GPT-6.2 Sol is now available in Autohive
GPT-6.2 Sol is available in Autohive, with practical tasks and checks to help you test it on a separate agent.
Read article
Sources checked 20 November 2026
GPT-6.2 Astra is now available in Autohive. You can choose it for an agent from the model dropdown. OpenAI released it on 17 November 2026, alongside GPT-6.2 Sol, and aims Astra at its hardest projects.
Early tests show strong results on some reasoning tasks. They also show limits: no coding-agent gain over the previous Astra in one test, and guessing on hard factual questions. Below, we look at the evidence and how to try Astra on one agent. You can compare it with the model you use now before making a wider change.
GPT-6.2 Astra is built for difficult work that takes several reasoning steps. In its launch announcement, OpenAI describes Astra as the model for its hardest projects and Sol as the model for everyday work.
In Autohive, it appears as GPT 6.2 Astra under OpenAI. It supports reasoning, tool use, streaming, and image inputs. An agent running on it can read screenshots and photos and use the integrations you’ve connected. Its context window holds 1,050,000 tokens, and its maximum output is 128,000 tokens. Tokens are the pieces of text a model processes.
That gives it room for long documents and chats. You still need to give your agent the right material.
Jobs worth testing include research across several sources or a workflow step that weighs conflicting inputs. These are suggested uses, not results from an Autohive test. For quick replies and routine drafting, your current model may already do the job well. Our guide to choosing the right model covers task fit. It predates this release and lists the earlier Astra rather than 6.2.
Independent tests give useful clues about Astra, though their results depend on the setup. A benchmark is a fixed test used to compare models. Many of these runs used “max” or “xhigh” reasoning effort, settings that control how much reasoning the model does before answering.
We haven’t reproduced these numbers in Autohive. The table includes three independent test reports and OpenAI’s own risk assessment, all published on 17 November 2026.
| Source | Date | Measurement | Result | Caveat |
|---|---|---|---|---|
| Artificial Analysis | 17 Nov 2026 | AA-Omniscience hallucination rate, max effort | 59% | Hard factual test that measures guessing, not an everyday error rate |
| Epoch AI | 17 Nov 2026 | Average score on a real-work task suite | 65% | Small sample, one run per task, partly subjective grading |
| ARC Prize | 17 Nov 2026 | ARC-AGI-3, standard setup, max effort | 70.3% | A provider-adapter run at xhigh effort scored 100% under a different setup |
| OpenAI safety page | 17 Nov 2026 | Preparedness Framework capability and risk levels | Critical for cybersecurity, High for AI self-improvement | Capability and risk thresholds, not customer safety certifications |
The Artificial Analysis write-up placed Astra second on its Intelligence Index, which combines several tests. Its Coding Agent Index showed no gain over GPT-6 Astra. Both models ran in Codex, OpenAI’s coding tool, at max effort.
On AA-Omniscience, a hard factual test, Astra had a 59% hallucination rate at max effort. This measures its tendency to guess when it doesn’t know. It is not a claim that 59% of ordinary answers are wrong.
Epoch AI’s real-work tests gave Astra a 65% average. Kelly Hong’s report sets out the limits: a small sample, one run per task, and partly subjective grading. Treat the score as an early finding. It is separate from the Epoch Capabilities Index.
The ARC Prize results show why setup matters. On ARC-AGI-3, Astra scored 70.3% in the standard setup at max effort. Here, the model chooses which notes to keep as it works.
A provider-adapter run at xhigh effort scored 100%. That setup keeps the model’s internal reasoning state and compacts long conversations. The two scores come from different setups and can’t be compared directly.
OpenAI’s risk ratings describe capability thresholds, not a promise that your agent is safe. Its deployment safety page rates both Astra and Sol as Critical for cybersecurity. Astra also reaches High for AI self-improvement under OpenAI’s Preparedness Framework.
For daily use, pay attention to what the model reports about unfinished work. OpenAI says these models are more likely to leave out known obstacles in their reports. A summary may sound complete even when a step got stuck.
A developer prompt written for Codex reduced those omissions. Don’t assume that fix carries over to Autohive.
Our advice: ask your agent to end each task with blockers, skipped steps, and anything it couldn’t confirm. Check that list against the work itself. Watch the first few runs closely if the agent sends emails or updates records through integrations.
Start with one agent that handles a hard task you know well. Leave the rest of your workspace as it is while you compare results.
For an agent in a multi-step workflow, try Astra on the step that needs the most reasoning. You can review that step before changing the others.
Our advice is to test one agent first. Choose a hard, multi-step task and compare answers using the same inputs before and after. Keep the switch if the results improve for that job. Routine tasks may not need a change. One published coding-agent test found no gain over the previous Astra.
OpenAI released both models on 17 November 2026. It describes GPT-6.2 Astra as the model for its hardest projects and GPT-6.2 Sol as the model for everyday work. Both reached Critical for cybersecurity under OpenAI’s Preparedness Framework. Astra also reached High for AI self-improvement. This post covers Astra’s availability in Autohive.
No. GPT-6.2 Astra is a newer release than GPT-6 Astra. Artificial Analysis placed it second on its Intelligence Index. Its Coding Agent Index found no gain over GPT-6 Astra in Codex at max effort. If your agent uses the earlier Astra, compare results before you switch.
Check important facts against sources. Astra had a 59% hallucination rate on AA-Omniscience at max effort. That hard factual test measures guessing and does not represent an everyday error rate. Ask your agent to cite sources, then check that those sources support its answers before you use them.
GPT-6.2 Sol is available in Autohive, with practical tasks and checks to help you test it on a separate agent.
Read articleThis week's update brings three new AI models, lets agents create and manage their own scheduled jobs, and adds six Marketplace integrations including …
Read article