What AI can now do·Jul 17, 2025·Agents

OpenAI treats ChatGPT Agent as 'High' biological-risk capability

OpenAI

41.6Humanity's Last Exam (pass@1)

ChatGPT Agent combined browsing, a terminal and connectors to act for users, scoring 41.6 on Humanity's Last Exam and 45.5% on SpreadsheetBench (vs 20.0% for Copilot in Excel); OpenAI treated it as 'High' capability for biological and chemical risk and shipped matching safeguards.

How far it has travelled

From a demonstration to real life, in five steps. Empty steps are the point: a demonstration is not a deployment.

Demonstratedshown once, by its developer or a researcher
yes · Jul 17, 2025 · Visual and text browsers, a terminal and app connectors; HLE 41.6 (44.4 with parallel attempts), FrontierMath 27.4% with tools, BrowseComp 68.9%, SpreadsheetBench 45.5%. source
Reproducedshown again by someone else
yes · Feb 5, 2026 · Comparable deep-research and agent modes followed across labs; Opus 4.6 reported HLE 53.0% with tools and BrowseComp 86.8% (Feb 2026). source
Accessibleavailable to the public or to customers
yes · Jul 17, 2025 · Rolled out to Pro, Plus and Team users (400 or 40 messages a month); not available in the EEA or Switzerland at launch. source
Reliableworks most of the time on a defined task
partial · Jul 17, 2025 · A capable general-purpose agent, but reliability over long tasks remained the open question METR's horizon work tracks; OpenAI applied always-on classifiers and reasoning monitors. source
Real consequenceshas changed something in the world, for good or ill
yes · Jul 17, 2025 · First consumer agent OpenAI classified 'High' for biological and chemical uplift (able to meaningfully help a novice), launched with dual-use refusal training, always-on classifiers and a bug bounty; GPT-5 followed at 'High' bio in August. source

What this could change for you

A mainstream assistant can now carry out multi-step online tasks for you, and its maker judged it capable enough in biology to need weapons-misuse safeguards.

Incidents it made possible

Sources

primary source.

A milestone is a dated, sourced result that moved what AI can do; a stage is upgraded only when a source shows it (a reproduction, a public release, production evidence, a documented consequence). “Unknown” is the honest default.