More Model Choice and Agent Scheduling: What's New, 14-20 August 2026
This 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
GPT-6.2 Sol is now available in Autohive. You can choose it as the model for your agent in its settings. OpenAI released Sol on 17 November 2026, alongside GPT-6.2 Astra. It describes Sol as a model for everyday work, from writing to tasks with several steps.
Sol supports actions, vision, reasoning, and streaming replies.
OpenAI describes Sol as an efficient, versatile model for everyday work. In its launch post on OpenAI’s community forum, OpenAI says Sol is meant for clear, natural writing, technical and non-technical tasks, and multistep agentic workflows.
An agentic workflow is a job where the AI works through several steps towards a goal. It might read a request, look something up in a connected app, draft a reply, and hand it back to you to check.
These are OpenAI’s descriptions of its own model. How well Sol handles your work depends on your instructions, knowledge sources, and connected actions. Test it on real tasks before you rely on it.
Start with tasks you already run, where you know what a good result looks like. These are proposed tests, not results we’ve seen from Sol. Each uses inputs you supply, so you can compare Sol’s output with work your team has already checked.
| Task | What you supply | Pass check |
|---|---|---|
| Weekly status summary | Last week’s project notes | Every update is covered, nothing is added, and open risks are flagged |
| Inbound request triage | Past tickets your team has already sorted | Each lands in your team’s category, with a reason tied to the ticket text |
| Explain a code or config change | The code change plus the ticket behind it | Matches what the change does, names the files touched, and flags what the code change can’t show |
| First draft of a customer reply | A past email thread and style notes | Answers every question and sticks to facts in the thread |
A weekly summary is a useful first test. Compare it with last week’s manual version. For background on recurring work and connected steps, see Autohive’s guides to scheduled jobs and creating workflows.
OpenAI and Artificial Analysis published findings on 17 November. They used different tests and compared different models. Neither result tells you how Sol will perform on your agent’s tasks.
OpenAI says Sol nearly matches Astra on FrontierCode 2.0. This tests whether a model can produce code changes that are ready to merge. It is OpenAI’s own test, comparing its two new models. Treat it as a vendor-reported result.
Independent evaluator Artificial Analysis reported an improved overall Intelligence Index result for Sol. Its Coding Agent Index result fell compared with GPT-6.1 Sol. Artificial Analysis ran that test in OpenAI’s Codex at maximum effort, which may differ from how your agent is set up.
Artificial Analysis traced the drop to one test, SWE-Atlas-QnA-AA. It says Sol was more likely than its predecessor to leave key facts out of its responses.
Check that Sol’s answers include every fact you need. Use the table’s completeness checks when you review them.
OpenAI says both GPT-6.2 models launched with tighter safeguards for higher-risk cybersecurity help. If your agents support security work, test them with real, authorised cases from your team. Examples could include reviewing alerts or explaining a vulnerability. Check how Sol handles the requests you rely on before using it for live work.
Use a separate test agent so you can try Sol without changing the agent your team relies on.
Test GPT-6.2 Sol in Autohive on a separate agent first, using tasks you’ve already checked. If it handles them well, with nothing missing and no unsupported claims, consider changing your live agent’s model. Moving one agent at a time makes problems easier to spot and limits how much work a change affects.
OpenAI released both models on 17 November 2026. It describes Sol as efficient and versatile, aimed at everyday work, writing, and multistep tasks. On OpenAI’s own FrontierCode 2.0 coding test, Sol comes close to Astra on merge-ready changes. Check your agent’s model settings to see which models you can pick.
Check that answers are complete, because Artificial Analysis saw Sol leave out key facts on one test. Check claims against your source material, and make sure multistep jobs finish every step. Keep a person approving anything that leaves your team until you’ve checked the results on your own tasks.
Pick one task from the table and run it on a GPT-6.2 Sol test agent this week.
This week's update brings three new AI models, lets agents create and manage their own scheduled jobs, and adds six Marketplace integrations including …
Read articleYou can now try interactive examples in Autohive Marketplace listings before signing up, and this week's updates also add chat branching, a summary …
Read article