QWE AI Academy By the end of this you’ll have Plus/Pro ChatGPT set up so the GPT-5.6 Sol reasoning slider sits at the right position for your work – and if you’re on Free, you’ll know exactly what Luna can and can’t do before you rely on it. No news recap. Just the setup.
OpenAI dropped this in early August 2026: Sol gets a reasoning slider for Plus and Pro users, Luna becomes the free default with unlimited text chats. Community reactions on X landed in two camps – praise for Sol at medium effort on coding tasks, frustration that the reasoning-level naming is confusing and the controls feel hidden, per a Digg summary of X reactions. So here’s what actually matters.
The end state you’re aiming for
Plus or Pro: one model (Sol) handles everything. You use the slider to dial reasoning up or down per task. Free: Luna is your default, Think button is there for harder questions – but it’s not a free pass. Use it selectively.
The split OpenAI is shipping, as of the official announcement: Plus and Pro get Sol updated for reliability and focused answers, with the new slider. Free and Go get Luna as the default – unlimited text chats – and the Think button rolls out the following week.
Set up GPT-5.6 Sol on Plus or Pro
Working backwards from a slider you actually understand:
- Open ChatGPT on web, desktop, or mobile (rolling out across all platforms, per the official announcement).
- Start a new chat. The model picker should already show GPT-5.6 Sol. No manual switch needed – Sol now handles both quick replies and longer reasoning through one model instead of two separate Instant and Thinking modes.
- Find the slider near the model name. In Chat it’s a simpler control than the full Codex ladder.
- Leave it at the default for your first few messages. Test a familiar task before adjusting. There’s no clean mapping from GPT-5.5 effort levels to this one.
- Push it up only when the answer feels thin. Sol got it right? Stay there. It hedged or skipped a step? Bump one notch and retry.
The naming across OpenAI’s surfaces is genuinely a mess. Turns out the API exposes six reasoning effort levels. In ChatGPT Work and Codex, the same idea shows friendlier labels: Light, Medium, High, Extra High, and Max – per the ChatGPT Learn docs. The consumer Chat slider is the simplified version of that same ladder. App’s Extra High = API’s xhigh. Max sits one rung above that.
Pro tip: Treat the slider like a router, not a quality dial. Ask “how expensive is a wrong answer here?” – Slack draft: keep it low. Migration decision or legal summary: push it up. Don’t max it by default – you’ll wait longer and get more verbose output for no reason.
What Free users should actually do with Luna
Free tier setup is simpler – no slider, just the model and a Think button.
- Default chats: Just type. Luna handles everyday questions, drafting, summaries. Luna replaces GPT-5.5 Instant as the free default (as of early August 2026).
- Hard questions: Tap Think before sending. It gives Luna more time to work through the answer.
- Don’t tap Think for casual chats. It’s rate-limited – treat it as a finite resource, not a permanent toggle.
4 gotchas worth knowing before you rely on this
These will actually trip you up.
1. Sol wasn’t updated everywhere. The catch: this chat-optimized Sol is only in the Chat experience. Sol in ChatGPT Work and Codex stays unchanged, per the official announcement. If Codex or Work is part of your daily workflow, last week’s behavior is still what you’ve got there.
2. “Luna is the new default” isn’t a straight upgrade. Luna is the trade – more chats, smaller brain. OpenAI’s GPT-5.6 system card explicitly notes that Luna (and Terra) are smaller and less capable than Sol on proxy evaluations. The Think button gives Luna more time – but not Sol’s underlying capacity. Free users who previously hit rate limits on Sol now get more messages. They also get a less capable model. Both things are true.
3. “Unlimited” means text-only. The announcement says unlimited. The fine print says: limits for image generation, file uploads, and other tools stay exactly as before, per Neowin’s coverage of the rollout. If you upload files or generate images regularly, nothing changed for you.
4. Maxing the slider costs tokens. API users especially: reasoning tokens count toward output token billing – no separate surcharge, but higher effort generates more tokens. A high-effort request can produce roughly 2x the tokens of a low-effort one for the same task, based on community documentation. In the ChatGPT app you’re not billed per token, but higher reasoning means slower responses.
Think of it this way: the slider is less like a quality dial and more like a fuel gauge. Turning it up doesn’t guarantee a better answer – it guarantees a more expensive process. Whether that process yields better output depends on whether the question actually needed the extra work.
Performance: what you’ll actually notice
The headline number – 68% fewer factual errors for Sol, 62% for Luna vs. GPT-5.5 Instant – is from internal OpenAI evaluations on financial, medical, and legal prompts (as of the August 2026 announcement). No independent benchmark has confirmed it yet. And “fewer factual errors” isn’t “no errors.” Verify anything consequential regardless.
What’s more useful day-to-day: Sol is designed to give more direct answers and adjust response length to the complexity of the request. Fewer walls of bullet points for a yes/no question. Good enough reason to update on its own.
One gap nobody’s answered yet: OpenAI hasn’t published a per-domain breakdown of where Luna falls short of Sol. The 62% vs. 68% headline suggests they’re close overall – but if you’re a Free user relying on Luna for medical or legal questions, treat that gap as unknown until independent evals appear. That’s not a warning to ignore Luna; it’s a reason to verify outputs the same way you always should.
When NOT to use this setup
- Codex-heavy work: The update doesn’t touch it. Keep your existing Codex settings.
- Mission-critical final decisions: This is still a chat model. Don’t build workflows where the model makes the last call on anything consequential.
- API automation: The Chat-tuned Sol isn’t the API model. Calling
gpt-5.6-solprogrammatically gets you the standard tuning, not the chat retune. - You preferred GPT-4o’s tone: Sol’s answers are more focused and less conversational. If the warmer style mattered to you, this may feel colder.
FAQ
Do I need to do anything to get GPT-5.6 Sol or Luna?
No. Automatic rollout. Plus/Pro see Sol and the slider now; Free/Go get Luna this week, Think button the week after.
Is the reasoning slider the same as the old Thinking mode toggle?
Not quite – and the difference matters if you had workflows built around switching modes. Previously you picked between two separate models: a fast Instant one and a slower Thinking one. Now it’s one model (Sol) that internally adjusts how much it reasons based on the slider position. Same outcome range, completely different architecture underneath. If you had prompt templates that explicitly referenced “Thinking mode,” those mental models need updating too.
Will Luna hallucinate more than Sol?
Probably yes on hard prompts. OpenAI’s own numbers: Luna at 62% fewer factual errors, Sol at 68%. Luna is smaller per the system card. For anything factual, verify – same rule as always, just more important now.
Next action: Open ChatGPT, pick a task you ran last week, run it at the default slider position. Then re-run it one notch higher. Compare. That five-minute test tells you more about your workflow than any changelog.