GPT-5.6, the second opinion.

There are two ways to be excellent. One is to be the best in the room at one thing. The other is to be very good at everything and completely unflappable. GPT-5.6 is the second kind, and once you see it that way, its results make sense.

On the scoreboard that tracks real working sessions, it holds fifth place, the best result of any non-Claude model. But buried in that same scoreboard are two details we find more interesting than the rank. First, GPT-5.6’s work draws more praise and fewer complaints from real users than any other model’s. Second, when a step fails mid-task, and in real work steps always fail, GPT models recover fastest.

Best at: drafts that clients read and just approve

The praise number deserves a moment. Millions of sessions, real people doing real work, and this model’s output made them complain the least. That is not a benchmark score, it is a customer satisfaction score. For the drafts your team reviews every morning, it translates directly: fewer rewrites, fewer eye-rolls, more clicking approve and moving on.

In Jyper, that makes GPT-5.6 a natural for revisions and for structured drafting. The family comes in three sizes, and OpenAI gave them tiers with names: Sol is the flagship thinker, Terra the middle tier for structured work, Luna the cheap fast one. Jyper uses Terra for revisions and extraction and Luna for the small quick reads, the same way you would not send a director to fix a typo.

Best at: not falling apart when something breaks

Real operational work is a parade of small failures. The supplier portal times out. The attachment will not open. The rate in the reply does not match the rate in the contract. A model that stalls or panics at the first broken step needs a human standing behind it, which defeats the point.

There is also a stress-test study called Boundary-Bench that measures something quietly important: how much worse a model gets when you lock it down with strict security rules, the way any serious company must. Every model gets worse. GPT-5.6 got worse the least, placing second overall. The score inside the guardrails is the score that predicts real life, because that is where these models actually work.

Speed, accuracy, cost

The three numbers. Accuracy: within touching distance of the very top on nearly every scoreboard, and first on user satisfaction. Speed: middle of the pack, around 73 words a second, per Artificial Analysis. Cost: flagship rates. AI is billed in tokens, small chunks of text, and Sol charges about $5 per million tokens read and $30 per million written, which is why OpenAI also sells Terra and Luna at a fraction of that.

The existence of those three tiers is OpenAI quietly telling you the same thing this whole series says: one model at one price is the wrong shape for real work. If you buy AI directly, your bill scales with usage at whatever rate the model you picked charges, so running everything on the flagship means paying flagship rates for work Luna would have done identically. Jyper mixes the tiers for exactly that reason, and can afford to, because Jyper charges on outcomes rather than passing a usage meter on to you.

Why we run it next to Claude instead of picking a winner

Two strong models disagreeing is the cheapest error detection ever invented.

Here is a trick from the way pilots and surgeons work: on decisions that matter, get two independent opinions. Jyper does this with models. On a request worth winning, Claude and GPT-5.6 can each take the same question. Where they agree, confidence is high. Where they disagree, a human looks, and the disagreement itself points at exactly what needs checking.

You could not afford to do this with people. With models it costs almost nothing, and it is a thing you can only do if your platform runs more than one lab’s models in the first place. As always in Jyper, both models draft and structure; the prices in the proposal come from Jyper’s own pricing engine, traced to your contracts.