# GPT-6.2 Sol Is a Regression for Deep Research

> Source: <https://www.vincentschmalbach.com/gpt-6-2-sol-regression-deep-research/>
> Published: 2026-11-19 08:32:40+00:00

**Recommendation:** Judge model upgrades with blind, order-swapped pairwise comparisons that weigh depth and count concrete details, not just key-fact accuracy or index scores, before switching a production pipeline.
More in the playbook "Evals and testing for LLM apps": <https://wpnews.pro/playbooks/evals-testing>

### Microsoft Shows How Not to Introduce AI in Your Business

Since Thursday afternoon I can't use my Outlook. On November 12 Microsoft 365 went down worldwide. Exchange Online, which runs the email,…

GPT-6.2 Sol is a clear regression from GPT-6.1 Sol for deep research. I switched deep research in [UnSpar](https://www.unspar.com/), the article tool I build, to 6.2 the day after OpenAI released it, and switched back the next morning. 6.2 writes research reports half as long as 6.1, with fewer details and fewer citations, and turning up the reasoning effort doesn't change that.

[OpenAI released GPT-6.2 Sol](https://community.openai.com/t/gpt-6-2-astra-and-sol-are-now-available/1417531) on November 17 at the same price as 6.1, $2 per million input tokens and $10 per million output tokens. Artificial Analysis gives it [57 points on its index at $1.01 per task](https://artificialanalysis.ai/models/gpt-6-2-sol), against [55 points and $1.27](https://artificialanalysis.ai/models/gpt-6-1-sol) for 6.1. It looked smarter and cheaper, so I switched.

My first test agreed. I ran eight research topics, once through each model and had Claude Opus 5.5 grade the reports blind. Both models got nearly all of the five key questions per topic right. 6.2 won three topics head to head and 6.1 won one, and 6.2 was 18% cheaper and more than twice as fast. I deployed it. The same test also showed 6.2's reports at 1,370 words against 2,580, with 18 cited links instead of 26. I had built the test to check whether the key facts were right.

For the second test I had the judge list every concrete detail in each report and weigh depth as much as correctness, without counting length. Each pair was judged twice with the order swapped, and a topic only counts when both rounds agree. This time 6.2 also got a higher reasoning setting than 6.1.

|  | GPT-6.1 Sol | GPT-6.2 Sol (higher reasoning) | 
|---|---|---|
| Topics won (of 8) | 4, plus 1 leaning | 1 | 
| Topics won on depth | 4 | 0 | 
| Relevant details per report | 51 | 46 | 
| Words per report | 2,610 | 1,390 | 

At the same reasoning setting, 6.2 used about 2,600 reasoning tokens per report where 6.1 used 7,100. With GPT-5.6 Sol I had the opposite problem, it [used twice the tokens](https://www.vincentschmalbach.com/gpt-5-6-sol-xhigh-uses-twice-tokens-gpt-5-5/) of the model before it. So I suspected the setting and turned it up. On two of the topics, 6.2 at max reasoned almost twice as much as 6.1 (10,600 tokens against 5,900), still wrote about 1,200 words, and listed 36 details per report, against 48 for 6.1 and 45 for itself one setting lower.

The price per token didn't change, so all of the savings come from 6.2 thinking and writing less. With the higher setting the saving shrinks to 8%. I don't count research that finds less for less money as cheaper.

So Deep research in UnSpar is back on 6.1.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com
