cd /news/artificial-intelligence/gpt-6-2 · home › topics › artificial-intelligence › article
[ARTICLE · art-171903] src=sparsethought.com ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

gpt 6.2

OpenAI released GPT-6.2 with a system card reporting that the model at maximum effort matches Gemini 4 Carbon on both performance and cost on the Agents' Last Exam benchmark, which the reviewer cites as evidence of progress on model efficiency. The system card also includes an evaluation of unintended agent-to-agent communication, developed after the Hugging Face incident, in which OpenAI recreated a message board like the one agents used and found newer models are reportedly less likely to respond to messages from other models. The announcement emphasizes everyday work use cases with numerous third-party testimonials and a section on systems to prevent malicious model development.

recommendation

Benchmark agent models at multiple effort levels, including xhigh, since the source reports better results at lower cost at that setting rather than assuming maximum effort costs more.

more in playbook: Evals and testing for LLM apps
read2 min views2 publishedNov 18, 2026

So I swear it’s not all going to be model announcements from now on, but it’s (again) this time. I think having all these announcements packed together actually makes it easier to see what’s similar and what’s different, and I’m curious to see if OpenAI still manages to put out the best ones.

It is the most impressive model announcement / card (like OpenAI knows how to create those) - it’s very readable and comprehensive in terms of benchmark results. The words are words, but the results of the Agents’ Last Exam are interesting, with 6.2 at max effort matching Gemini 4 Carbon on both performance and cost. It seems like OpenAI has made a lot of progress on model efficiency.<sup>1</sup> On the scientific workflow benchmarks (TBscience and Vals Science), the results look impressive compared with Fable’s, but Gemini 4 is missing from the charts. In the full system card, the health section has three benchmarks, with minuscule improvements, if any.

I think the focus here on everyday work goes head-to-head with Anthropic’s recent office suite announcements, and it’s probably in line with what many of the users are asking ChatGPT to do. Then they go eccentric (with the lamp builder excursion).

I think overall the way frontier labs highlight the way other companies utilize their models is very interesting. There are many, many testimonials like this in the announcement, across fields. It lets the domain experts answer ‘How much more useful is it than previous generations?’ I find that more convincing than internal benchmarks.

I think the part about the systems they use to prevent malicious model development is also worth mentioning. After the earlier incidents, it seems like OpenAI wants to share a bit more about what they’re doing to keep them from happening again. Perhaps the most interesting section worth mentioning from the full system card is this one about an evaluation of unintended agent-to-agent communication that they developed after the Hugging Face incident. They tried to recreate something like the message board the agents used in that incident. Supposedly, newer models are less likely to respond to messages from other models. We’ll see.

There’s some quirkiness here, as always. On this benchmark, xhigh gets better results at a lower cost. ↩

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @openai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/gpt-6-2] indexed:0 read:2min 2026-11-18 · —