OpenAI's GPT-5.6 Lands: Sol, Terra, Luna and a New Coding State of the Art
OpenAI launched GPT-5.6 — Sol, Terra and Luna. Sol claims a new coding state of the art on the Artificial Analysis Coding Agent Index with far fewer tokens; Terra and Luna push frontier-class coding down to $6–$15 output. Here's how to route between them.
Table of contents
OpenAI has launched GPT-5.6, a three-model family aimed squarely at the work developers actually do: coding, long-horizon agentic tasks, and enterprise knowledge work. After a limited preview on June 26, the models went public on July 9, 2026, across ChatGPT, Codex, and the OpenAI API. The headline for anyone shipping code with AI: OpenAI is claiming a new coding state of the art — and, unusually, it's backing the claim with an independent benchmark.
The three variants
The family splits into three tiers, named for a sun-earth-moon motif:
| Model | Role | Price (input / output per 1M tokens) |
|---|---|---|
| Sol | Frontier reasoning + long-horizon agentic work; the coding workhorse | $5 / $30 |
| Terra | Balanced everyday model, GPT-5.5-class at ~2× lower cost | $2.50 / $15 |
| Luna | Fastest, most affordable | $1 / $6 |
The tiering matters more than it looks. Most day-to-day coding — small refactors, test generation, boilerplate — doesn't need frontier reasoning, and paying Sol's $30/M output for it is waste. Terra and Luna are the models you'll wire into high-volume, latency-sensitive loops; Sol is what you escalate to for the hard, multi-step agentic runs.
The coding claim
OpenAI calls Sol its "best coding model yet," and points to the Artificial Analysis Coding Agent Index, where it reports Sol "sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less." OpenAI also says Sol is 54% more token-efficient on AI coding tasks than its predecessor.
Two things stand out for a working developer. First, the efficiency framing — fewer output tokens and less wall-clock time — is arguably more valuable than the raw index point: agentic coding costs are dominated by output tokens and by how long a run ties up your loop. Second, the comparison is against a live rival (Anthropic's Fable 5) on a third-party index rather than an OpenAI-only eval, which is a stronger signal than the usual launch-day bar chart. Treat "best yet" with the normal grain of salt, but the numbers are more checkable than most.
Down the stack, OpenAI positions Terra above Fable 5 and Luna as outperforming Opus 4.8 — i.e. the budget model is pitched against last generation's frontier. If that holds up in real repos, the interesting story isn't the top of the range; it's how cheap "good enough for most tasks" just got.
Beyond code: agents, security, and Office work
Sol is built for long-horizon agentic work — the multi-step, tool-using runs where a model plans, calls tools, reads results, and keeps going. OpenAI also calls it its "strongest cybersecurity model yet," reaching frontier performance with far fewer tokens on defensive tasks: threat modeling, code review, patching, and blue-teaming. For teams already using AI in code review, that's a natural extension of the same loop.
Alongside the models, OpenAI shipped ChatGPT Work — a workplace surface for drafting documents, spreadsheets, and presentations across desktop, web, and mobile — signaling the family isn't only for engineers but for the knowledge work around them.
What to actually do with it
- Default to Terra or Luna for your agent's routine steps; escalate to Sol only for the hard planning/reasoning turns. The price gap ($6 vs $30 output) rewards routing.
- Watch output-token budgets, not just the index. Sol's pitch is doing more with fewer tokens — the win shows up on your bill and your run time, so measure both before and after.
- Re-benchmark on your own repo. A coding-agent index is a signal, not your codebase. Run your existing eval/harness against Terra and Sol before switching a production loop.
- Consider Sol for security review passes if you already run AI over diffs — the same run can do threat modeling and patch suggestions.
The takeaway
GPT-5.6 is less a single model than a routing decision. Sol takes the state-of-the-art coding crown (on a third-party index, with a genuinely useful efficiency story), but the quieter news is Terra and Luna pushing frontier-class coding down to $6–$15 output. For anyone building agentic workflows, the smart move isn't "switch everything to the best model" — it's wiring the cheap tiers into the common path and saving Sol for when the task earns it.
