AI Field Guide
Chapter One

Sol, Terra & Luna

OpenAI's GPT-5.6 family, taken apart tier by tier — what the benchmarks say, what OpenAI didn't publish, and which one you should actually point your agents at.
Compiled 14 July 2026 · GPT-5.6 released 9 July 2026 · every figure sourced

Chapter OneThe three tiers


The short version

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.

The naming scheme is the actual news

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.

The comparison table

 LunaTerraSol
API model IDgpt-5.6-lunagpt-5.6-terragpt-5.6-sol
Aliasbare gpt-5.6 routes here
Context window1.05M1.05M1.05M
Max output128K128K128K
Input / 1M tokens$1.00$2.50$5.00
Output / 1M tokens$6.00$15.00$30.00
Cache read90% discount across all three · cache write 1.25× input · 30-min minimum cache life
Reasoning effortnone · low · medium · high · xhigh · max — six levels, all tiers
Modalitiestext + image in → text out
Knowledge cutoff16 February 2026
Intelligence Index515559
Cost per task$0.21$0.55$1.04
Coding Agent Index74.677.480.0
Terminal-Bench 2.184.7%87.4%88.8% (91.9% Ultra)
Agents' Last Exam50.350.453.6
MRCR long-context41.3%89.6%91.5%
Built forHigh-volume: classification, summarising, drafting, routine automationThe default. Everyday interactive work and agentic codingLong-horizon agents, large-codebase reasoning, security research
The one number that will bite you

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.

"Max" and "Ultra" are not the same knob

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.

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.

Counterintuitive migration advice — from OpenAI

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.

A caution on the speed numbers

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.

Chapter OneThe benchmark that OpenAI attacked


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.

What OpenAI published

BenchmarkSolTerraLunaClaude Fable 5
Terminal-Bench 2.188.8% (Ultra 91.9%)87.4%84.7%86.0%
Coding Agent Index80.077.474.677.2
Agents' Last Exam53.650.450.340.5
ExploitBench 173.5%
SWE-Bench Pro64.6%80.0%

What OpenAI did not publish

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.

The SWE-Bench Pro incident — read this one twice

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.

What this means for you specifically

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.

Where Sol genuinely is the best model in the world

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.

A detail worth internalising

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.

Chapter OneThe lineage, and what happened to 5.3


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.

VersionDateWhat changed
GPT-5Aug 2025Unified system — fast base model plus a reasoning layer behind a real-time router
5.1Nov 2025Adaptive reasoning: dynamic compute allocation, 2–3× faster on easy queries. Warmer default tone
5.2Dec 2025Aimed at professional knowledge work. First model past 90% on ARC-AGI-1
5.3Feb–Mar 2026Codex + Instant only, no flagship. Instant jumped to 400K context, 26.8% fewer hallucinations with web search
5.4Mar 2026Native computer use, 1M-token context, tool search
5.5Apr 2026Big jump in agentic coding and long-horizon work. More expensive, but far more token-efficient
5.6Jul 2026Luna / Terra / Sol. Programmatic tool calling, multi-agent beta, max effort, Ultra mode

The o-series is dead — and the migration path is a trap

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.)

If you have anything running on o3

OpenAI maps o3gpt-5.5 and o3-progpt-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.

Deprecations — one of these is nine days out

23 July 2026gpt-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.

Chapter OneCodex ate the workspace


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.

Chapter OneWhat I could not verify


Stated plainly, because a guide that hides its gaps is worse than useless:

Chapter OneSix things to actually do


  1. Default your agents to gpt-5.6-terra, not Sol. Half the price, near-identical long-context recall, within a few points on everything else.
  2. Keep Claude in the loop for repo-level coding. SWE-Bench Pro 80% vs 64.6% is not a rounding error, and the independent hands-on reports agree with the benchmark rather than the launch table.
  3. Never point Luna at a long context. 41.3% recall. It has a window it cannot use.
  4. When migrating off 5.5, try one reasoning level lower. OpenAI recommends this themselves.
  5. Grep your code for gpt-5.1-codex, gpt-5-chat-latest, gpt-5.1-chat-latest before 23 July.
  6. Spike on programmatic tool calling. The model writes JavaScript that runs in a sandboxed V8 with no network access, orchestrating tool calls in code instead of one round-trip per call. Of everything in 5.6, this is the feature most likely to change how you actually build agents.

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.