Default to Terra. It lands within 2–3 points of the flagship on almost everything, at half the price, with essentially the same long-context recall.
Never point Luna at a long document. It has a 1M-token window it cannot actually use — see the cliff below.
Sol is the flagship, but it is not the workhorse. That distinction matters and OpenAI's own guidance agrees.
You remembered two of the three names. The full family is Luna (budget), Terra (mid), and Sol (flagship). They arrived on 9 July 2026 — five days before this document was written — after an unusual limited preview in which OpenAI shipped only to trusted partners first, at the request of a US-government safety review.
OpenAI has thrown away mini and nano. In their words: the number identifies a model's generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence.
That last clause is the point. Under the old scheme, gpt-5.4-mini could not ship until gpt-5.4 shipped — the tiers were welded to the generation. Now Terra can get a mid-cycle upgrade without forcing a rename of the whole family. The suffixes also described size rather than capability, which flattered nobody: "nano" implied a toy, and Luna is not a toy.
| Luna | Terra | Sol | |
|---|---|---|---|
| API model ID | gpt-5.6-luna | gpt-5.6-terra | gpt-5.6-sol |
| Alias | — | — | bare gpt-5.6 routes here |
| Context window | 1.05M | 1.05M | 1.05M |
| Max output | 128K | 128K | 128K |
| Input / 1M tokens | $1.00 | $2.50 | $5.00 |
| Output / 1M tokens | $6.00 | $15.00 | $30.00 |
| Cache read | 90% discount across all three · cache write 1.25× input · 30-min minimum cache life | ||
| Reasoning effort | none · low · medium · high · xhigh · max — six levels, all tiers | ||
| Modalities | text + image in → text out | ||
| Knowledge cutoff | 16 February 2026 | ||
| Intelligence Index | 51 | 55 | 59 |
| Cost per task | $0.21 | $0.55 | $1.04 |
| Coding Agent Index | 74.6 | 77.4 | 80.0 |
| Terminal-Bench 2.1 | 84.7% | 87.4% | 88.8% (91.9% Ultra) |
| Agents' Last Exam | 50.3 | 50.4 | 53.6 |
| MRCR long-context | 41.3% | 89.6% | 91.5% |
| Built for | High-volume: classification, summarising, drafting, routine automation | The default. Everyday interactive work and agentic coding | Long-horizon agents, large-codebase reasoning, security research |
Luna scores 41.3% on MRCR long-context recall. Sol scores 91.5%.
Luna advertises the same 1.05M-token context window as its siblings — and cannot use it. Feed it a long document and it will confidently lose the middle of your prompt. This is the sharpest capability cliff anywhere in the family, and OpenAI does not foreground it. Luna is for short, well-scoped, high-volume work. Nothing else.
GPT-5.6 adds a sixth reasoning level, max, sitting above xhigh. Separately — and confusingly — there is Ultra, which is not a reasoning-effort value at all. Ultra is an execution mode that spawns four subagents in parallel and merges their results.
max = think harder, in one chain.Cost scales the way you'd fear: Ultra runs roughly 3× a single Max call, because token spend stacks across every subagent, and it buys about three points on Terminal-Bench. Reach for it rarely.
When you move off 5.5, keep your current effort as a baseline and then test one level lower. GPT-5.6 reaches frontier quality using fewer output tokens, so whatever effort level you tuned for 5.5 is probably now overkill — you'd be paying for reasoning you no longer need.
Artificial Analysis measured Sol at max effort at 67.7 output tokens/sec with a ~131-second time-to-first-token. That is not network latency — it is reasoning time, and it is what max costs you in wall-clock. If you are running Sol at max inside an interactive loop, that pause is the model thinking, and you should expect it.
Per-tier speed figures for Terra and Luna have not been published by anyone. Various blogs quote confident numbers ("Sol ~180ms TTFT, Terra ~320ms") — these are internally contradictory and almost certainly fabricated. Ignore them. A Cerebras-hosted Sol at up to 750 tok/s launched in July 2026 if you need a genuinely low-latency tier.
This is the most important section in the chapter, because it is the one place where the marketing and the evidence come apart — and it happens to land directly on what you do all day, which is agents working inside real repositories.
| Benchmark | Sol | Terra | Luna | Claude Fable 5 |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.8% (Ultra 91.9%) | 87.4% | 84.7% | 86.0% |
| Coding Agent Index | 80.0 | 77.4 | 74.6 | 77.2 |
| Agents' Last Exam | 53.6 | 50.4 | 50.3 | 40.5 |
| ExploitBench 1 | 73.5% | — | — | — |
| SWE-Bench Pro | 64.6% | — | — | 80.0% |
SWE-bench Verified. GPQA Diamond. AIME. MMLU. ARC-AGI-2. FrontierMath. All of them were omitted from the launch table — every one of the traditional academic benchmarks that defined frontier comparison right up through GPT-5.5.
That omission is the tell. OpenAI leads on the agentic benchmarks and trails on several academic ones, so the former were foregrounded and the latter quietly dropped.
Sol scores 64.6% on SWE-Bench Pro. Claude Fable 5 scores 80%. That is a fifteen-point loss on the single benchmark that most closely resembles "can this thing work inside my actual repository."
The day before announcing GPT-5.6, OpenAI published a critique of SWE-Bench Pro arguing that roughly 30% of its tasks are broken and advising developers to scrutinise results from it.
As Vellum put it, and it is hard to improve on: when a lab leads on a benchmark, they cite it; when they lose on it, they question its methodology.
The independent corroboration matters here. Simon Willison — who has no stake in either company — reports that Sol "hasn't struck me as better than Fable" on complex coding tasks, despite Sol's wins on the benchmarks OpenAI chose to show. His hands-on impression matches SWE-Bench Pro, not the launch table.
You build web apps with coding agents all day. On repo-level coding — the thing you actually do — the independent evidence still favours Claude Fable 5 over Sol, and OpenAI's response to that evidence was to attack the ruler rather than the result. That is worth knowing before you switch anything.
None of the above means Sol is weak. On ARC-AGI-3 it is the only performant model in existence as of July 2026, and the first ever to win a public game (87% on ft09). Its verified ARC-AGI-2 score at max effort is 92.5%. It hits 94.6% on GPQA Diamond and 89.0% on FrontierMath v2.
And on Artificial Analysis's Intelligence Index, Sol at max scores 59 — one point behind Claude Fable 5 at 60, but at roughly a third of the cost per task ($1.04 vs ~$3.00). Treat the top of that leaderboard as a statistical tie with a large price gap in Sol's favour.
On ARC-AGI-1, Sol peaks at xhigh (97.5%) and gets slightly worse at max (96.5%). More reasoning is not monotonically better. If you reflexively crank effort to maximum, you are sometimes paying more for a worse answer.
You didn't ask about this, but it resolves a genuine confusion: GPT-5.3 does exist. I assumed it didn't, and I was wrong.
What actually happened is that the Instant / Thinking / Codex lines stopped versioning in lock-step. GPT-5.3-Codex shipped in February 2026 and GPT-5.3 Instant in March — but no 5.3 flagship "Thinking" model was ever released. The March launch paired 5.3 Instant with 5.4 Thinking. So the version number is very much in use; it just never had a frontier model attached to it.
| Version | Date | What changed |
|---|---|---|
| GPT-5 | Aug 2025 | Unified system — fast base model plus a reasoning layer behind a real-time router |
| 5.1 | Nov 2025 | Adaptive reasoning: dynamic compute allocation, 2–3× faster on easy queries. Warmer default tone |
| 5.2 | Dec 2025 | Aimed at professional knowledge work. First model past 90% on ARC-AGI-1 |
| 5.3 | Feb–Mar 2026 | Codex + Instant only, no flagship. Instant jumped to 400K context, 26.8% fewer hallucinations with web search |
| 5.4 | Mar 2026 | Native computer use, 1M-token context, tool search |
| 5.5 | Apr 2026 | Big jump in agentic coding and long-horizon work. More expensive, but far more token-efficient |
| 5.6 | Jul 2026 | Luna / Terra / Sol. Programmatic tool calling, multi-agent beta, max effort, Ultra mode |
There is no longer a separate reasoning line. Every active ChatGPT model belongs to the GPT-5 family; reasoning_effort is the o-series now, absorbed into the parameter. (There was never an o4 flagship — only o4-mini.)
OpenAI maps o3 → gpt-5.5 and o3-pro → gpt-5.5-pro. But o3 was a dedicated reasoning model with its own thinking paradigm, and gpt-5.5 is a chat-completion model with configurable reasoning. A drop-in ID swap is not behaviourally equivalent. Validate the behaviour; don't just diff the cost and latency and assume you're done.
23 July 2026 — gpt-5-chat-latest, gpt-5.1-chat-latest, all gpt-5.1-codex variants, legacy audio/realtime models, and the deep-research variants. If any of those strings appear in your code, that is a this-week problem.
23 October 2026 — the GPT-4 family, gpt-4-turbo, o1, o3-mini, o4-mini, gpt-3.5-turbo.
11 December 2026 — the original GPT-5 snapshots (gpt-5-2025-08-07 and its mini/nano/pro siblings) and o3 itself.
The structural change worth knowing: there is no gpt-5.6-codex model. The dedicated Codex fork is dead. Codex ran its own models from gpt-5-codex (Sept 2025) through 5.3-codex (Feb 2026), and then stopped — from GPT-5.5 onward, the mainline models ship straight into Codex. You now pick Sol, Terra, or Luna inside Codex directly.
On the same day GPT-5.6 launched, Codex was merged into the ChatGPT desktop app — it is no longer a separate application. PR review moved inside it, inline diff editing arrived, and Codex Remote went GA (start work on a connected Mac, approve actions from your phone). OpenAI's own framing is that Codex has gone from a coding tool to "a broader workspace for getting work done with AI."
If that sounds familiar, it should — it is a direct answer to the Claude Code desktop app, and we will compare them properly in Chapter Four.
Stated plainly, because a guide that hides its gaps is worse than useless:
gpt-5.6-sol. The evidence points to a mode rather than a model, but I could not confirm the parameter name.gpt-5.6-terra, not Sol. Half the price, near-identical long-context recall, within a few points on everything else.gpt-5.1-codex, gpt-5-chat-latest, gpt-5.1-chat-latest before 23 July.Sources. OpenAI model docs and deprecations pages · OpenAI GPT-5.6 launch · Artificial Analysis (Intelligence Index, cost/task, measured throughput) · Vellum (benchmark analysis) · ARC Prize verified results · Simon Willison · TechCrunch · GitHub Copilot changelog. Figures current as of 14 July 2026 and will move; the live version of this chapter is kept updated.