ChatGPT Work
- Category
- Browser/web agent
- Type
- commercial
- Scope
- Agent and cloud-browser tasks
- Source
- Vendor release notes
- Last reviewed
- 2026-10-18 (what this means)
Citation check incomplete — recheck later.
Summary
ChatGPT Work is OpenAI's current agent for multi-step tasks, including a cloud browser that can act on websites. It replaced ChatGPT agent, which itself replaced Operator; both earlier products are retired. Independent tests exist for Operator and ChatGPT agent but are not counted. Two independent tests of ChatGPT Work were found (a hands-on review and an integration vendor's benchmark), fewer than the three required, so no overall score is given. Insufficient evidence. Informational only; not a guarantee, endorsement, or certification.
Status: Insufficient evidence — no overall score is given (fewer than three independent performance sources). Independent performance sources: 2 for ChatGPT Work (3 required).
Factor notes (no scores)
| Factor | Weight | Score | Note |
|---|---|---|---|
| Completion rate | 30% | — | Two independent tests of ChatGPT Work were found (fewer than the 3 required): AI Agents Library, 10 business tasks, mean 9.15/10, one task (publishing a webpage) not completed (Aug–Sep 2026); Composio Golden Eval, 45 of 47 business-app scenarios passed (July 2026). Earlier products, for reference only (not counted): Operator 61.3% (Online-Mind2Web), 68.6% (Emergence WebVoyager), 10% full success (Mind2Web 2); ChatGPT agent 65.8% (BearCubs). |
| Disputes | 20% | — | No lawsuit, regulatory action or CVE specific to Operator, ChatGPT agent or ChatGPT Work was found. Researchers demonstrated prompt-injection data leaks (Operator) and a CAPTCHA-policy bypass (ChatGPT agent); both were reported to OpenAI. |
| Adherence to limits | 25% | — | The vendor documents site-access settings, confirmations before consequential actions and takeover for sign-ins; not independently tested for ChatGPT Work. |
| Evidence quality | 25% | — | Two independent performance sources for the current product (one hands-on review, one integration-vendor benchmark), neither peer-reviewed. Five tests of the retired predecessors are listed but not counted. |
See the methodology for the insufficient-evidence rule and the note on retired products.
Product timeline
- January 2025: Operator launched as a research preview.
- 17 July 2025: ChatGPT agent launched as “a natural evolution of Operator and deep research”; the Operator site was to be sunset within weeks and is no longer accessible.
- 9 July 2026: ChatGPT Work launched, including a cloud browser.
- By August 2026: ChatGPT agent retired. OpenAI’s help page says it “is no longer available” and points to ChatGPT Work; OpenAI support confirmed this on the community forum (12 August 2026). No exact retirement date was published.
Scope and overlap
This entry covers ChatGPT Work’s agent and cloud-browser task execution. Operator and ChatGPT agent are retired, so they are not given scored entries; their independent tests are listed below as historical context only and are not counted, because ChatGPT Work uses different models and a different setup. Deep research reports are covered by the separate OpenAI Deep Research entry.
Uncertainties
- OpenAI has published no exact retirement date for ChatGPT agent; user reports place it in early August 2026.
- ChatGPT Work's models changed three times within about three months (GPT-5.6 Sol, GPT-6 Sol, GPT-6.1 Sol).
- Vendor benchmark figures (system cards, BrowseComp, OSWorld 2.0) are not used as independent evidence.
- Citation check incomplete — recheck later (0052: Semantic Scholar unavailable; OpenAlex title screen only).
- Both counted tests are single-run, not peer-reviewed, and used GPT-5.6 Sol; the product has since moved to newer models.
- Composio ran the test through its own integration product; AI Agents Library used the author's own Skills in 2 of 10 tasks.
- Composio's page labels the OpenAI harness as ChatGPT Work in its setup table but as Codex CLI in its results table and one paragraph, so which OpenAI product ran the test is not fully clear.
Sources
Independent performance sources (counted)
- AI Agents Library hands-on test (Aug 25–Sep 3 2026, ChatGPT Plus, GPT-5.6 Sol where recorded): 10 business tasks scored on 5 criteria, mean 9.15/10; webpage publishing not completed; one accuracy error (inferred follow-ups); self-funded; 2 of 10 tasks used the author's own Skills — independent performance, accessed 2026-10-18
- Composio Golden Eval (20 July 2026): ChatGPT Work with GPT-5.6 Sol passed 45 of 47 business-app scenarios, one run each, via Composio MCP; published by an integration vendor whose product was used in the test — independent performance, accessed 2026-10-18
Historical context: tests of earlier products (not counted)
- Online-Mind2Web (COLM 2025): Operator 61.3% success on 300 live tasks, human-evaluated — historical context (earlier product, not counted), accessed 2026-10-17
- Mind2Web 2 (NeurIPS 2025 D&B; Amazon AGI co-authors): Operator 0.26 partial, 10% success — historical context (earlier product, not counted), accessed 2026-10-17
- BearCubs (preprint): ChatGPT agent 65.8%, Operator 23.4%, humans 84.7% — historical context (earlier product, not counted), accessed 2026-10-17
- Emergence WebVoyager (preprint 2026; competing authors): Operator 68.6% vs 87% vendor claim — historical context (earlier product, not counted), accessed 2026-10-17
- Dark Patterns Meet GUI Agents (preprint 2025): Operator highest dark-pattern avoidance of 6 agents, not complete — historical context (earlier product, not counted), accessed 2026-10-17
Other sources (not counted)
- Simon Willison (30 August 2026): Work Cloud has internet-enabled code execution, a full browser, persistent shared workspace, subagents and scheduled tasks; combines private data, untrusted content and an outbound channel (prompt-injection risk; no attack demonstrated) — other (hands-on documentation of capabilities, no performance test), accessed 2026-10-18
- Embrace The Red (Feb 2025): prompt injection made Operator leak PII; disclosed and mitigated — other (security test of earlier product), accessed 2026-10-17
- SPLX (Sep 2025; security vendor): staged prompt made ChatGPT agent solve CAPTCHAs against policy — other (security test of earlier product), accessed 2026-10-17
Vendor statements
- Vendor launch post (Operator merged into ChatGPT agent; Operator site sunset) — vendor or maintainer, accessed 2026-10-17
- Vendor help page: "ChatGPT agent is no longer available"; Operator site no longer accessible — vendor or maintainer, accessed 2026-10-17
Excluded sources
- OpenAI system-card and launch figures (prompt-injection resistance, BrowseComp, WebVoyager, WebArena, OSWorld 2.0): vendor-reported, not independent.
- SusBench: tested OpenAI’s API sample app, not the Operator product.
- Composio Golden Eval: interested-party test; its results table describes a Codex CLI run rather than ChatGPT Work.