GPT-5.6 Sol just got better in one place and stayed the same everywhere else
The New Stack Amanda Caswell ● Covered by 9 sources
OpenAI quietly upgraded GPT-5.6 Sol, but only inside regular ChatGPT chat—Codex and Work stay untouched. That means the same model name behaves differently depending on where you run it.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI slipped an update into GPT-5.6 Sol on Thursday, but only for the version living inside ChatGPT's everyday chat experience. The company was explicit about the split: the Sol that powers Codex and ChatGPT Work isn't changing as part of this release. So if you've been testing prompts in consumer ChatGPT before shipping them into a coding or work environment, don't assume what you see will hold.
The update also folds ChatGPT's quick-answer and deep-thinking modes into a single Sol model, which is why Plus and Pro users now get a slider — usable on web, mobile, and desktop — to dial in how much reasoning effort they want on a given answer. Developers already make this tradeoff through the API, choosing when a slower, pricier answer is worth it. Now everyday ChatGPT users get a version of that same decision baked into the interface.
Behind the scenes, OpenAI's July System Card describes classifiers that watch Sol and Terra generate answers in real time, flagging and holding responses that look off before another system reviews them. Those classifiers get tuned separately per model, which helps explain why identical prompts might land differently across ChatGPT, Codex, and Work. The same System Card also notes something engineers will recognize: GPT-5.6 occasionally does more than what was asked, making unrequested changes. OpenAI says this happens more than it did with GPT-5.5, though it still calls the behavior rare.
On the accuracy front, OpenAI claims the updated Sol cuts down on factual slip-ups. In internal testing across financial, medical, and legal questions, answers containing at least one error showed up 68% less often than with GPT-5.5 Instant. Luna, soon to be the default model for Free and Go users, posted a roughly 62% reduction in the same kind of errors. Those numbers sound solid until you notice what's missing — no released prompts, no methodology detail, and a comparison baseline of GPT-5.5 Instant rather than the earlier version of Sol itself. That last choice makes it impossible to say how much Sol has actually improved versus its predecessor in ChatGPT.
Which leaves engineering teams back where they always end up: running their own prompts in the actual environment they'll deploy in, and saving the configuration so results can be checked and repeated. OpenAI's numbers describe a lab. Production is a different building.
My take — AI-written commentary, not fact-checked reporting
Announcing a 68% error reduction without publishing the test prompts or picking an honest baseline isn't really a benchmark — it's a press release wearing a lab coat. If OpenAI wants developers to trust these numbers, comparing new Sol against old Sol shouldn't be optional, it should be the whole point. Until then, the only benchmark that matters is the one teams run themselves, in the exact environment they're shipping to.
Read more about this at: The New Stack