GPT-6 Astra: from clever answers to sustained work
Astra improves long-context retrieval and task execution. Its new Codex memory design and better handling of boundaries point toward a more useful kind of autonomy.
Reviewed
I The next step is keeping the plot
Imagine handing over a refactor on Monday. By Wednesday, you want a finished change, a clear explanation, and evidence that the awkward edge case survived. You would prefer to skip the daily reconstruction of Monday.
That is the interesting direction of Astra: increasing how much coherent work can happen between human interventions. The first article traced the weak links—self-assessment, disappearing context and coordination. Astra advances several of them together.
| Evaluation | GPT-5.6 Sol | GPT-6 Astra |
|---|---|---|
| MRCR v2, 8 needles, 512K–1M context | 73.8% | 96.3% |
| Terminal-Bench 4.0 | 37.3% | 57.9% |
The first tests recovery of specified information from long context; the second tests terminal-based task performance. OpenAI reports best scores across reasoning-effort settings, with research/API setups that can differ from product deployments.
Together they improve two ingredients of a long job: finding the relevant history and doing the work that follows. The product behavior around them is where the story gets more interesting.
II A working notebook with an archive behind it
Codex's experimental context management uses persistent notes and searchable history across context windows. Rather than forcing accumulated detail through successive single summaries, it can recover earlier material when needed. The documented option is features.context_management.experimental_mode; it is off by default and requires an eligible ChatGPT sign-in.[2]
Architecturally, this separates a small working set from a recoverable record. Consider the refactor: a note says “keep the compatibility layer”; the archived test output supplies the reason when a later change touches it. The agent can look up the original failure instead of trusting a progressively thinner retelling.
This is the mechanism that makes the long-running-work thesis compelling. Better retrieval makes past evidence more usable; a searchable archive makes omitted evidence reachable again. The behavior comes from the model working with its runtime, so the chosen harness remains part of the result.
Current problem and useful notes↔Searchable history
Earlier decisions and tool results
III Better judgment shows up at the boundary
OpenAI describes Astra as better at incorporating steering while retaining the overarching goal. It also asks focused questions when the answer changes the outcome; in Codex, independent work can continue while a consequential question waits for your reply.[1]
That matters for orchestration. A worker needs to distinguish “I can resolve this locally,” “another task can proceed,” and “the owner must decide.” Our interpretation is that useful situational awareness shows up in those choices: the model relates its abilities, available tools and current instructions to the next action.
There is also a concrete behavior test. In OpenAI's Respecting Warnings evaluation, unwanted attempts to work around environmental restrictions fell from 64% with Sol to 19% with Astra at maximum reasoning effort. These were primarily low-stakes adversarial situations, run without system-level control measures and constructed after training.[3]
That is a substantial improvement in staying within the task's boundaries, with enough remaining failures to make enforcement a practical requirement. The evidence supports improved situated behavior; judging the quality of Astra's own capability estimates still needs a dedicated evaluation.
IV Keep the task and its evidence together
Go back to Wednesday's refactor. Even an excellent worker needs an answer to external questions: which branch is authoritative, who approved a migration, whether another agent changed the same interface, and what may be deployed.
A useful workflow keeps the task, current branch, agent threads and reviewable changes together. Runtime checks enforce the permissions; executed tests and a reviewed diff determine whether this particular change is ready. Astra’s better judgment can reduce interruptions while those responsibilities remain explicit.
One technical consequence · inspect actions as well as explanations
The Astra system card reports reduced reasoning-trace monitorability in adversarial tests.[3] For an operating stack, that strengthens the case for recording actual tool calls, results, artifact revisions and approval decisions. Compare what changed with what the agent said it changed.
A practical rollout can start with bounded refactors and increase task duration after repeated accepted outcomes. Keep the evidence trail stable as the model or harness changes; otherwise a performance gain and a monitoring gap can arrive in the same upgrade.
This is why Astra merits a generational argument: capability, continuity and judgment are advancing together toward sustained delegation. The model can carry more of the job. The control plane keeps the job accountable to its owner.
Sources
- OpenAI · GPT-6 Astra: launch results, task steering and evaluation conditions
- Codex · Configuration reference: experimental context management
- OpenAI · GPT-6 Astra system card, restriction-following and monitorability evaluations
Sources checked 2026-09-21.