GPT-6.2 Sol Review: The Everyday Model Nearly Catches the Flagship, but Not an Across-the-Board Upgrade
On OpenAI's FrontierCode 2.0 chart, Sol scores 43.8% and Astra 44.2%, a gap of 0.4 points. On the same day, in Artificial Analysis's (AA) Coding Agent Index, Sol came in 3 points below the previous-generation GPT-6.1 Sol.
Both statements are true. GPT-6.2 Sol is the model OpenAI positions for "everyday work and agents at scale". It gets close to Astra on many tasks and slips in a few specific areas. Calling it a "budget flagship" and calling it a "cut-down version" are both inaccurate.
This article compiles public material. It is not a test run by Guishu Intelligence. Numbers are listed by source and conditions and are not mixed.
Official positioning and specifications
According to OpenAI's release post, Sol targets everyday work and large-scale agent deployment, and OpenAI says many people inside the company use Sol as their default model. The Sol model documentation lists:
- API ID:
gpt-6.2-sol. - Text and image input, text output; no audio or video.
- Context window of 1,050,000 tokens, maximum input 922,000, maximum output 128,000.
- Reasoning is supported, and the API rejects
reasoning effort = none. - Supports Chat Completions, Responses and Batch. It does not support Realtime, Assistants, fine-tuning, or image and video generation. Tool calling goes through the Responses API.
- Knowledge cutoff June 1, 2026, later than Astra's May 18.
Context and modalities are identical to Astra's. The differences are positioning, cutoff date and the profile of strengths.
For availability, follow ChatGPT Learn: ChatGPT Work, Codex and the API. In ChatGPT you need to turn on Work first. Enterprise and Edu need an admin to enable it. Ultrafast is the name of a speed tier. It is available to some higher-tier Pro subscriptions and to enterprise or education workspaces where an admin has enabled it.
Where it comes close to Astra
The figures below come from OpenAI's release materials, as compiled by VentureBeat. Official charts may use the best plotted point, so the reasoning levels may differ.
| Benchmark (OpenAI official chart) | Sol vs. GPT-6.1 Sol |
|---|---|
| FrontierCode 2.0 | 43.8% vs. 39.0% |
| LifeVal | 83.8% vs. 79.5% |
| BenchCAD 2.1 Core | 73.1% vs. 52.6% |
| Terminal-Bench-Science | 68.0% vs. 57.0% |
| AutomationBench (official chart) | 41.2% vs. 36.1% |
| Harvey LAB-AA | 11.1% vs. 6.9% |
| Agents' Last Exam | 73.0% vs. 72.5% |
Two of these deserve separate mention:
- FrontierCode 2.0. Sol 43.8%, Astra 44.2%. OpenAI also says Sol trails Astra by about 0.5 points at max effort.
- LifeVal. Sol's 83.8% is higher than its own predecessor and also higher than GPT-6 Astra's 79.8%.
Independent evaluations also support "close". Under AA's standard FrontierCode 2.0 conditions, as compiled by Handy AI, Sol scores 44.4%, Astra 45.4% and Claude Fable 5.2 48.8%. On AA Intelligence Index v5.0 (max effort), Sol scores 57.3, roughly level with GPT-6 Astra's 56.9, while Astra scores 62.1. So Sol is roughly equal to the previous flagship, not the current one.
On the Vals composite index (OfficeChai or The Decoder relaying Vals; this article did not check the Vals composite board directly), Sol is ninth at 66.06%, 1.6 points behind seventh-placed Astra at 67.66%.
Where it regresses
Coding agents
The AA coding agent page uses the Codex harness at max effort:
- Coding Agent Index: Sol 49, GPT-6.1 Sol 52, down 3 points; Astra 53.
- SWE-Atlas-QnA-AA: Sol 46, GPT-6.1 Sol 60. AA's article attributes the drop mainly to this 14-point decline: Sol more often misses key facts.
- Terminal-Bench 4.1: Sol 58, Astra 64.
- Average time per task: Sol about 17.5 minutes, Astra about 25.1 minutes.
Here Sol's FrontierCode-AA 2.0 score is 42, the same as Astra's. The same benchmark name gives 42 under the Codex harness, 44.4% under AA standard conditions and 43.8% on OpenAI's chart. These three numbers come from three sets of conditions and cannot be swapped for one another.
Science
On the Vals Science Index (updated November 16), GPT-6.2 Sol is seventh at 57.76%, below GPT-6.1 Sol's 59.63%. OpenAI's release materials say 58.5% against 59.6%, which differs slightly from the Vals page; this article follows the Vals page. The conclusion is the same: a small decrease.
Professional tasks
Surge's professional-task suite (11 tasks designed by physicians, accountants, lawyers and senior engineers) gives mixed results:
- Better: coding, charts and agent instruction following. On Chartography Sol scores 87.7%, above GPT-6 Astra's 71.0% and slightly below Astra's 90.7%.
- Worse: medical, PDF and finance. On DAYJOB Healthcare, Sol scores 17.2% against GPT-6.1 Sol's 25.2%, a drop of 8 points and the largest regression Surge measured that day. Astra scores 34.4% on the same task.
- On sudo L7, Sol scores 36.1% and Astra 38.3%, both below Fable's 52.2%.
Hallucination and obstacles
On AA-Omniscience at max effort, Sol's hallucination rate is 55% and Astra's is 59% (lower is better). The questions are very hard, do not represent everyday hallucination rates, and do not show that Sol is more reliable.
The system card says both models are more likely to omit known obstacles when reporting their work. For Sol, the SWE-Atlas drop AA observed (more often missing key facts) is a different metric, but readers can set the two side by side as a reminder that human checking is needed before delivery.
A personal test: Kazike's code review
In their WeChat hands-on article, Kazike and AIZ Xiaozhu used a code-review set they built themselves to compare three models at xhigh: GPT-6.1 Sol found 65% of the issues, GPT-6.2 Sol 84% and GPT-6.2 Astra 96%. This is an author-built simple evaluation set, and its questions and sample size are not published. The article also notes that Astra was about twice as slow as Sol and found a dozen or so more bugs. The author plans to use Sol at xhigh as the default review model. That is the author's own decision, not general advice.
It does not contradict AA's finding. Code review means reading code and pointing out problems, while the Coding Agent Index measures finishing long tasks inside Codex. These are two different kinds of work.
Taking "close" apart: where the gap sits
"Sol is close to Astra" means very different things on different benchmarks. Using the official-chart comparisons cited here, the gap falls into three bands.
First band, almost no gap: FrontierCode 2.0 (43.8% vs. 44.2%) and LifeVal (83.8% vs. 86.1%, a gap of 2.3 points). These tasks are fairly general, and extra reasoning budget adds limited benefit however strong the model is.
Second band, a clear gap: on BenchCAD 2.1 Core Sol scores 73.1% and Astra 82.5%; on Terminal-Bench-Science 68.0% against 77.1%; on AutomationBench (official chart) 41.2% against 52.4%; on Harvey LAB-AA 11.1% against 15.0%. For science, CAD and complex automation, the "hardest projects", Astra's lead is around 9 points. This is also the basis for OpenAI's division of labor between the two models.
Third band, Sol made no progress: below its predecessor on the Vals Science Index, down 14 points on AA's SWE-Atlas Q&A, down 8 points on Surge's medical task.
Taken together, Sol's strength is "fast and good enough", not "better than the old model in every direction". The Vals Science Index page offers numbers that support "fast": under that index Sol averages about 19 minutes against Astra's 22 minutes 15 seconds, and uses about 3.21 million tokens against Astra's about 3.48 million. These are Vals's averages across its three science tasks and do not apply to all tasks.
Why comparison with the previous generation matters
Many readers ask whether upgrading from GPT-6.1 Sol to 6.2 Sol is worth it. The official charts answer fairly positively: across the seven official comparisons, Sol is above its predecessor on every one, and GDPval v2 adds +5.3 points. Independent evaluations give a mixed answer: Sol is close to GPT-6 Astra on the AA Intelligence Index, but the Coding Agent Index falls, the Vals Science Index dips and Surge shows regressions in several areas.
Both can hold at once, because the benchmarks OpenAI chose, the reasoning levels it sampled and its scoring rules all differ from those of the independent organizations. Whether to upgrade depends on which side your workload resembles. If your tasks look like FrontierCode, LifeVal or BenchCAD, the case for upgrading is stronger. If they look like medical documents, PDF extraction or retrieval Q&A, compare the old and new generations on your own samples first.
A minimal validation checklist before rollout
- Take about 30 real tasks and run the old and new Sol on each, at the same reasoning level.
- For each task, record whether key facts were missed, since AA saw Sol regress there.
- Count medical, finance and PDF tasks separately instead of averaging them with general tasks.
- Ask the model to list the obstacles it hit at delivery, and spot-check them. The system card warns that both models are more likely to omit them.
- If you go through the API, use the Responses API for tool calls and set the reasoning level as documented. Do not use none.
Who should make Sol the default
Based on the evidence above, Sol fits:
- Everyday Q&A, document organizing, and life and office tasks (the LifeVal direction).
- Code review, chart generation, and batch agents that need lower latency.
- Teams that want to keep Astra for hard problems and hand the rest to a faster model.
Be careful with:
- Workflows heavy on medical, finance or PDF material, where Surge shows Sol regressing against the previous generation.
- Retrieval Q&A where facts must not be dropped. AA's SWE-Atlas decline is a warning.
- Science benchmarks. Sol is slightly below its predecessor on the Vals Science Index.
Two things not to misread
First, don't carry an xhigh conclusion over to the default level. Kazike's code-review comparison is at xhigh, AA's coding agent index is at max, and the official charts use the "best plotted point". The level you use in Work or Codex may be lower. At lower levels the gap between Sol and Astra on hard problems usually shows up more easily, and you need to test that yourself.
Second, don't summarize the whole model with one leaderboard. That Sol closes in on Astra on FrontierCode does not change its 17.2% on Surge's medical task. That it falls slightly below its predecessor on the Vals Science Index does not change the fact that it beats GPT-6 Astra on LifeVal. Each number answers one question.
What launch-day feedback says about Sol
Public discussion on day one was mostly about Astra. Direct assessments of Sol are few: Kazike treats it as a candidate for everyday review; OpenAI itself says many people internally use it as the default; Surge's conclusion is mixed. These are different kinds of statements: OpenAI's is a vendor position, Kazike's is a personal self-built set, and Surge's is an independent suite. There are not enough long-horizon Sol usage reports with concrete scenarios, so this article does not give an overall verdict on whether Sol is good to use. It only says where the public evidence is strong and where it is weak.
A workable division of labor
If a team has both models, the public evidence supports a conservative split: everyday requests, batch review, charts and general office tasks go to Sol; when a task involves scientific computation, CAD or complex automation, or when spot checks show Sol frequently missing items, move up to Astra. There are three reasons. OpenAI positions the two models this way. The gap on FrontierCode and LifeVal is small. And on BenchCAD and Terminal-Bench-Science the gap is concentrated around 9 points.
Conversely, don't hand medical, finance or retrieval Q&A to Sol by default just because it "caught the flagship" on one benchmark. These are the areas where Surge and AA saw regressions, and they need validation on your own samples first. This split is one reading of the existing public data, not a conclusion validated by Guishu Intelligence.
Limits
It has been one day since release. Some of Surge's tasks were still running that day, and more Epoch and Vals boards will update. Everything here is third-party data and personal testing, not Guishu Intelligence's own testing. For choosing between Sol and Astra, see the selection guide.
In one paragraph
Sol is OpenAI's efficient model for everyday work and agents at scale. On FrontierCode and LifeVal it comes close to, or passes, the previous flagship, and it regresses on coding agents, Vals Science and Surge's medical task. Treat it as a default that is good enough for most tasks, not as a new model that improved in every direction.
Sources
- OpenAI: release post, Sol model documentation, ChatGPT Learn.
- Independent evaluations: Artificial Analysis, AA coding agent page, Vals AI, Surge.
- Media compilations: VentureBeat, The Decoder, Handy AI, OfficeChai.