GPT-5.6 Sol: OpenAI's New Flagship, Explained

GPT-5.6 Sol went generally available on July 9, 2026, rolled out simultaneously to ChatGPT, Codex, and the API. It arrived with a new naming scheme worth thirty seconds of your attention, because it finally makes OpenAI's lineup legible: the number is the generation, the name is the capability tier. GPT-5.6 is the generation; Sol, Terra, and Luna are the tiers within it.

Sol is the flagship, the API identifier is gpt-5.6-sol, and it is the only tier that gets the two headline features: max reasoning effort and ultra mode. Where it sits among every current coding model, and against Anthropic's frontier tier, is what the rest of this guide covers.

Data verified July 18, 2026 against OpenAI's published pricing and launch materials. The full family, side by side:

The Terra line deserves a call-out: OpenAI positions it as roughly GPT-5.5 performance at about half the price, which quietly makes it the value pick of the generation, more on that in the decision guide below and in the best Codex model.

Sol's two exclusives are what separate it from Terra beyond raw capability. Max reasoning effort is the highest thinking budget in the family, for problems where you want the model to grind. Ultra mode is more interesting: Sol natively fans a task out to parallel subagents, multiple copies of itself working a problem concurrently. OpenAI's own numbers show what that buys: 91.9% in ultra mode on Terminal-Bench 2.1, up from 88.8% standard, per OpenAI's launch materials.

Note what ultra mode parallelizes: work within one task. It does not parallelize your backlog. Running several independent stories at once, each on its own branch, is an orchestration-layer job, which is exactly the layer AIDEN provides on top of the Codex CLI, see AIDEN vs Codex CLI.

OpenAI led its launch with Terminal-Bench 2.1 and, notably, did not publish SWE-bench scores. Here is the verified picture, including how Sol compares with Claude Fable 5:

Data verified July 18, 2026. OpenAI also claims leads on ALE (53.6) and the Artificial Analysis Coding Agent Index (80.0); treat both as OpenAI's claims until independently reproduced.

In Codex, the current generation is Sol, Terra, and Luna, and Codex auto-selects a model by default, most sessions never touch the picker. Also still selectable: GPT-5.5, GPT-5.4, GPT-5.4 Mini, and GPT-5.3 Codex Spark, a text-only research preview limited to ChatGPT Pro. Two housekeeping notes from OpenAI: gpt-5.2 and gpt-5.3-codex are deprecated for ChatGPT sign-in, and GPT-5.5, the previous flagship, now overlaps almost entirely with Terra at twice Terra's price.

If you have tuned workflows on GPT-5.5, they keep working, but for new work the current generation is the sane starting point. The tier-by-tier picking logic, including reasoning effort, lives in the best Codex model guide.

One more axis: you do not have to pick a vendor at all. AIDEN runs your Codex CLI next to your Claude Code CLI on one kanban board, each story on its own branch, so you can route terminal-heavy stories to Sol and repo-scale refactors to Claude Fable 5, and let the per-story results, not the launch keynotes, decide who gets the next card.