results when given screenshots and either hallucinates content or enters retry loops, as seen with 66-minute wasted runs on document translation. OpenAI: GPT-5.6 Sol is the next-ranked alternative. Use for computer use and 3D manipulation tasks where it shows particular strength, including proactive binary reverse engineering and APK analysis.
About this recommendation
- Updated
- Sep 25, 2026
- Evidence through
- Sep 25, 2026
- Sources
- 20
- Revision
- v79
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
20
live candidates
Evaluation feeds
4
task-weighted
Winner coverage
60%
intended feed weight
Largest provider share
1 of 5
Anthropic
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| OSWorld Verified | 65% | not measured | 3/20 |
| LMArena Agent | 15% | #11 | 16/20 |
| Berkeley Function Calling | 10% | not measured | 3/20 |
| OpenRouter usage | 10% | 97/100 | 20/20 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic1 model
- OpenAI1 model
- Qwen1 model
- xAI1 model
- Z.ai1 model
Decision table
Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.
| Rank | Model | Relative score | Coverage | Practitioner evidence | Strongest measured reason |
|---|---|---|---|---|---|
| 01 | GLM 5.2Z.ai | 70 | 60% | 7 threads · 5 families · 1 cautions | #11 LMArena Agent |
| 02 | GPT-5.6 SolOpenAI | 69 | 84% | 9 threads · 7 families · 7 cautions | #2 OSWorld Verified · #4 LMArena Agent |
| 03 | Claude Fable 5Anthropic | 62 | 60% | 2 threads · 2 families · 1 cautions | #2 LMArena Agent |
| 04 | Qwen3.8 Max (0902)Qwen | 59 | 60% | 2 threads · 1 families · 1 cautions | #13 LMArena Agent |
| 05 | Grok 4.6xAI | 57 | 60% | 3 threads · 3 families · 2 cautions | #15 LMArena Agent |
Relative score combines normalized benchmark quality and signal coverage; independent practitioner evidence and freshness are bounded tie-breakers. It is an ordering score, not an absolute quality percentage. The writing model receives this order and cannot change it.
GLM 5.2 functions as a capable text and code model for agent workflows but lacks native vision support, requiring pairing with multimodal models for screenshot-based computer use.
Best when: Use for code-heavy agent tasks where you can feed text context from other tools, as it handled IDA reverse engineering workflows effectively when given pre-processed information.
Tips
- Use for code-heavy agent tasks where you can feed text context from other tools, as it handled IDA reverse engineering workflows effectively when given pre-processed information.
- Deploy in multi-model agent setups where Kimi or another vision-capable model handles screenshots while GLM 5.2 handles reasoning and code generation.
Watch out for
- Avoid for any task requiring direct image input; the model receives empty results when given screenshots and either hallucinates content or enters retry loops, as seen with 66-minute wasted runs on document translation.
- Watch for routing limitations in Codex CLI where vision sidecars default to OpenAI or Anthropic backends, forcing fallback routes when GLM 5.2 is the primary model.
GPT-5.6 Sol demonstrates strong computer use capabilities with benchmark validation but exhibits reliability issues in long-running agent sessions and safety check false positives.
Best when: Use for computer use and 3D manipulation tasks where it shows particular strength, including proactive binary reverse engineering and APK analysis.
Tips
- Use for computer use and 3D manipulation tasks where it shows particular strength, including proactive binary reverse engineering and APK analysis.
- Expect solid benchmark performance for desktop automation, ranking #4 on LMArena's agentic arena and achieving 28.1% accuracy on OSWorld Verified.
Watch out for
- Watch for premature stopping, scope drift, and dangerous workaround attempts when the model gets stuck in multi-step tasks.
- Avoid for long-running agent sessions with automatic context compaction, as execution fails to reliably resume after compaction events.
- Expect HTTP 400 errors with Subagents V2 enabled in Codex Security Deep Scan workflows, and verify zsh snapshot compatibility if using KSH_OPTION_PRINT.
Claude Fable 5 ranks highly for agentic tool use and demonstrates spatial reasoning in 3D coordinate tasks, though environment variable injection affects daemon workflows.
Best when: Use for multi-step tool use and agentic tasks where it ranks #2 on LMArena's agentic arena with strong human preference scores.
Tips
- Use for multi-step tool use and agentic tasks where it ranks #2 on LMArena's agentic arena with strong human preference scores.
- Deploy for 3D spatial reasoning tasks requiring precise coordinate manipulation, as shown in brush tool positioning for aircraft rendering.
Watch out for
- Watch for environment variable pollution when restarting daemons from Claude Code sessions; NO_COLOR=1 and AI_AGENT=... propagate to child processes and may break color-dependent or identity-sensitive applications.
Qwen3.8 Max participates in security audits and agentic benchmarks but shows middling performance on multi-step task evaluations.
Best when: Use for security-focused code analysis where it has identified secret exposure and cleartext transport vulnerabilities in Jenkins sandbox implementations.
Tips
- Use for security-focused code analysis where it has identified secret exposure and cleartext transport vulnerabilities in Jenkins sandbox implementations.
Watch out for
- Expect weaker performance on general agentic tasks, ranking #12 on LMArena's agentic arena with a score of 3.3, below top-tier alternatives.
Grok 4.6 supports computer use and browser automation in cloud environments but suffers from context overflow loops and local execution failures in Codex CLI integrations.
Best when: Use for cloud-based agent workflows with spawn, computer-use, and GitHub MCP capabilities in controlled VM environments.
Tips
- Use for cloud-based agent workflows with spawn, computer-use, and GitHub MCP capabilities in controlled VM environments.
Watch out for
- Avoid for screenshot-heavy sessions; reading multiple PNGs can trigger irreversible context_overflow loops with failed automatic compaction, requiring session restart.
- Do not use with Codex CLI local exec tools; third-party routing causes immediate aborts before shell entry, with models retrying failed calls and wasting tokens.
Frequently asked
- What is the top-ranked model for Computer & Browser Use?
- Z.ai: GLM 5.2 ranks first in the current evidence-weighted comparison. Use for code-heavy agent tasks where you can feed text context from other tools, as it handled IDA reverse engineering workflows effectively when given pre-processed information.[1]
- What should I watch out for with Z.ai: GLM 5.2?
- Avoid for any task requiring direct image input; the model receives empty results when given screenshots and either hallucinates content or enters retry loops, as seen with 66-minute wasted runs on document translation.[2]
Sources
- 1
“I have tested this exact scenario, and it works. Opus 5 had access to IDA over MCP, and I simply asked it HOW certain things were done in the target binary. Purely informational, educational, discovery, it was very helpful creating context documents. Then I took those over to GLM-5.2 to actually accomplish something.”
gfosco · Hacker News · Jul 28, 2026 - 2
“**提交人**: 盖世大宝 **客户端版本**: 0.1.20 --- ## 现象 在 claude-code agent 下使用纯文本模型（z-ai/glm-5.2，官方仅文本/代码模态，走第三方 Anthropic 兼容直通路由）时，Read 工具读取图片后，模型表现为收到空结果：无任何报错、无任何“图像无法送达”提示，模型要么反复重读，要么直接臆造图片内容。实测单任务浪费 66 分钟，并产出包含错误金额与人名的法律文件译文。 ## 复现步骤 1. 配置 claude-code agent，模型路由为 z-ai/glm-5.2（api-key-header 分支的 Anthropic 兼容直通，经 anthropic-compat-proxy 字节透传至 api.z.ai/api/anthropic 等 Anthropic 兼容端点）。 2. 让模型处理一张本地图片（如 Read 一张扫描件）。 3. 观察：模型不报错，行为如同工具返回了空内容。 ## 期望行为 图像无法送达时应被显式阻断或降级为说明性占位文本（“图像无法送达本路由，禁止臆测”），模型应立即告知用户，而非静默继续…”
cindy-issue[bot] · GitHub · Jul 28, 2026 - 3
“Astra fails in similar ways, and at similar frequency, as GPT 5.6 Sol does. It often goes way out of scope, or just stops prematurely, or tries to find odd and even dangerous workarounds when it gets stuck. It's phenomenal at computer use and 3D stuff. I've been using it less and less for coding.”
enraged_camel · Hacker News · Sep 21, 2026 - 4
“5.6 Sol has happily reverse engineered and decompiled binaries for me. Heck it has proactively asked me if I wanted it to tear apart APKs that remote control some HW I have.”
com2kid · Hacker News · Sep 21, 2026 - 5
“Yes, you are right (as far as I'm aware). For things where you need the LLM to look at screenshots, photos or other images you can use Kimi-K2.6 K2.7 - comparable pricing, somewhat comparable performance and quality. You can even probably combine two models (e.g Kimi and GLM) in one agent, using Kimi for multimodal inputs and GLM for everything else, although 1) I'm not sure if this will not cause some kind of context poisoning with low-quality patterns for better performing model (e.g. in some…”
dryarzeg · Hacker News · Jun 17, 2026 - 6
“**English translation** *(original: ko)* ## Client or integration Codex CLI (opencode-go session, `glm-5.2 max` routing) ## Area Proxy and routing (directly related to Provider adapter area) ## Summary The `backend` of the web search sidecar and vision sidecar is fixed to only two values: `openai` / `anthropic`. Therefore, when the **main routing model is a provider other than OpenAI·Anthropic (e.g., z.ai GLM-5.2)**, the turns requiring search/vision must necessarily be executed through the Ope…”
github-actions[bot] · GitHub · Jul 24, 2026 - 7
“Ranks #4 of 36 on LMArena's agentic arena (score 7.1), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Sep 15, 2026 - 8
“Achieves 28.1% binary accuracy on OSWorld Verified (#2 of 8) with a 500-step budget, measuring computer use in real desktop environments.”
OSWorld Verified · Benchmark · Jun 24, 2026 - 9
“### 📦 Environment - LobeHub Desktop: 2.2.13 (stable) - Deployment: self-hosted `lobehub/lobehub:latest` - OS: macOS 26.5.1, arm64 - Model: GPT-5.6-Sol with function calling - Task type: long-running, multi-step server operation using local shell/MCP tools ### 🐛 Description When the conversation reaches automatic context compaction while an agent task is still unfinished, the compacted context correctly records pending actions and even an outstanding shell session, but execution does not relia…”
brlanweb · GitHub · Aug 6, 2026 - 10
“## Environment - CCSwitchMulti: `3.16.5-22` - Codex CLI: `0.146.0` - ChatGPT/Codex Desktop: `26.727.40816` - macOS: `26.5.2 arm64` - Route: Codex MultiRouter -> OpenAI Official - Models: `gpt-5.6-sol`, `gpt-5.6-terra` - Endpoint: `/responses` ## Problem Normal Codex conversations using the same route and models return HTTP 200 successfully. After enabling native Subagents V2 and starting a Codex Security Deep Scan, requests consistently fail with HTTP 400: The complete error shown by CCSwitchMu…”
sbh923 · GitHub · Jul 31, 2026 - 11
“### What version of Codex CLI is running? 0.146.0 ### What subscription do you have? ChatGPT Pro ### Which model were you using? gpt-5.6-sol ### What platform is your computer? Linux 7.1.5-zen1-2-zen x86_64 unknown ### What terminal emulator and version are you using (if applicable)? tmux 3.7b; zsh 5.9.2 (x86_64-pc-linux-gnu) ### Codex doctor report ### What issue are you seeing? When Codex creates a shell snapshot for zsh, snapshot validation fails if the zsh option KSH_OPTION_PRINT is enabled…”
Aloxaf · GitHub · Aug 1, 2026 - 12
“Ranks #2 of 36 on LMArena's agentic arena (score 8.8), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Sep 15, 2026 - 13
“how about trying to draw an airbus a320 in 3d space using only one brush tool that can be moved to specific x,y,z coordinates (and its color, size & hardness can be changed). i think fable 5.1 did quite a good job (reasoning high, cost $0,261): https: files.catbox.moe umx102.png for comparision, this is fable 5: https: files.catbox.moe ihl4m1.png”
leumon · Hacker News · Sep 1, 2026 - 14
“## Environment - Claude Code v2.1.220 (npm global install), Windows 11 Pro, Windows Terminal + pwsh 7.6.4 - Model: claude-fable-5 ## Summary Claude Code injects `NO_COLOR=1` (and `GIT_TERMINAL_PROMPT=0`, `AI_AGENT=...`) into the environment of Bash/PowerShell tool subprocesses. That is reasonable for keeping tool output clean, but the variable is inherited by *anything* those subprocesses launch. If a user restarts a long-lived daemon from inside a Claude Code session (a very common workflow: "…”
wlstjdxo0 · GitHub · Jul 31, 2026 - 15
“## Latest finding (multi-model security audit) - **Inbound agent JNLP secret exposed in sandbox process command line and via the exec API** (_confidence: Medium; models: deepseek/deepseek-v4-pro:max, qwen/qwen3.8-max-0902, xiaomi/mimo-v2.6-pro) - Category: Secret Handling / Information Exposure (CWE-214, CWE-522) - Location: src/main/java/sh/createos/jenkins/sandbox/CreateOSLauncher.java:248-262 (`String secret = computer.getJnlpMac()` interpolated into agentCommand as `-secret `, then handed t…”
pratikbin · GitHub · Sep 24, 2026 - 16
“## Latest finding (multi-model security audit) - **Plain-HTTP CreateOS API endpoint accepted; API key and tunnel traffic sent in cleartext** (_confidence: Medium; models: deepseek/deepseek-v4-pro:max, inclusionai/ling-3.0-flash, openrouter/z-ai/glm-5.3-flash:xhigh, qwen/qwen3.8-max-0902, x-ai/grok-4.7, xiaomi/mimo-v2.6-pro) - Category: Cryptography & Transport Security (CWE-319) - Location: src/main/java/sh/createos/jenkins/sandbox/CreateOSCloud.java:419-429 (doCheckApiUrl accepts `http://` and…”
pratikbin · GitHub · Sep 24, 2026 - 17
“Ranks #12 of 36 on LMArena's agentic arena (score 3.3), measuring tool-use and multi-step task performance from human preference.”
LMArena agentic arena · Benchmark · Sep 15, 2026 - 18
“from: SPUR to: OFFER id: spur-offer-cursor-cloud-20260820-01 subject: harness --- PLAIN: Cursor Cloud Grok 4.6 — spawn, computer-use, git-into-env, GitHub MCP, /goal, Agent Skills. Not the laptop. HARNESS: Cursor Cloud Agent · Grok 4.6 WHERE: cloud VM — not the owner's PC CAN: spawn subagents (explore, computerUse, generalPurpose); computer-use/browser; pull this repo into a fresh env and branch; GitHub issues/contents/PRs/gh; ntfy write roads; Agent Skills in .agents/skills; Cursor /goal CANNO…”
cursor[bot] · GitHub · Aug 20, 2026 - 19
“## What happened A Desktop session on `grok-4.6` (`xai-oauth`) entered a loop it cannot leave: every turn fails with `context_overflow` ("上下文窗口已超出限制") plus the recovery hint "没有执行工具，可直接重试", and automatic history compaction reports "Context summary failed" and keeps failing open. Retrying reproduces it exactly — the prompt is byte-identical on every attempt, so "retry directly" can never succeed. The session had read three PNG screenshots with `Read`. From the captured provider request for the f…”
Astro-Han · GitHub · Aug 21, 2026 - 20
“**提交人**: CC **客户端版本**: 0.1.46 --- ## 现象 Cindy 内置 Codex agent（cli 0.145.0）接 Grok / DeepSeek 等第三方模型时，本地 `exec` 在进入 shell 之前就失败。界面表现为工具被 abort；模型不断重试同一次调用，浪费 token，最终仍跑不成任何本地命令。网络类工具（如 web_search）可以成功。MCP / 会话历史等能力也挂在 `exec` 里，一并坏掉。 ## 复现步骤 1. 新建会话，Agent 选 Codex，模型选 Grok 4.6 或 DeepSeek-V4-Flash（经 cliproxy / cindy_gateway）。 2. 让 agent 执行任何本地操作（例如查 Git 状态、读项目文件、跑一条 shell）。 3. 观察 `exec` 是否立刻 abort，以及模型是否空转重试。 ## 期望行为 - 第三方模型能正常调用本地命令 / 读写文件；或 - payload 不符合时返回可纠正的 schema 错误，而不是 Fatal abort；并阻止同一回合无参重试…”
cindy-issue[bot] · GitHub · Aug 14, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.