GPT-5.6 Sol just got better in one place and stayed the same everywhere else
The New Stack Amanda Caswell ● Covered by 7 sources
OpenAI updated GPT-5.6 Sol, but only inside regular ChatGPT chat, not Codex or Work. Same model name, different behavior depending on where you use it — testing prompts in one place won't tell you how they'll run in another.
OpenAI dropped an update to GPT-5.6 Sol on Thursday, and the headline detail is where it doesn't apply. The refreshed model lives only in ChatGPT's everyday chat experience. Codex and ChatGPT Work keep running the older version, untouched. OpenAI said as much twice in its own announcement, which suggests the company anticipated confusion — and probably support tickets — from teams who assumed "GPT-5.6 Sol" meant one consistent thing everywhere.
It doesn't. ChatGPT has quietly diverged from its sibling products for a while now, but this update makes the split harder to ignore. Plus and Pro subscribers get a new slider that lets them dial up or down how much reasoning effort Sol puts into a response, available across web, mobile, and desktop. That's the same tradeoff API developers have been managing manually for months, just wrapped in a friendlier interface. Free and Go users get a different model entirely, called Luna, set as their new default.
OpenAI's July System Card for GPT-5.6 adds a layer worth understanding: both Sol and Terra run behind classifiers that watch answers as they're generated and can pause output for a second check if something looks off. Those classifiers get tuned per model, meaning Sol's guardrails in ChatGPT aren't identical to Sol's guardrails in Codex. The same document flags something engineers will recognize instantly — GPT-5.6 has a slightly higher tendency to make unrequested changes than GPT-5.5 did. OpenAI calls it rare. Anyone who's watched an agent quietly refactor a file nobody asked it to touch might disagree on the definition of rare.
The benchmark numbers OpenAI is touting sound impressive until you look at the comparison baseline. The company says the new Sol cuts factual errors by 68% on financial, medical, and legal questions, with Luna posting a 62% reduction. But those figures are measured against GPT-5.5 Instant, not the previous ChatGPT version of Sol, and OpenAI hasn't published the actual prompts. So there's no way to isolate how much of that gain comes from Sol improving versus from Instant simply being a weaker starting point. It's a comparison built to look good on a slide, not one built for anyone trying to reproduce the result.
Which leaves the practical takeaway squarely on engineering teams. Nothing about the API version of GPT-5.6 Sol has changed, so there's no signal yet for anyone building on top of it outside ChatGPT. If you're testing prompts, do it in the actual environment where they'll run in production, save the configuration alongside each result, and don't trust a headline percentage until you've watched it hold up against your own baseline.
My take
Splitting a model's name across three different products without splitting the version number is a marketing decision dressed up as an engineering one, and it's going to keep tripping people up. If OpenAI wants developers to trust its benchmarks, publishing eval prompts and comparing new versions against their direct predecessors — not a conveniently weaker stand-in — would be a start. Until then, treat every percentage in a System Card as an ad, not a spec sheet.
Read more about this at: The New Stack
