What Is GPT-5.6? The Sol, Terra, and Luna Tiers, the Codex App Merger, and How It Feels to Use

On 10 July 2026, OpenAI’s newest model, GPT-5.6, arrived. The Codex app was merged into the ChatGPT app at the same moment, Claude answered with an update of its own, and competition in the AI industry stepped up another notch. This article assembles the picture from official information, the wider reaction, and my own use.

What GPT-5.6 is

GPT-5.6 is the newest generation of OpenAI’s GPT series. A limited preview began on 26 June 2026, and general availability followed on 10 July. It serves as the brain behind ChatGPT and as the engine for Codex, the developer-facing agent.

It is a milestone release in which a generational change of model and a reorganization of the apps, discussed below, happened together.

The headline feature: three tiers, Sol, Terra, and Luna

GPT-5.6 comes in three model tiers for different purposes. In plain terms, from the official announcement:

• Sol: the flagship. For long-running coding and agent work that needs planning, tool use, and persistence. For jobs where accuracy and completeness come first
• Terra: the balanced tier. Said to match the previous generation, GPT-5.5, at about half the cost, and intended as the workhorse for everyday tasks
• Luna: light, fast, cheapest. For high-volume work where the task is well defined, while scoring close to the flagship of a short while ago

API pricing per million tokens is $5 input and $30 output for Sol, $2.5 and $15 for Terra, and $1 and $6 for Luna. The three-step arrangement — top tier for hard work, balanced tier for daily use, fast tier for simple volume — is becoming an industry standard shared with other vendors’ line-ups.

The reaction: praised for agent performance and value

On benchmarks, Sol leads the agent-oriented tests (88.8% on Terminal-Bench 2.1, 92.2% on BrowseComp), and on composite measures of general capability it comes within about a point of Claude’s top model, Fable 5, at roughly a third of the cost. That combination of performance and value drew most of the attention.

Ease of migration was also well received. Terra’s position — GPT-5.5 quality at half the price — makes it easy for existing users to move across. On the other hand, more tiers has brought some uncertainty about which one to use.

The Codex app merged into the ChatGPT app

The other change on the day GPT-5.6 arrived was the disappearance of the standalone Codex app for developers, merged into the ChatGPT app.

I had found the Codex app unsatisfying for a while and mostly used the CLI, but the merged ChatGPT app is considerably more comfortable. Chat and agent work completing in one app brings it closer to Claude’s desktop app, which got there first.

Claude updated the same day

What is interesting is that Claude moved on the same day. Fable 5’s allowance was reset — rumored to be a response to GPT-5.6 — and the desktop app merged chat with Cowork, its agent feature. Cowork has also been announced for mobile. The era in which chat was the only point of contact with an AI is ending.

My impression: Sol reads between the lines

I have been using Sol heavily since release. In one sentence, Sol has the between-the-lines intelligence that runs through Fable 5. The number of occasions where it takes the intent from a loose instruction and produces something good has clearly increased.

It arrived toward the end of the period when Fable 5 was available within a subscription, which makes my honest assessment this: an adequate successor to Fable 5, and in places beyond it. How to divide work between Claude and Codex (GPT-5.6) is covered in detail in sorting out the division of labor between Fable 5 and GPT-5.6.

Where to use what: intensifying competition favors users

The rule of thumb is simple. Sol for hard work and long agent runs, Terra for everyday writing and research, Luna for routine high-volume processing. For ordinary ChatGPT use an appropriate model is assigned automatically, so newcomers need not memorize the tier names.

New models, merged apps, loosened allowances — the vendors pushing against each other translates directly into convenience for users. Hold on to two threads, that every vendor now has tiers by purpose and that chat and agents are converging, and AI news from this period becomes much easier to read.

For the GPT series itself, and for underlying terms such as model and LLM, see the AI glossary on this site.

Performance and pricing information in this article is as of 17 July 2026. Check OpenAI’s own announcements for current details.

This article is adapted and expanded for AI Jiten from testing notes originally published on the author’s blog, Delaymania.