GPT‑5.6 Launch: Three Model Tiers, New ChatGPT & Codex Features, and How to Choose

Model tiers

On July 9 2026 OpenAI released GPT‑5.6 with three long‑term capability layers:

GPT‑5.6 Sol – flagship model, $5 / M tokens input, $30 / M tokens output.

GPT‑5.6 Terra – balances capability and cost, $2.5 / M tokens input, $15 / M tokens output.

GPT‑5.6 Luna – fastest and cheapest, $1 / M tokens input, $6 / M tokens output.

All three provide a 1.05 M token context window, up to 128 k token output, and accept text and image inputs.

Three pricing systems cannot be directly converted.

Pricing and usage limits

API pricing is per‑million‑token as listed above. Cache reads receive a 90 % discount on the input price; cache writes cost 1.25 × the normal input price, and cache entries live for at least 30 minutes.

Three separate cost structures apply:

ChatGPT subscription (Free, Go, Plus, Pro 5x, Pro 20x, Business).

Codex and ChatGPT Work usage quotas.

API usage billed by token.

Subscription details (prices in USD):

Free – no charge, limited to Terra in Work/Codex.

Go – regional pricing, primarily Terra.

Plus – $20 / month, grants access to Sol; Work and Codex can select any of the three models and includes an Ultra mode that runs four parallel agents.

Pro 5x – $100 / month, ~5 × Plus usage, includes Sol Pro.

Pro 20x – $200 / month, ~20 × Plus usage, includes Sol Pro.

Business – $25 / person / month (annual $20), provides team workspaces and baseline Codex usage.

ChatGPT Work and Codex changes

The desktop clients for ChatGPT and Codex were merged into a single interface with two modes: ChatGPT Work and ChatGPT Codex. ChatGPT Work behaves as a persistent work agent that can ingest files, call tools, and produce editable deliverables such as documents, spreadsheets, presentations, websites, and charts. Users can monitor progress, supply additional information, or change direction without restarting the conversation.

Model availability in Work:

Free and Go users primarily use Terra.

Plus, Pro, Business, and Enterprise users can choose Sol, Terra, or Luna; Plus also offers an Ultra mode that coordinates four parallel agents.

Codex evolves into a full‑featured development‑agent workspace supporting:

Multiple agents working in parallel.

Control of local applications and browsers.

Real‑time rendering checks of front‑end pages.

Image generation.

GitHub review and multi‑terminal interaction.

SSH remote development.

PDF, table, slide, and document preview.

Memory, scheduled tasks, and automation.

90+ plugins and integrations (see Codex official docs [4]).

Benchmark comparisons

OpenAI’s internal evaluation (Agents’ Last Exam, SWE‑Bench Pro, Terminal‑Bench 2.1) shows:

Agents’ Last Exam – Sol 52.7 %, Terra 50.4 %, Luna 50.3 % vs GPT‑5.5 46.9 %.

SWE‑Bench Pro – Sol 64.6 %, Terra 63.4 %, Luna 62.7 % vs GPT‑5.5 59.4 %; Claude Fable 5 80.0 %.

Terminal‑Bench 2.1 – Sol 88.8 %, Terra 87.4 %, Luna 84.7 % vs GPT‑5.5 85.6 %.

OpenAI notes that inference intensity, tool usage, and test environments differ, so the numbers illustrate capability trends rather than absolute rankings.

Pricing comparison (per M tokens): Sol $5 input / $30 output, Terra $2.5 / $15, Luna $1 / $6. Claude Fable 5 costs $10 / $50, and Gemini 3.5 Flash is optimized for speed at $1.5 / $9 [5].

When to choose each model

Architecture design, complex reviews, high‑risk changes → Sol.

Daily feature development, API work, data analysis, routine knowledge work → Terra.

Batch edits, simple sub‑tasks, low‑risk high‑volume operations → Luna.

Large migrations or multi‑day autonomous development → Sol or Claude Fable 5.

High‑frequency rapid coding loops → Luna or Gemini 3.5 Flash.

In practice most developers benefit from a three‑model combo: Sol for high‑failure‑cost tasks, Terra for the bulk of daily work, and Luna for cheap, high‑volume batch jobs.

Additional highlights

Max (alias Extra High) extends reasoning time to explore alternatives.

Ultra runs four parallel agents for faster, more reliable outcomes.

Programmatic tool calling lets GPT‑5.6 write and execute code in‑memory, reducing round‑trips between model and tools.

Computer Use now renders actual pages, enabling layout, spacing, font, and responsive checks.

OpenAI performed extensive red‑team testing, consuming ~700 k A100e GPU‑hours before public release.

Model selection guidance

Select a model based on the cost of failure, task frequency, and risk:

High failure cost → use Sol.

Standard, repeatable tasks → use Terra.

Low‑risk, high‑volume tasks → use Luna.

Example allocation (pre‑formatted for clarity):

Sol: planning, architecture, difficult bugs, final verification
Terra: everyday development, data work, knowledge tasks
Luna: batch processing, simple sub‑tasks

Other notable upgrades

max

(extra high) gives the model more reasoning time. ultra coordinates four agents in parallel rather than simply thinking longer.

Programmatic tool calling reduces token overhead when chaining tools.

Computer Use can visually validate front‑end output.

Red‑team testing involved ~700 k A100e GPU‑hours.