GPT 5.6: What Shipped, How to Access It, and What It Costs

GPT 5.6 is not one model. It is a family of three, launched on 9 July 2026, and the one you should pick depends entirely on how much of the work is actually hard.

See Synexa → Official site

Three models, one launch

The 9 July 2026 post announces general availability of the family after a limited preview, and names all three variants in a single sentence: Sol as the new flagship, Terra as a balanced model for everyday work, and Luna as the most cost-efficient. That structure is the most useful thing on the page, because it means the interesting decision is routing rather than adoption. Sol is described as achieving state-of-the-art results across coding, knowledge work, cybersecurity and science while using fewer tokens than previous and competing frontier models. The framing throughout is performance per dollar: either more finished work for the same spend, or comparable work for less. The separate developer documentation carries its own page for the Luna variant, which is where the callable identifier lives.

What ultra actually does

Alongside the three models, the release introduces a setting called ultra, described as the highest-capability option, which coordinates multiple agents across parallel workstreams so that complex tasks finish faster. Read that carefully: it is a way to spend more compute on demand, not a fourth model. The post pairs it with two other claims about Sol, stronger computer use and better design judgment, and argues the combination makes it better at inspecting and refining its own output rather than handing back a first draft. The page sections back this up in their titles, with one on end-to-end knowledge work and another on cyber and science. If your workload is a single short prompt, none of this changes anything for you. If it is a half-day research task, it is the whole point of paying for the flagship.

The number the announcement leads with

One benchmark is quoted prominently: on Agents' Last Exam, an evaluation of long-running professional workflows spanning 55 fields, Sol sets a new high of 53.6, which the post says eclipses a competing frontier model. Two things are worth holding on to. First, it is an agentic workflow benchmark rather than a knowledge quiz, so it maps better than most scores to the long, tool-using tasks people actually hand these models. Second, it is a vendor-reported figure in a vendor announcement, which is normal and still worth remembering when you are comparing against someone else's marketing page. Nothing else on the captured page publishes a second score, and no other benchmark number appears here for that reason.

Pricing moved twice after launch

The announcement carries two dated updates above the fold, and both are price reductions rather than feature news. On 30 July 2026 the price of Luna was cut by 80 percent and Terra by 20 percent. On 21 August 2026, the API and credit pricing of Sol was reduced by over 20 percent, explicitly for the following three months. That last detail matters more than the percentage: a time-boxed reduction is a promotion, so a budget built on it should assume the earlier rate returns. No absolute per-token figures appear anywhere on the captured page, which is why none are quoted here. The developer pricing page is the only place to read the current rate for the identifier you plan to call.

It is a mature release, not the newest one

The page now opens with a note stating that this post introduced the family in 2026 and pointing readers to GPT-6 as the newer model. That changes how you should read everything below it. Older flagship models frequently stay available, get cheaper, and remain the sensible default for production work long after a successor lands, so the banner is not a reason to migrate on its own. It is a reason to check the deprecation page before you standardise on an identifier. And if the work in front of you is generating images, video or audio rather than text, the whole naming question is beside the point: Synexa exposes those models through one REST endpoint and a Python SDK, billed per run, on a release cycle that has nothing to do with this one.

What the release actually contains

Text family versus hosted generation models

Getting on it without guessing

1. Decide the tier before the model
   Route easy and high-volume calls to the cost-efficient variant and reserve the flagship for work that genuinely needs it. The family exists to make that split possible.
2. Copy the identifier from the docs
   The developer documentation carries a model page per variant. Take the exact string from there, because the friendly name and the callable identifier are rarely identical.
3. Read the current rate yourself
   Two reductions landed after launch and one of them is explicitly a three-month window. Open the pricing page before you build a forecast on a percentage you read in a blog post.
4. Check the deprecation page
   The announcement now points to a newer flagship. Older models usually stick around, but confirm the retirement schedule before standardising on one for production.

FAQ