GPT-5.6 Sol, Terra, and Luna: which tier a business should actually buy

GPT-5.6 arrived on July 9, 2026 as a family of three: Luna, Terra, and Sol, smallest to largest, after a limited preview that started June 26. All three carry a million-token context window, 128,000 tokens of output, and a February 16, 2026 knowledge cutoff. The names are the only whimsical thing about the release. The pricing ladder is where the decisions live.

What OpenAI says changed, and what independent testers say

OpenAI's headline claim is agentic: on its Agents' Last Exam benchmark, which scores long-running professional workflows across 55 fields, the company reports Sol at 53.6, which it says beats Claude Fable 5 by 13.1 points, and it positions Terra and Luna as beating Fable 5 at around one-sixteenth the cost. The new ultra mode runs subagents in parallel for long jobs, and the API gained programmatic tool calling, multi-agent orchestration, explicit prompt-cache breakpoints, and an original-resolution image setting.

Read the independent numbers next to the vendor ones. Simon Willison's launch-day review called the models very competent but noted they had not yet passed Claude Fable on the complex coding tasks he runs, and on SWE-Bench Pro the published comparison has Fable 5 at 80 percent against Sol's 64.6. So the honest summary: GPT-5.6 is the strongest OpenAI has shipped, it is priced aggressively at the bottom two tiers, and the top tier is not an automatic win over the competition for code. That combination is exactly why tier choice matters more than brand choice.

Update, September 3, 2026: Anthropic shipped Claude Fable 5.1 on September 1 at the same $10 in, $50 out list prices as Fable 5. On Terminal-Bench 4.0 it scores 55.8 against Sol's 37.3 in the figures Anthropic published and Tech-ish reported. The tier logic below does not change; the competitor at the top of the ladder got stronger, and the case for Luna and Terra on volume work got no weaker.

How to pick a tier without overpaying

• Route by cost of a miss, not by prestige. If a wrong answer costs a re-run, use Luna. If it costs a client, use Sol.
• Put Terra in the middle of every workflow and let the edges argue their way up or down. Most intake, follow-up, reporting, and content operations never need to leave it.
• Cache deliberately. The new cache breakpoints mean a long system prompt or a client knowledge base is paid for once per window, not per call. Design prompts so the stable part comes first.
• Use the million tokens for context, not for laziness. Feeding a model the whole CRM export because you can is how a $15 job becomes a $150 one.

Codex: the agent grew up over the summer

Codex, OpenAI's coding agent, had a busy two months, and the changelog reads like a product finding its shape. Late July brought Sign in with ChatGPT for six connected plugins (Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel) and multi-repository review in a single pass. Coverage in the first days of August added portable agent plugins, persistent conversation sections, automated approvals, and support for the July 28 revision of the MCP protocol. And on July 31 OpenAI announced that the GPT-5.4 models would be retired from Codex for ChatGPT sign-ins on August 31, to be replaced by Terra and Luna, which tells you where OpenAI thinks the agent's default should sit on the ladder.

For a business the interesting words in that list are not the coding ones. Persistent sections and automated approvals are the ingredients of an operator that can be left running on a defined job, with a human gate where money or a customer is involved. That is the same architecture Monolith uses when it installs agents on the client side, whether the harness is Codex, Claude Code, or Hermes: name the inputs, name the tools, name the approval gate, then let it run.

Images 2.0: the part of the release that changes creative budgets

ChatGPT Images 2.0 launched April 21, 2026, model ID gpt-image-2 in the API, and it is the quiet giant of OpenAI's year. Third-party testing reports a 316-point jump in text-rendering scores over the prior generation, which in practice means signage, packaging, UI mockups, and multilingual layouts come out with readable, correctly spelled text. Aspect ratios run from 3:1 to 1:3, output goes to 2K standard, and batches of up to eight images render at once. Paid ChatGPT plans get images with thinking, where the model plans the composition before it draws.

Twenty-one cents for a hero-grade image is not a creative budget, it is a rounding error. The cost has moved to the part that was always the real cost: knowing what to ask for, keeping a brand consistent across two hundred outputs, and saying no to the first draft. That is a creative director's job, and it is what Monolith's AI creative production sprints are built around.