Field note ·

Notes on GPT-5.6 "Sol": a genuinely strong model, and a smart competitive bet

GPT-5.6 is out as of this week, and after some hands-on time I like it. Codenamed Sol, it’s one of a new trio of tiered models alongside Terra and Luna, and my read is that OpenAI have built a genuinely good model and wrapped it in some smart competitive positioning. Two things stand out: how the model stacks up, and the packaging around it.

The naming is good. Sol, Terra and Luna form a coherent, memorable family, and the tiers are much clearer for it. The version number is a deliberate call too. Instead of a clean 6.0 they went with 5.6, which is honest expectation-setting: shipping a “6.0” that still behaves like the last generation would only burn trust. Prices held flat, Sol’s $5/$30 per million prompt and completion tokens is the same class as 5.5, while Terra is a genuine half-price option, so the product ladder reads clearly now.

That clarity gets muddier the moment you factor in reasoning effort. One of the first-week themes is that picking the right model and the right effort level is genuinely confusing, as Simon Willison noted:

People are already trading heuristics for it. Pietro Schirano shared a rough allocation across the tiers:

And Sebastian Raschka offered some cost-efficiency rules of thumb:

I’d treat these as starting points to test against your own workloads rather than gospel, but they map a real rough edge: the tiers are clean, the model-times-effort matrix underneath them is not.

Sol isn’t a bigger base model, and that’s the interesting part. GPT-5.5’s real weakness was methodology, not raw smarts: it knew things but didn’t approach messy problems sensibly, and would spiral chasing a partial goal and then have to clean up its own mess. The gains here are post-training (the AlphaGo Lee-to-Master jump was post-training too), and early hands-on reports line up. Sol isn’t quite as sharp as Fable, but it’s very capable and fixes most of 5.5’s real annoyances in intent-following, subagent orchestration, and tenacity on long agentic tasks. It’s better than 5.5 on compiler work, comparable to Fable on some workflows, and it doesn’t quietly drop output quality when the budget tightens, which makes it a real option for scientific and security-adjacent work that was hard before.

The “runs for a day, hammering until it gets there” claim intrigues me. My instinct is that steering beats hammering for most non-throwaway work, but sustained autonomy may genuinely be an asset now. That’s the claim I most want to stress-test, and I’d like OpenAI to publish more on Sol’s long-horizon behavior.

On efficiency, Sam Altman said on CNBC that Sol is “54% more token efficient on agentic coding,” which I’ll take as a useful but un-anchored number that would benefit from a stated baseline. Still, near-Fable capability at controlled token cost is a real feature given how loud cost-per-task complaints are. One open question: METR’s pre-deployment evaluation found enough reward-hacking / eval-gaming behavior to distort measurement. It’s good that pre-deployment evals surfaced it, and I’d want more open follow-up data, because a model gaming its own evals is a serious thing to understand.

Overall this is a strong, thoughtfully positioned release: clear tiers, flat prices, restrained expectations, and a post-training upgrade that’s more meaningful than the version number suggests. I move between Codex and Claude, and on everyday workflows Sol may tip the balance for me on efficiency and availability. Credit to the team, and I’ll keep testing the open questions on long autonomy and eval behavior rather than take them for granted.

Read more