TL;DR: OpenAI released the GPT-5.6 family — Sol, Terra, and Luna — on 9 July 2026. Sol is the new flagship, but Terra is the real story: it beats GPT-5.5 on almost every agentic benchmark at exactly half the price ($2.50/$15 vs $5/$30 per million tokens). I got access on day one through Codex. Here is the full comparison, with the catches OpenAI’s charts do not show.
Last month I wrote about Claude Fable 5 within days of its launch. Now it is OpenAI’s turn to make me drop everything.
OpenAI announced GPT-5.6 for general availability on 9 July 2026, after a limited preview that began on 26 June. This is not one model. It is three — Sol, Terra, and Luna — and picking the wrong one either wastes your money or wastes your time.
My ChatGPT plan gave me all three in Codex on day one, and I have spent since then switching between them on my own projects and digging through every benchmark table OpenAI published. Here is what I found.
What Is GPT-5.6, and Why Are There Three Models Now?
GPT-5.6 is one generation split into three permanent tiers. The number is the generation; Sol, Terra, and Luna are capability tiers that OpenAI says will carry forward into future releases. Sun, earth, moon — flagship, balanced, budget.
- GPT-5.6 Sol (
gpt-5.6-sol) — the flagship. The plaingpt-5.6API alias routes to Sol. Built for long, complex, high-stakes work. - GPT-5.6 Terra (
gpt-5.6-terra) — the balanced tier. OpenAI’s own words: performance competitive with GPT-5.5 at a lower cost. - GPT-5.6 Luna (
gpt-5.6-luna) — the fast, cheap tier for high-volume, repeatable tasks.
All three share the same core specs, straight from OpenAI’s model pages: a 1,050,000-token context window, 128,000 max output tokens, and a 16 February 2026 knowledge cutoff. That context window is up from GPT-5.5’s 1 million.
GPT-5.6 Sol vs Terra vs Luna vs GPT-5.5: Quick Comparison
Before the benchmarks, here is the practical picture in one table:
| GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | GPT-5.5 | |
|---|---|---|---|---|
| Position | Flagship | Balanced default | Fast & cheap | Previous flagship |
| API price (per 1M tokens) | $5 in / $30 out | $2.50 in / $15 out | $1 in / $6 out | $5 in / $30 out |
| Context window | 1,050,000 | 1,050,000 | 1,050,000 | 1,000,000 |
| Best for | Hardest multi-step work, research, security | Everyday coding and knowledge work | Bulk extraction, classification, summaries | Legacy workflows only |
Notice the first thing: Sol costs exactly what GPT-5.5 cost. OpenAI did not raise flagship pricing — it added two cheaper tiers below it. Cache reads keep the 90% discount, cache writes are billed at 1.25x input, and there is now a guaranteed 30-minute minimum cache life.
How Do the GPT-5.6 Benchmarks Compare to GPT-5.5?
These numbers are from OpenAI’s official launch tables, published 9 July 2026. I have picked the rows that actually matter for real work:
| Benchmark | Sol | Terra | Luna | GPT-5.5 |
|---|---|---|---|---|
| Agents’ Last Exam (long-horizon work) | 52.7% | 50.4% | 50.3% | 46.9% |
| Artificial Analysis Coding Agent Index | 80 | 77.4 | 74.6 | 76.4 |
| SWE-Bench Pro (real coding tasks) | 64.6% | 63.4% | 62.7% | 59.4% |
| Terminal-Bench 2.1 (command line) | 88.8% | 87.4% | 84.7% | 85.6% |
| BrowseComp (agentic browsing) | 90.4% | 87.5% | 83.3% | 84.4% |
| GPQA Diamond (PhD-level science) | 94.6% | 92.9% | 92.3% | 93.6% |
| FrontierMath Tier 4 (hardest maths) | 83% | 68.3% | 58.5% | 72.5% |
| MRCR v2 long-context recall (256–512K) | 91.5% | 89.6% | 41.3% | 81.5% |
| ARC-AGI-3 (abstract reasoning) | 7.78% | 0.8% | 0.18% | 0.43% |
| HealthBench Professional | 60.5% | 57.7% | 55.7% | 49.5% |
Three things jump out of this table:
- Sol wins every single row. No surprises — it is the flagship, and Sol Ultra (more on that below) pushes Terminal-Bench to 91.9% and BrowseComp to 92.2%.
- Terra beats GPT-5.5 on nearly every agentic benchmark at half the price. Coding, terminal work, browsing, long-horizon agent tasks, health — all better or equal.
- Luna is shockingly good until it is not. A $1/$6 model that scores 50.3% on Agents’ Last Exam is remarkable. Then look at the long-context row: 41.3%. That is a cliff, and it matters.
Is GPT-5.6 Terra Really Better Than GPT-5.5?
For most work — yes, and this is the headline of the whole launch, even if OpenAI’s marketing puts Sol on the poster.
Terra scores higher than GPT-5.5 on SWE-Bench Pro (63.4% vs 59.4%), Terminal-Bench (87.4% vs 85.6%), BrowseComp (87.5% vs 84.4%), Agents’ Last Exam (50.4% vs 46.9%), and long-context recall (89.6% vs 81.5%) — while costing $2.50/$15 instead of $5/$30. Notion’s co-founder Simon Last put it plainly in OpenAI’s launch post: many of their agents running GPT-5.5 perform just as well on Terra for half the cost and 16% fewer tokens.
Where does GPT-5.5 still hold on? Pure academics. It edges Terra on GPQA Diamond (93.6% vs 92.9%) and beats it on FrontierMath Tier 4 (72.5% vs 68.3%). If your daily work is competition mathematics, keep GPT-5.5. For everyone else, there is no reason left to pay GPT-5.5’s price.
My take: Terra is the model OpenAI should have named the default, and in Codex they effectively did — it is what Free and Go users get. If you were paying for GPT-5.5 via the API yesterday, switching to Terra today is a straight 50% cost cut with a small quality gain. That almost never happens in one release.
Where Does Sol Actually Earn Its Price?
If Terra is this good, why pay double for Sol? Because on the genuinely hard stuff, the gap is not small — it is a different class.
- Abstract reasoning: ARC-AGI-3 at 7.78% vs Terra’s 0.8% — nearly 10x. This benchmark is brutal for every model, and Sol is the only one that registers.
- Hard maths: FrontierMath Tier 4 at 83% vs Terra’s 68.3% and GPT-5.5’s 72.5%.
- Computer use: 62.6% on OSWorld 2.0 — a new state of the art, and OpenAI says it did it using 85% fewer output tokens than Claude Opus 4.8.
- Cybersecurity: 73.5% on ExploitBench vs GPT-5.5’s 47.9% — one of the largest single-generation jumps in the whole launch table.
Sol also brings two new capability settings. max gives the model more reasoning time than the old xhigh setting for a single hard task. ultra runs four agents in parallel by default, splitting a complex task across subagents — that is what lifts Terminal-Bench from 88.8% to 91.9%. Ultra burns roughly 3x the tokens of a single-agent run, so treat it as a special weapon, not a default.
One honest note before OpenAI fans celebrate too loudly: Claude Fable 5 still leads repo-level coding. On SWE-Bench Pro, Fable 5 scores 80% against Sol’s 64.6%, and it narrowly tops the GDPval knowledge-work Elo too. OpenAI’s counter-argument is efficiency — Terra and Luna beat Fable 5 on Agents’ Last Exam at around one-sixteenth the cost. Both claims are true. Pick your fighter based on the work.
How Do You Get Access to GPT-5.6?
Access depends on where you use it, and this part genuinely confused people on launch day. As of 10 July 2026:
- ChatGPT (normal chat): Plus, Pro, Business, and Enterprise users get Sol at medium and higher effort. Pro and Enterprise additionally get Sol Pro for the highest-quality answers. Terra and Luna are not selectable in normal chats, and Free/Go users do not get GPT-5.6 in chat at all.
- ChatGPT Work and Codex: Free and Go users get Terra. Plus and above can choose all three and set a reasoning effort for each. In Codex, type
/modelto switch — same as Claude Code. - API: all three models are live as
gpt-5.6-sol,gpt-5.6-terra, andgpt-5.6-luna. The new Programmatic Tool Calling in the Responses API lets the model run small in-memory programs to coordinate tools, and it is Zero Data Retention compatible.
ultra is available in Codex on Plus plans and above, and in ChatGPT Work for Pro and Enterprise. GPT-5.6 has also already become the preferred model in Microsoft 365 Copilot and landed in GitHub Copilot on the same day.
My First Day Running All Three in Codex
Let me be upfront: GPT-5.6 launched yesterday, so this is a day-one report, not a month-long review. I will update this post as I log more hours.
What I did on day one: switched my Codex sessions between the three tiers on my own work — content tooling for my sites, a research workflow, and a couple of scripted comparisons of the same task run tier by tier.
- Terra feels like GPT-5.5 with the lag removed. For scoped tasks — write this script, fix this function, draft this section — I could not reliably tell Terra and Sol apart. That matches the benchmark story.
- Sol shows up when the task is vague. Give it a messy, underspecified problem and it plans more before touching anything, and its first attempt lands closer to done.
- Luna is the bulk-work model. Renaming, extraction, summaries — instant and dirt cheap. I would not hand it a large codebase though, and the 41.3% long-context score says why.
Interesting detail: when I asked Codex itself which exact model snapshot it was running, it refused to guess — its session only identifies as “Codex based on GPT-5”. Even the tools are humble about model IDs now.
What Is the Catch?
Four things you should know before you move real work to GPT-5.6.
- Luna’s long-context cliff is real. 41.3% on MRCR long-context recall vs Terra’s 89.6%. Luna is priced for volume, not for reading your 300-page contract. Past roughly 200K tokens of context, pay for Terra.
- The 1M context is not flat-priced. Prompts above 272K input tokens are billed at 2x input and 1.5x output for the whole request. Budget accordingly before you dump a monorepo into it.
- The safety layer is stricter — and you will occasionally feel it. OpenAI ran roughly 700,000 GPU-hours of automated red teaming, and Sol’s cybersecurity safeguards now block about 10x more potentially harmful activity than before. Genuine security research needs identity verification through the Trusted Access for Cyber programme, with hardware passkeys mandatory from 1 September. For flagged prompts, ChatGPT and Codex offer a one-click retry on a lower-capability model.
- Independent safety testers flagged Sol’s reward hacking. METR’s pre-deployment evaluation found Sol has the highest reward-hacking rate among public models they have tested — it will sometimes exploit evaluation bugs to “pass” a task. In agent setups, keep your approval gates on. Do not loosen sandboxes just because the model got smarter.
FAQ
What is the difference between GPT-5.6 Sol, Terra, and Luna?
Same generation, three capability tiers. Sol is the flagship for the hardest work ($5/$30 per 1M tokens), Terra is the balanced everyday model ($2.50/$15), and Luna is the fast budget model ($1/$6). All three share a 1,050,000-token context window.
Is GPT-5.6 Terra better than GPT-5.5?
On agentic work, yes — Terra beats GPT-5.5 on SWE-Bench Pro, Terminal-Bench, BrowseComp, and Agents’ Last Exam at half the price. GPT-5.5 keeps a small edge only on academic benchmarks like GPQA Diamond and FrontierMath.
Can free ChatGPT users use GPT-5.6?
Not in normal chat. Free and Go users get GPT-5.6 Terra only inside ChatGPT Work and Codex. In regular ChatGPT conversations, GPT-5.6 Sol requires a Plus plan ($20/month) or higher.
What are the max and ultra modes in GPT-5.6?
max gives a single model extra reasoning time beyond the old xhigh setting. ultra coordinates four parallel agents on one task by default — it lifted Sol’s Terminal-Bench score from 88.8% to 91.9%, but costs roughly 3x a single-agent run.
Is GPT-5.6 better than Claude Fable 5?
Depends on the work. Sol leads on efficiency, browsing, computer use, and long-horizon agent tasks. Claude Fable 5 still clearly leads repo-level coding — 80% vs 64.6% on SWE-Bench Pro — and edges GDPval knowledge work.
What is GPT-5.6’s knowledge cutoff and context window?
All three GPT-5.6 models have a 16 February 2026 knowledge cutoff, a 1,050,000-token context window, and 128,000 max output tokens, per OpenAI’s official model pages.
Summing Up!
GPT-5.6 is less about one smarter model and more about pricing intelligence properly. Sol wins every benchmark it enters, but Terra is the model most people should actually use — GPT-5.5-class-or-better results at half the cost. Luna is a gift for bulk work as long as you keep contexts short.
My recommendation: make Terra your default, escalate to Sol when the task is genuinely hard or vague, and keep Luna for high-volume repeatable jobs. If you are choosing between ecosystems, read my Claude Fable 5 hands-on next — Anthropic still owns repo-level coding. If you are building your own agents, my write-up on Grip AI, my open-source agent platform pairs well with this, and students should check the Google AI Pro student offer before paying for anything.
Which tier have you tried so far — and did Terra surprise you the way it surprised me? Tell me in the comments.
