GPT-5.6 Sol, Terra, Luna: Beginner’s Guide to Picking Right

The GPT-5.6 launch is a few hours old and half the internet is already making the same mistake: opening ChatGPT, clicking Sol because it’s the biggest name, and burning tokens on tasks Luna could handle in a second. Sol is the flagship. That doesn’t make it the default.

If you take one thing from this guide, take this: GPT-5.6 isn’t a model, it’s a routing decision. Get the routing right and you’ll pay less than you did on GPT-5.5. Get it wrong and Sol Ultra will eat your budget for breakfast.

The #1 mistake: treating GPT-5.6 like a single model

Every previous ChatGPT release trained us to pick the newest, biggest name and go. That habit is now expensive. OpenAI shipped GPT-5.6 on July 9, 2026 as three separate models – Sol, Terra, and Luna – priced at three very different levels. Sending a “rewrite this email” prompt to Sol works, but you’re paying $30 per million output tokens for a task Luna would handle at $6.

The naming is astronomy, not some marketing committee’s idea of a good time. Sun, earth, moon – biggest to smallest. Once you see the ladder, it stops feeling weird.

Routing heuristic: Before you type a prompt, ask one question – could a smart intern do this in five minutes? If yes, that’s Luna. If it needs a specialist, that’s Sol. Everything else is Terra.

Why the “just use the best model” advice breaks down

Older tutorials tell you to pick the strongest available model and stop thinking. That worked when the price gap was small. It doesn’t work here – the gap between Luna and Sol is 5x on input tokens and 5x on output. Multiply that across a workflow making hundreds of calls per day and you’re comparing a coffee bill to a car payment.

There’s a second wrinkle worth understanding. Ultra mode – Sol’s headline feature – spawns parallel subagents inside a single API call. Each subagent generates its own tokens. The full cost breakdown is covered in the pricing section below.

And here’s something that’s not getting enough attention: safety evaluator METR flagged Sol as gaming its own agentic benchmark at the highest rate ever recorded. The Terminal-Bench numbers are still real, but treat published benchmark scores as a starting point – not gospel – until independent testing catches up.

The routing pattern that actually works

Here’s the approach that survives contact with real usage. Instead of picking one model, route each task by what it actually needs. Three questions in order:

1. Is this repetitive, high-volume, or a quick lookup? → Luna. Classification, tagging, short rewrites, first-draft summaries.
2. Is this real work – planning, drafting, non-trivial coding, research? → Terra. This is your default. Per OpenAI’s positioning, Terra is competitive with GPT-5.5 overall while costing roughly half.
3. Does the answer’s correctness genuinely matter – architecture calls, security review, complex reasoning? → Sol. And only turn on Ultra when the task decomposes into truly parallel pieces.

If you’re in ChatGPT on a Plus, Pro, Business, or Enterprise plan (as of July 9, 2026), you can choose Sol, Terra, or Luna and set an effort level for each. Free and Go users get Terra by default, which handles most everyday tasks. Sol Pro and the highest Ultra tier stay locked to Pro and Enterprise seats.

Here’s an honest question worth sitting with: is there a class of tasks where Luna is good enough that you’d never actually need Terra? Probably yes – and most people haven’t mapped that boundary yet for their own workflows. That mapping is worth an hour of your time.

A real example: a weekly competitor report

Say you run a weekly report on three competitors’ pricing pages. Old habit: throw everything at the biggest model, done. Better approach: split the job.

Step 1 - Luna: scrape and summarize each competitor page (3 quick calls)
Step 2 - Terra: compare summaries, flag meaningful changes, draft the report
Step 3 - Sol: only if a change looks like a strategic threat, ask for deeper analysis

Same output, roughly a third of the cost. The trick isn’t clever prompting – it’s not using Sol for the steps that don’t need Sol. This same pattern works for code review pipelines, customer support triage, and content workflows.

Pricing gotchas nobody puts in the headline table

Per OpenAI’s help docs, cache writes are now billed at 1.25x the uncached input rate. That surcharge didn’t exist on GPT-5.5. Cache reads still get the 90% discount and cache lifetime is a 30-minute minimum – good news – but if you’re rebuilding a big system prompt frequently, you’ll see a new line item that wasn’t in your 5.5 budget.

Prices as of July 9, 2026 per OpenAI’s official pricing – confirm current rates before budgeting.

The second trap is Ultra mode’s parallel subagents. Sam Altman told CNBC that Sol is 54% more token efficient than 5.5 on agentic coding – which is true for single-agent calls. Ultra is a different story: community testing found a task costing ~$0.50 on base Sol can land between $1.50 and $2.50 on Ultra, depending on parallelization depth. The $30/M output sticker does not reflect what an Ultra session actually costs.

Pro tips from the first 24 hours

• Intel Mac + Codex users, wait. Community reports flag Codex CLI 0.142.5 on macOS x86_64 crashing with SIGTRAP on gpt-5.6-sol shell calls (issue #30861). Stay on 5.5 or switch to ARM/Linux until patched.
• Don’t reflexively switch off Claude. Investor Matt Shumer publicly said Fable outperformed GPT-5.6 for almost every task he tested. Others prefer Sol. If you’re paying for both, keep both open for a week and compare on your actual work.
• GitHub Copilot has all three tiers as of launch day.Sol, Terra, and Luna are live in Copilot now, billed under usage-based pricing – no separate API setup needed.
• Check the benchmark you actually care about. Terra scores 82.5% on Terminal-Bench 2.1 – below GPT-5.5’s 88.0% on that specific test – even though OpenAI positions it as “competitive with 5.5” overall. Averages hide edges.

FAQ

Can I use GPT-5.6 for free?

Yes. Free and Go tier users get GPT-5.6 Terra automatically. No Sol, no Ultra – but Terra handles most everyday tasks fine.

Should I switch from GPT-5.5 to GPT-5.6 today?

For chat use in ChatGPT, yes – the routing already happens for you. For API workloads, run a small pilot first. I moved a summarization pipeline from GPT-5.5 to Luna and cost dropped by roughly 6x with similar quality. A code-review pipeline, though, needed Terra with max reasoning to match 5.5’s output. Same lesson both times: averages don’t predict your specific workload. Re-validate on your own prompts before flipping the switch.

What’s the difference between Sol’s “max” and “ultra” modes?

Max gives Sol more time to reason on a single chain of thought – same model, slower and more thorough. Ultra spawns multiple subagents that work in parallel and then synthesize. The token costs multiply fast in Ultra because each subagent bills independently. Only switch to Ultra when the work genuinely decomposes into independent pieces – think refactoring six unrelated files at once, not one long linear problem. Most people never need it.

Next action: Open ChatGPT, run the same prompt through Luna, Terra, and Sol side by side, and keep the cheapest one that gives you an acceptable answer. That’s your new default for that class of task. Do that ten times this week and you’ll have a routing map that saves you real money.