About this research · Sources, method & limitations
Choose Instant for straightforward questions, quick rewrites and routine classification. Choose Thinking when the task has several constraints, needs a careful plan, involves code or documents that must be reconciled, or benefits from a visible reasoning effort. A Thinking answer can still be wrong, so verify important facts, calculations and actions. OpenAI's September 14, 2026 release note says Plus and Pro users globally no longer get automatic Instant-to-Thinking switching; the user can select an available option in the model picker.
The change is about control, not a new guarantee of accuracy. OpenAI says it is retiring automatic switching from Instant to Thinking for Plus and Pro users and removing the Higher intelligence setting on the web for those plans. ChatGPT may still switch automatically for safety purposes. The exact options and wording can vary by plan, account and rollout.
This guide is based on official OpenAI documentation, not an account test or benchmark. It does not promise that Thinking is always better, faster or available on every device. Use the mode as a workflow choice, then check the result yourself.
Instant versus Thinking in one table
| Task shape | Start with | Why | Verify |
|---|---|---|---|
| One-step rewrite or summary | Instant | Speed and a clear instruction are usually enough | Names, tone and omitted caveats |
| Multi-constraint plan | Thinking | More room to reconcile requirements before drafting | Every requirement and assumption |
| Code diagnosis or refactor plan | Thinking | The task benefits from tracing state and edge cases | Run tests and inspect the diff |
| Source comparison or long document | Thinking | A deliberate pass can organise evidence and conflicts | Citations, quotations and missing pages |
| Sensitive or consequential action | Either, then human review | Mode choice does not replace approval | The exact action, recipient and final state |
What changed on September 14, 2026?
OpenAI's ChatGPT release notes record three relevant points:
- automatic switching from Instant to Thinking is being retired globally for Plus and Pro users;
- users can still choose an available option in the model picker to give ChatGPT more time to reason;
- the Higher intelligence setting is being removed from the web for those plans, while automatic switching may still occur for safety.
This is not the same as saying a new model launched on that date. It also does not describe a change to every API model, every plan or every mobile surface. Check the current model picker and account documentation when the choice matters.
Earlier OpenAI model-picker guidance describes Instant as fast responses for everyday questions and Thinking as deeper reasoning for more complex tasks. It also describes optional thinking effort when Thinking or Pro is selected. Labels, availability and limits can evolve, so use that guidance as a mental model rather than a permanent UI contract. See the model-picker notes.
A four-question decision tree
October 7 check: thinking levels and product surfaces
The current OpenAI guide describes a Thinking slider on eligible paid plans, with Medium, High and Extra High options depending on the plan. Free and Go use a separate Think option. Asking an Instant chat to “think harder” does not automatically change the selected thinking level on Plus or Pro. Enterprise users may retain the older picker during rollout. These are documented controls, not an interface we tested. Current thinking-level guide.
Choose the actual control shown in your account rather than expecting a universal button called Thinking. The decision table below describes task suitability, not a promise that every account has the same labels or models. The September 14 automatic-switching note remains relevant, but it is now historical context rather than the full current interface description.
The current documentation distinguishes Chat from ChatGPT Work and Codex, including the GPT-6.1 Sol rollout in those latter surfaces. Do not infer a change to Chat's thinking controls from a Work or Codex model announcement. Model access and reasoning options depend on the plan and workspace. OpenAI model and availability guide.
For Chat-specific availability, use the current GPT-5.6 and GPT-6 Pro guide, rather than an older model-launch article. Record the product surface, selected model and account limits when comparing outputs. This source refresh does not include a signed-in interface test, and the September 14 automatic-switching change remains the basis of this guide.
Before selecting a mode, ask:
- How many conditions must stay true? If the answer has one obvious transformation, start with Instant. If it must satisfy a rubric, schema, codebase constraint or several stakeholders, Thinking is a reasonable starting point.
- What happens if a hidden assumption is wrong? A low-stakes caption can be revised. A tax, legal, medical, security or financial decision needs independent verification regardless of mode.
- Can I test the result? A code change can run tests; a data extraction can reconcile source rows; a plan can be checked against a checklist. If there is no verification path, do not treat extra thinking time as evidence.
- Is the task expensive in tokens, time or attention? Use the smallest mode that can meet the requirement, then escalate with the original prompt and the failed output when needed.
Three real-world workflows
For writers and marketers
Use Instant to turn an approved outline into a first draft, change a headline or produce short variations. Use Thinking when the brief includes a target audience, source boundaries, legal wording, internal links, a word range and a house style that must all coexist.
Prompt for the decision, not for hidden certainty:
Draft a 900-word explainer from the supplied sources. First list the claims that need citations, the claims that are inferences and any missing evidence. Keep every number tied to a source. Then write the draft and leave a short verification checklist.
The checklist is more useful than asking the model to “be accurate.” Read the cited pages, check names and dates, and remove any sentence that outruns the evidence.
For developers
Instant is useful for explaining a compiler error or sketching a small function. Thinking is more appropriate when you need a migration plan, a race-condition analysis, a test strategy or a refactor across several files. Provide the repository constraints and ask for a plan before edits.
Inspect the error and propose the smallest change. List the files, assumptions and tests first. Do not claim the fix passes until the named command runs.
Even with Thinking selected, run the tests, inspect the diff and check the production build. A reasoned explanation is not a substitute for execution evidence.
For research and analysis
Use Thinking when comparing several documents, reconciling conflicting figures or designing a method that must be repeatable. Supply the source set and ask the model to separate source facts, inferences and unknowns. If the source is a live webpage, verify the page date and the quoted passage yourself.
Define the task
Step 1 of 4Write the deliverable, constraints, source set and what a correct result must contain.
Compare three product policies and identify conflicts, not just a summary.
Stop and review: The task has no success test or includes a decision the user has not authorised.
Choose the mode
Step 2 of 4Start with Instant for a bounded transformation; choose Thinking when the work must plan, reconcile or inspect edge cases.
Thinking is selected because the comparison needs a source matrix and conflict log.
Stop and review: The mode is being chosen as a substitute for fact checking or approval.
Inspect the result
Step 3 of 4Ask for citations, assumptions and an uncertainty list, then compare material claims with the sources.
A missing policy page is marked unknown instead of filled with a plausible sentence.
Stop and review: A citation does not support the claim or a key assumption is hidden.
Run the final check
Step 4 of 4Test code, recalculate figures, read the exact output and approve any external action separately.
The final brief keeps the conflict table and a human signs off before sharing it.
Stop and review: The output changed after review or the final destination is unclear.
Common mistakes with Thinking mode
“Thinking” means the answer is verified
It does not. The mode controls how the system allocates reasoning time or selects an available reasoning option. It does not turn a generated claim into a source. Ask for evidence and check it.
Always choose the slowest option
That can waste time and limits. A short, reversible task does not need a long planning pass. Start small, escalate when the task shape or failed output justifies it.
Hide a vague brief behind a stronger model
If the prompt has no audience, input boundary or acceptance test, more reasoning can produce a longer version of the same ambiguity. Write the decision criteria first.
Treat plan labels as universal
OpenAI states that model choices depend on account and plan. A label seen in a screenshot, another country or an older release note may not match the current picker. Record the date and surface when documenting a workflow.
FAQ
Did OpenAI remove Thinking from ChatGPT?
No. The September 14 note says Plus and Pro users can still select an available option in the model picker to give ChatGPT more time to reason. It retires automatic Instant-to-Thinking switching globally for those plans and removes the Higher intelligence setting on the web.
Will Thinking always be more accurate?
No. It may be a better fit for complex tasks, but it can still misunderstand a source, make a calculation error or follow a bad assumption. Verify the output with tests, primary sources or a qualified reviewer.
Does this change the OpenAI API?
The cited release note describes ChatGPT Plus and Pro behavior. It does not, by itself, establish an API pricing, model or quota change. Check the relevant API documentation before making an engineering decision.
Can I use this workflow in India and the United States?
Yes as a decision method, but product access, plan names, limits, data handling and regulatory context can differ. Do not publish a plan comparison or promise availability without checking the account and region you are writing for.
Bottom line
The useful skill is not “always use Thinking.” It is recognising when a task needs deliberate planning and then attaching a verification step to the result. Use Instant for bounded work, Thinking for multi-constraint reasoning, and human approval for anything consequential. Record the model choice and date when the workflow matters because the picker and plan rules can change.
Reader discussion
Comments are public and require a GitHub account. By taking part, you agree to the Terms and Conditions and acknowledge the Privacy Policy.