GPT-6 Sol and Opus 5.2 are simultaneously in gray testing. Sam Altman and Dario Amodei just called for hitting the brakes over the weekend, but stepped on the gas the moment they got back to work.

Who is the one shouting for brakes over the weekend, but secretly hits the gas pedal as soon as work resumes on Monday?

Sam Altman and Dario Amodei, neither of you can really brag, one is running a gray test for GPT-6 Sol, the other is running a gray test for Opus 5.2.

Some netizens previously tested rocket launch animation generation with Opus 5, and they tried it again after hearing that their requests might now be routed to version 5.2 — the improvement is visible to the naked eye.

He also added that he had tested this scenario many times before, and the realism of the animation this time has truly reached a new level.

Amodei's side is in a hurry to roll out a model that can compete with GPT-6 Astra, after all, Anthropic is planning an IPO this year, and it cannot afford to lose momentum.

Altman's side, on the other hand, is rushing to release a cheaper model with Astra-level capabilities, to reactivate its $200 subscription tier.

The two sides reached an unspoken agreement, and decided to go live this week.

There are not many leaked test results for GPT-6 Sol, which are mainly distinguished by knowledge cutoff date.

The GPT-5.6 series has a knowledge cutoff date of February 16, 2026, and it considers the latest Claude Opus model to be version 4.6.

If it can answer that the latest Opus model released on April 16 is Opus 4.7, congratulations, you have been selected for the gray test.

Why am I never lucky enough to get access to this kind of good stuff?

Opus Skips 5.1 and Jumps Straight to 5.2

Since Opus 5.2 is built to push performance limits, let's compare its strongest suit with GPT-6 Astra: SVG animation generation.

While the material lighting effects generated by GPT-6 Astra look more three-dimensional, some users prefer the aesthetic style of Opus 5.2.

Others have found that you can only get access to Opus 5.2 on Claude Code, and the web version is still most likely running version 5.0.

Another intuitive comparison is static SVG modeling of complex objects. Opus 5 cannot even get the structure right, and while 5.2 is still not as good as its counterpart Astra, it produces pretty decent results.

By the way, the PS controller drawn by Astra with SVG looks exactly the same as the real physical product...

Some users have tested it for a whole day, and concluded that Opus 5.2 has made significant progress compared to Opus 5, runs much faster, is competitive with Astra, and has greatly improved front-end and 3D generation capabilities.

But why are we sure it's called Opus 5.2 instead of 5.1? After all, Anthropic just released Fable 5.1 two weeks ago.

It turns out that this version number was dug out from Microsoft's code.

The Entire GPT-6 Family Is Coming Together

Tibo, the well-known industry insider, said that this week's release is comparable to last year's OpenAI DevDay.

Combined with the ongoing GPT-6 Sol test, it is very likely that the full series of models will be released together. Could it be Thursday again?

Some enthusiastic users have even made the promotional poster for this release.

There are not many leaked test results for now, but all the leaked results show that the new models are no worse than Astra, and will be more affordable. So it is absolutely worth looking forward to.

One More Thing

Out with the old, in with the new.

OpenAI has just announced that the GPT-5.5 series will be retired on October 14.

Speaking of which, 5.5 is a legendary model that helped Codex users outperform Claude Code, but it will be discontinued in October just after its release in April.

It has to be said that the AI industry is developing extremely fast in 2026.

Back then (actually last December), when Gemini 3 Pro was dominating the market, Altman sounded the red alert and rushed to develop an internal model codenamed "Garlic", which later became GPT-5.5 half a year later.

Now Gemini has become a cute pet, and GPT-5.5 has successfully fulfilled its mission.

Speaking of GPT-5.5, OpenAI held an offline party for it on May 5. More than 8,000 people signed up, but due to venue limitations, OpenAI gave 10 times the regular Codex quota to all applicants who were not selected to attend the event.

So for the OpenAI DevDay at the end of September this year, you can sign up first no matter if you can attend in person, who knows if they will give away extra quota again?

References:

[1]https://x.com/Ananth7e/status/2099831306603684294?s=20

[2]https://x.com/Mr_Salio/status/2099503024292970966?s=20

[3]https://x.com/ChatGPT/status/2099954190600876533?s=20

This article is from the WeChat public account "QbitAI", written by Meng Chen, and authorized for release by 36Kr.