Breaking

OpenAI Adds Codex Notes and Searchable History for GPT-6 Astra Long Tasks

September 5, 2026 at 15:16 EDT

Software Dev & Coding
Foundation Models
AI Agents

OpenAI is changing how GPT-6 Astra handles long Codex sessions: instead of repeatedly compacting the conversation into a single summary, the coding agent can keep notes across context windows and search earlier messages and tool outputs.

September 2026 · OpenAI Codex

Notes that survive a full context window

GPT-6 Astra can keep Codex notes across windows and search earlier messages and tool outputs, instead of crushing a long session into one summary.

OSWorld task time

Astra finished simulated computer-use tasks in about 40 minutes, versus about 75 minutes for GPT-5.6 Sol. Column height is proportional to minutes.

75 min

GPT-5.6 Sol

40 min

GPT-6 Astra

15 blocks vs 8 blocks · 75:40 minutes

Mind2Web completion speed

Astra plus the updated Codex harness completed the benchmark 1.9× faster than the current Sol experience. These figures measure the model and harness, not the notes experiment.

1×

Sol baseline

1.9×

Astra + harness

When a long Codex task overflows

Context fills

→

Notes persist

→

Search earlier windows

What developers like

Notes keep failed-test reasons, requirements, and tool results that compaction used to flatten into a vague recap.

What to watch

The same flag has been reported to make Luna or Terra lose the thread after compaction. The mode is still experimental.

Launch limits

Opt-in now. Astra default later.

Enable with features.context_management.experimental_mode = true, then start a new task
ChatGPT Plus, Pro, and Pro Lite sign-in on supported Codex clients
Not available with Business, Enterprise, or API-key sign-in at launch
1.05M-token Astra window; 128K max output tokens

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.