ChatGPT Think button [What It Changes]

TL;DR

• The ChatGPT Think button is a Free and Go control that gives GPT-5.6 Luna extra time on a harder question; it does not move those accounts onto GPT-5.6 Sol.

• OpenAI announced the change on August 6, 2026: Luna becomes the Free and Go default that week, and unlimited text chats plus Think follow the week of August 10, subject to abuse guardrails.

• Unlimited covers everyday text chats only. File uploads, images, voice, data analysis, and tools such as Deep Research stay capped.

• Plus and Pro get a retuned GPT-5.6 Sol and an effort slider on August 6. Work and Codex are carved out of this Chat-surface release.

Key Takeaways

• Name the control correctly: Think is a two-state Luna switch for Free and Go; the paid slider is a different control on Sol.

• Treat the announcement as three dates, not one present-tense launch. Sol and the slider were live on August 6; Luna as default was “this week”; Think and uncapped text were “next week.”

• Recheck the paid-seat case. The free floor moved (better default model, uncapped text). Remaining paid differentiators are tools, uploads, context, and Sol-class effort control.

• Do not paste Think output into invoices, medical notes, or legal letters without a human pass. OpenAI still forbids unreviewed high-stakes advice and still says output can be wrong.

• If your production agents live in Work, Codex, or the API, this Chat-surface retune is informational unless you change those model ids yourself.

What the ChatGPT Think button is

The ChatGPT Think button is OpenAI’s free-tier control that, on a single message, gives GPT-5.6 Luna more time to work through a harder question instead of answering in Instant mode.

That is not a slogan for “ChatGPT got smarter.” It is a product switch with a named model, a named audience, and a named limit. OpenAI’s August 6, 2026 product post says Free users get unlimited text chats on Luna and, for questions that need more thought, a Think button that “lets you access higher reasoning for harder questions.” The matching ChatGPT release notes repeat the same sentence: Luna becomes the default for Free and Go “this week,” and unlimited text chats plus Think start “next week,” subject to abuse guardrails, with limits still applying for file uploads, images, and other tools.

A 2-truck HVAC shop should care because the old friction was operational, not academic. The person writing a same-day estimate, a warranty reply, or a parts note was the one who hit the free-tier wall and either stopped, opened a second assistant, or paid. Uncapped text chats mean that person can keep iterating the quote in one thread. Think is the extra tap when the question is a multi-step diagnosis rather than a one-line rewrite.

A 10-person marketing agency should care because client work is mostly text: briefs, variants, objection handling, and follow-ups. The agency already has form-to-CRM automation for intake. What changed is the cheap assistant sitting next to that intake. Staff can keep asking without a conversation counter, and they can press Think when the brief is messy. They still cannot lean on free-tier uploads or image generation as if those meters disappeared.

A solo-run clinic should care in a narrower way. Patient-facing emails, intake scripts, and after-visit summaries are text. Think can slow Luna down on dense insurance language. It does not make Luna a clinician, and OpenAI’s usage policies still bar tailored advice that requires a license without a licensed professional in the loop. The useful change is time: the office manager is no longer rationing messages to finish a rewrite before the cap.

What shipped, and when

As of August 6, 2026, OpenAI folded three effective dates into one post. Digital Applied’s August 7 recap keeps the vendor’s own tense: Sol and the effort slider were described as available that day; Luna as the Free and Go default was described as happening that week; unlimited text and Think were described as starting the week of August 10.

Sources: OpenAI, August 6, 2026; ChatGPT release notes, August 6, 2026; Digital Applied timeline.

By August 14, 2026, the same changelog told Free and Go users to “Select Think on the web when you want ChatGPT to reason through a harder question,” which is the first dated confirmation that the control had a web placement, not only a future-tense promise.

The GPT-5.6 family itself is older than this button. OpenAI launched Sol, Terra, and Luna for general availability on July 9, 2026, describing Luna as the family’s fastest and most affordable model and the three names as durable capability tiers rather than frozen version numbers. Think is how that cheap tier shows up in consumer ChatGPT, not a fourth model.

How Think works in plain language

Think is extra deliberation on the model you already have, not a secret upgrade to the paid flagship.

OpenAI’s help article GPT-5.6 in ChatGPT is the mechanism page. It states, in one sentence that should end most arguments: “Free and Go users can use Think for harder questions. Think uses GPT-5.6 Luna, not GPT-5.6 Sol.” Instant on Free and Go is also Luna. So the free user is not routing into Sol when they tap Think. They are telling Luna to spend more time.

Paid users get a different control. On eligible paid plans, a slider sets Instant, Medium, High, Extra High, or Pro. Instant through Extra High are GPT-5.6 Sol; Pro is GPT-5.6 Sol Pro. Plus includes Medium and High and does not include Extra High or Pro. Pro, Business, and Enterprise include Extra High and Pro. Free and Go are marked “Not included” for those Sol reasoning options.

That split matters if a shop is deciding whether to buy Plus because “we need thinking.” On Free and Go, thinking is the Think button on Luna. On Plus, thinking is Sol with a slider. Those are not the same model, and OpenAI’s own internal eval quotes different error reductions for each.

The idea of spending more compute on a hard question is not new. When OpenAI introduced GPT-5 in August 2025, it described a unified system with a fast model, a deeper reasoning model, and a router that could take an explicit “think hard about this” cue. The August 2026 Think button is the free-tier, two-state version of that same trade: latency and cost in exchange for a longer pass. If you need the longer pass, tap it. If you need a one-line subject line, do not.

What “unlimited” actually covers

The word that fails in headlines is “unlimited.” OpenAI’s own sentences keep three fences around it.

Fence one: text chats. The August 6 post says limits still apply for file uploads, images, and other tools. The live ChatGPT pricing page lists Free as “Unlimited text chats with GPT-5.6 Luna” and then a stack of “Limited” lines: messages with uploads, slower image generation, voice chats, deep research, memory and context, Codex, and ChatGPT Work on desktop. Go adds more of those tools, not a blank check.

Fence two: abuse guardrails. The same pricing page footnotes unlimited text chats as “subject to abuse guardrails.” The Pro-tiers help article describes those guardrails as a response to misuse such as automatically or programmatically extracting data, sharing credentials, or reselling access, and says they can produce a temporary usage restriction. OpenAI’s rest-of-world Terms of Use, effective January 1, 2026, separately prohibit circumventing rate limits or protective measures. Uncapped is not unmonitored.

Fence three: plan scope. Unlimited text is described for Free and Go, not as a new Plus or Pro entitlement. The August 6 post describes no usage-limit change for those paid plans. Plus and Pro still buy tools, context, and Sol.

Deep Research is a concrete example of a tool that did not get uncapped. OpenAI’s Deep Research help article says usage varies by plan, that an in-product counter shows remaining tasks, and that plans with a fixed monthly allowance reset every 30 days from first use. Think is not Deep Research. Think is extra Luna time on a chat turn. Deep Research is a multi-step, cited report with its own meter.

The GPT-5.6-in-ChatGPT help page restates the Free and Go rule in current tense: “Free and Go users have unlimited everyday text chats, subject to abuse-prevention safeguards. Separate limits continue to apply to file uploads, image generation, voice, data analysis, and other tools.”

Accuracy: two vendor figures, not a blank check

OpenAI published a pair of internal numbers that will be copy-pasted without their qualifiers. Keep the qualifiers.

In an internal evaluation of financial, medical, and legal prompts requiring factual detail, according to OpenAI, responses containing at least one factual error were about 62% less common with GPT-5.6 Luna and 68% less common with GPT-5.6 Sol than with GPT-5.5 Instant. That sentence is the source of the two percentages. It is not an independent audit. It is not an absolute error rate. It is not a claim about your product FAQ.

About 62% fewer error-containing Luna replies than Instant is the free-tier number from that same linked eval, not the 68% Sol figure.

Every week, according to OpenAI, 1 billion people turn to ChatGPT for questions, searches, planning, research, advice, and decisions. 1 billion people use ChatGPT each week is the scale OpenAI attached to this access change, not a Census count.

Independent benchmarks sit next to those vendor evals. According to Artificial Analysis, GPT-5.6 Luna (max) scores 52 on the Intelligence Index, against a comparable-class median of 18, with a listed 1M-token context window and output around 127.8 tokens per second. The same lab’s Intelligence Index methodology page currently ranks Claude Fable 5.1 (max with fallback) at 66 and lists GPT-5.6 Sol (max) among the leaders, with Luna further down the same table. According to Unite.AI, Luna scored 74.6 on the Artificial Analysis Coding Agent Index in OpenAI’s July launch tables.

The August 6 system-card update is the safety companion, not a quality brochure. It treats this August Sol and Luna release as High capability in cybersecurity and in biological and chemical domains, says neither reaches High in AI self-improvement, and notes dedicated under-18 evaluations for the first time in this card series. It also states that users accessing GPT-5.6 Sol and Luna in Codex and ChatGPT Work are still on the July builds.

OpenAI’s consumer terms still tell you not to treat output as a sole source of truth. The row Terms of Use require human review as appropriate and bar using output about a person for credit, employment, housing, insurance, legal, or medical decisions. Think does not rewrite that paragraph.

Why now: Luna got cheap, then it got default

OpenAI did not publish a sentence that says “we removed the text-chat cap because Luna’s list price fell.” The public timeline is still the reason the economics became plausible.

On July 30, 2026, according to Unite.AI, OpenAI cut Luna API list prices by 80% and Terra by 20%, leaving Sol’s then-stated $5 / $30 rates untouched. OpenAI’s own July 30 price-performance post confirms those percentages and the new Luna list of $0.20 per million input tokens and $1.20 per million output tokens. Luna API input costs $0.20 per 1M tokens on that linked rate card.

The live API pricing page is the checkable current sheet. Short-context standard rates there are Luna $0.20 / $1.20, Terra $2.00 / $12.00, and Sol $4.00 / $20.00, with Sol’s promotional pricing available at least through November 21, 2026, and a 10% uplift on eligible regional-processing endpoints for models released on or after March 5, 2026.

Source: OpenAI API pricing, read at writing. These are developer list rates, not ChatGPT subscription prices.

OpenAI’s July 29 efficiency post is the engineering backdrop: the company says it has scaled to 1 billion active users and more than 2 million businesses, that GPT-5.6 Sol helped rewrite production kernels and cut end-to-end serving cost by 20%, and that speculative-decoding work raised token-generation efficiency by more than 15%. The same post says the agent harness caps tool output at 10,000 tokens by default. Those are serving-stack numbers. They are not a consumer “why we shipped Think” quote. They are why a high-volume Luna default is cheaper to run than it was in early July.

USTA analysis: what the live card implies for a shop that can choose a model

USTA analysis (derived only from the figures already cited). Inputs: live short-context list prices from the API pricing page — Luna $0.20 in / $1.20 out; Sol $4.00 in / $20.00 out; Terra $2.00 in / $12.00 out. Historical Luna list before the July 30 cut, from OpenAI’s July 30 post and Unite.AI: $1.00 in / $6.00 out.

Arithmetic a reader can check: $4.00 ÷ $0.20 = 20; $20.00 ÷ $1.20 ≈ 16.67, shown as 16.7; $2.00 ÷ $0.20 = 10; $12.00 ÷ $1.20 = 10. Historical cut on Luna: ($1.00 − $0.20) ÷ $1.00 = 0.80, and ($6.00 − $1.20) ÷ $6.00 = 0.80. Error-containing-response index from OpenAI’s relative reductions, with Instant set to 100: Luna remaining index = 100 − 62 = 38; Sol remaining index = 100 − 68 = 32.

What that means for a shop: if you can send the same prompt to Luna or to Sol on the API, Sol’s short-context input list is twenty times Luna’s. ChatGPT Free is not billing you those API dollars, but the ratio is why OpenAI can default the largest cohort onto Luna and still offer a Think pass. It is also why Think on Luna is not “almost Sol.” Sol is priced, and eval’d, as a different tier.

Teams already routing documents through US Tech Automations workflows will plug Luna-plus-Think in as a model swap, not a rebuild: the workflow still fetches the file, still writes the draft, still waits for a human on anything that hits a customer.

What each consumer plan actually lists

The ChatGPT pricing table is the consumer surface. The copy we opened is denominated in pounds, so those list prices stay in pounds rather than a guessed dollar conversion. According to Digital Applied, ChatGPT Go is an $8/mo consumer plan in that outlet’s reconstruction; the live pricing page we fetched lists Go at £7 / month.

Sources: ChatGPT pricing; Pro usage multiples in About ChatGPT Pro tiers ($100 Pro = 5× Plus usage, $200 Pro = 20× Plus). Context windows are the Instant / Reasoning rows on the pricing comparison, not API context.

Free Instant context is listed at 27K on that pricing comparison, against 54K for Go and Plus and 128K for Pro. Think does not change those windows. A long paste still hits the plan’s input ceiling, and large pastes can be handled as attachments, which on Free are limited.

Plus is the first plan the pricing page describes as including “Advanced reasoning models with GPT-5.6,” plus expanded uploads, image creation, deep research, memory, Projects, scheduled tasks, custom GPTs, Codex, and ChatGPT Work across desktop, web, and mobile. That is the honest paid case after this release: not more text turns, which Free now also claims, but tools and Sol.

A rival consumer card is useful only as a ceiling check, not as a conversion forecast. According to Anthropic, Claude Pro is $20 if billed monthly ($17 per month on the annual prepay) with a 200k context window on Free through Max, and Max starts at $100 per month. Claude Free still lists extended thinking. OpenAI’s move is not “we invented thinking on a free plan.” It is “we put our current cheap model, uncapped text, and a named Think control on the same free plan that already has a billion-user weekly habit.”

For a U.S. operator counting firms, the Census Bureau’s Statistics of U.S. Businesses last features establishment data for 2022, and the Annual Business Survey data page was revised May 6, 2026. Those pages do not measure ChatGPT seats. They are the reminder that most U.S. employer firms are small, which is why a free-tier default change is an operations story rather than an enterprise-IT story.

What this does not change

Work and Codex are the easy miss. OpenAI says the retuned Sol is optimized for everyday chats and is only in the Chat experience. The version of GPT-5.6 Sol that powers Work and Codex is not changing in this release. Help-center availability matches that: Work lists Sol, Terra, and Luna for Plus, Pro, Business, and Enterprise; Codex lists Terra for Free and Go, and Sol, Terra, and Luna for the paid consumer and business plans. Terra and Luna are “not selectable in standard ChatGPT conversations.”

If your shop’s production path is an agent in Work, a Codex task, or an API call, do not rewrite the runbook because a Think button appeared in Chat. Change the model id when you mean to change the model id.

Governance did not loosen. According to NIST, the AI Risk Management Framework was released on January 26, 2023 for voluntary use. NIST’s Generative AI Profile (NIST AI 600-1) followed in July 2024 and tells organizations to treat confabulation, data privacy, and human-AI configuration as first-class risks. A Think tap that spends more tokens on a wrong premise is still a wrong premise. Put a review step on anything that leaves the building.

OpenAI’s usage policies, effective October 29, 2025, still disallow automation of high-stakes decisions in legal, medical, employment, credit, housing, and similar areas without human review. A shop that auto-sends Think drafts to customers is not “using the new button.” It is skipping the review the vendor already required.

How a small shop should use Think this month

Use Instant Luna for the pile: subject lines, first-pass emails, checklist rewrites, “turn this voicemail into a job note.” Use Think when the thread has several constraints (code, brand voice, parts list, warranty language) and a wrong answer is expensive to unwind. Do not use Think as a ritual on every message; it costs time, and OpenAI still reserves the right to throttle patterns that look automated.

Keep uploads on a paid seat or on a workflow that already stores the file. The free plan still meters attachments. If the job starts as a PDF invoice, a photo of a nameplate, or a spreadsheet, Think will not be the bottleneck. The upload cap will.

Put the human on the last mile. Executive-assistant automations already describe this pattern: the model drafts, the person sends. Think only changes how long the draft thinks. It does not change who is responsible for the send.

A shop that already maps intake forms into a CRM through US Tech Automations can keep that routing and tell staff to press Think only on the messy edge cases — disputed invoices, multi-property scopes, exception emails — while Instant handles the template replies.

For agencies, the same rule applies on client work. Marketing-agency workflow tools still matter more than a chat button when the bottleneck is handoff, not prose. Use Think to pressure-test a brief. Do not use it as the system of record.

If you are comparing this moment to the rest of small-business automation, start with the state of small-business automation and then come back to this page for the ChatGPT-specific control. The button is one step. It is not a stack.

Signal vs Speculation

Demonstrated fact, sourced above: OpenAI announced on August 6, 2026 that GPT-5.6 Luna becomes the Free and Go default, that unlimited text chats and a Think button follow the week of August 10 subject to abuse guardrails, and that uploads, images, and other tools stay limited. Plus and Pro received an updated GPT-5.6 Sol and an effort slider the same day, on web, mobile, and desktop. Work and Codex are not part of the Chat-surface retune. OpenAI’s internal factual-error eval quotes 62% (Luna) and 68% (Sol) fewer error-containing replies versus GPT-5.5 Instant on a financial, medical, and legal prompt set. Luna’s API list price was cut 80% on July 30. Live short-context API list is $0.20 / $1.20 for Luna and $4.00 / $20.00 for Sol. Independent Artificial Analysis currently scores Luna (max) at 52 on its Intelligence Index. Consumer Free still shows limited uploads, images, voice, and Deep Research on the pricing page.

Our read: if X holds — if uncapped text on a current-generation cheap model becomes the free-tier floor across assistants over the next 12 to 36 months — then small and mid-size businesses will stop paying consumer Plus seats merely to finish a long email thread. Paid consumer spend will concentrate on tools (uploads, images, Deep Research, Codex, Work) and on Sol-class effort control. Shops that already run document and CRM workflows will keep those workflows and treat ChatGPT as a draft engine with an optional longer pass. Shops that used the message cap as an accidental brake on unsupervised sending will need an explicit review step, because the cap is no longer doing that job. We do not claim OpenAI will keep unlimited text forever, that Luna will match Sol on your domain, or that a 62% relative drop in error-containing replies on an internal eval will show up in your warranty letters. Those are forecasts or category errors. They stay in this section.

Digital Applied also notes that, as of June 2026, personalization drawing on past chats and connected accounts had already reached Free and Go, which would make this the second free-tier expansion in three months. That June stamp is context for “why the floor keeps moving.” It is not a third unlimited-chat announcement.

Frequently asked questions

What is the ChatGPT Think button?

The ChatGPT Think button is a Free and Go control that gives GPT-5.6 Luna more time on a harder question. It is not GPT-5.6 Sol, and OpenAI’s help center says Think uses Luna, not Sol.

Does unlimited ChatGPT mean I can upload files all day?

No. Unlimited is described for everyday text chats, with separate limits for file uploads, images, voice, data analysis, Deep Research, and other tools. The pricing page still marks those rows limited on Free.

Do Plus and Pro get the Think button?

OpenAI describes Think for Free and Go. Plus and Pro get an effort slider on updated GPT-5.6 Sol, with Instant through Extra High (plan-dependent) and Sol Pro on Pro. Do not assume the free control was copied upward.

When did Think actually appear on the web?

The announcement on August 6, 2026 put Think in the week of August 10. The August 14, 2026 release notes then told Free and Go users to select Think on the web for harder questions.

Did this release change ChatGPT Work or Codex?

No. OpenAI states the retuned Sol is Chat-only and that the Sol version powering Work and Codex is not changing in this release. Help-center availability still lists Terra on Codex for Free and Go.

Can I script the Think button to drain free-tier reasoning?

OpenAI’s terms prohibit automatically or programmatically extracting data and prohibit circumventing rate limits or safeguards. Abuse guardrails can restrict an account even when text chats are described as unlimited. Build on the API or a paid workspace if you need a machine-to-machine loop.

Should a clinic or law office ship Think answers without review?

No. Usage policies still disallow tailored licensed advice without a professional in the loop, and the Terms still say output may be wrong and must be evaluated before use. Think only lengthens Luna’s pass.

Glossary

• ChatGPT Think button — Free and Go per-message control that spends extra time on GPT-5.6 Luna for a harder question.

• GPT-5.6 Luna — OpenAI’s fastest, lowest-cost GPT-5.6 tier; the new Free and Go default in Chat, and the model behind Think.

• GPT-5.6 Sol — Flagship GPT-5.6 tier; updated on August 6 for Plus and Pro Chat, with an effort slider; Work and Codex keep the prior Sol build in this release.

• GPT-5.6 Terra — Mid GPT-5.6 tier for everyday work; listed in Work and Codex availability, not as the Free Chat default.

• Effort slider — Plus and Pro Chat control for Instant, Medium, High, Extra High, or Pro reasoning on Sol / Sol Pro.

• Abuse guardrails — OpenAI’s reservation that unlimited text can still be throttled or restricted when use looks automated, shared, or otherwise abusive.

• Deep Research — Separate ChatGPT tool that produces a cited report and keeps its own plan-based task meter; not what Think turns on.

• Go plan — Low-cost ChatGPT consumer tier that follows Free onto Luna, uncapped text, and Think, with more tool headroom than Free.

The live homepage is the place to see current product surfaces. US Tech Automations documents this control so an operator can map it onto an existing agentic workflow — draft, review, send — rather than treating a new chat button as a new operating system.

When you are ready to put Luna-speed drafts and a human review step on the same path, open the agentic workflow builder and wire the model swap there.