Overview

GPT-5.6 landed on 9 July 2026 in three tiers. Sol is the coding and reasoning flagship, Terra is the balanced default and Luna is the volume tier. All three take a 1 million token context window on the API and all three accept a reasoning effort setting on every request, which is a cleaner design than maintaining separate reasoning models. Sol adds an Ultra mode that runs four parallel subagents.

Pricing moved hard on 30 July. Luna fell 80% to $0.20 per million input and $1.20 output, Terra fell 20% to $2 and $12, and Sol held at $5 and $30. Batch takes another 50% off and cached prefix reads bill at 10% of rate, so a pipeline with a stable system prompt lands well under the sticker price. On data handling, the API does not train on customer content by default, standard retention is 30 days and Zero Data Retention was extended to frontier models for eligible accounts on 19 August 2026.

The obvious use is agentic work: long tool-calling chains, terminal tasks, browser automation, anything where the model has to stay coherent across dozens of steps. Sol's 88.8% on Terminal-Bench 2.1 shows up in practice as fewer dead-ended runs. Teams doing bulk document analysis on a budget should price Gemini 3.1 Pro at $2 input before committing, and anyone reaching for Luna on long documents should look at its 41.3% MRCR recall first. One practical note for teams inside mainland China: this is not officially served there, so the deployment story runs through overseas infrastructure.

Key Features

✓ GPT-5.6 shipped 9 July 2026 in three tiers: Sol for coding and reasoning, Terra as the balanced default, Luna for volume

✓ 1 million token context window on all three API tiers, with 128K maximum output

✓ Reasoning effort configurable per request from none through to max, plus a Sol Ultra mode that fans out four parallel subagents

✓ Prompt caching at a 90% discount on cached reads with a 30 minute minimum, Batch at 50% off, Flex at half rate and Priority at 2.5x

✓ Built in tools callable from the Responses API: web search, code execution and file search

✓ Zero Data Retention for eligible API customers, extended to frontier models on 19 August 2026

My Take

The 30 July price cut reset the maths. Luna dropped 80% to $0.20 per million input and Terra 20% to $2, which pushes a lot of extraction and classification work from marginal into obviously worth doing. Sol still holds the agentic crown at 92.2% on BrowseComp, and the per-request reasoning dial from none to max is the most useful cost lever in any frontier API. The catch is that the 1M context window is an API property, not a ChatGPT one: Pro reasoning caps at 400K, so the $200 seat does not buy the long context the launch material advertises.