GPT-6 Reaches the ChatGPT Chat Tab. The Chat Tab Is Now One Generation Behind Codex.

When GPT-6 Sol and Luna shipped on September 22, we noted the odd part: the models were in the API, in ChatGPT Work, and in Codex, but not in the Chat tab where most of ChatGPT's users live. Today, October 7, OpenAI fixed that. "GPT-6 with Intelligent UI starts rolling out globally to ChatGPT Plus, Pro, Business and Enterprise tiers today in the Chat tab," the post says, with Free and Go "starting tomorrow."

The headline feature is Intelligent UI. OpenAI's sentence: "We've trained GPT-6 to compose responses using text, visuals and interactive elements, choosing how they fit together based on your question." Ask how to split a bill and you may get a bill splitter instead of a paragraph. Ask to compare two phones and you may get a table you can sort.

That is the news. The fine print is which GPT-6 you are getting, and the answer is: not the one Codex has. We read the announcement, OpenAI's X posts, the ChatGPT release notes, two help-center articles that disagree with the announcement, the 24-page October system card, and the API pricing page. Here is what held up.

Which Model, Which Tier, Which Tab

"GPT-6 Instant" is a setting, not a model. The help center is explicit: "When GPT-6 is selected, GPT-6 Sol powers Instant, Medium, High, and Extra High; Pro uses GPT-6 Pro, powered by GPT-6 Astra." Instant is the lowest thinking level. Several outlets wrote about GPT-6 Instant as if it were a new model ID; there is no such ID.

Chat is one build behind, and one model behind. The October system card says it plainly: the Chat tab gets new October builds that "will replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT," while "users accessing GPT-6 Sol and GPT-6 Luna in Codex, and via ChatGPT Work, are still using previously released versions." And GPT-6.1 Sol, which shipped to Work and Codex on September 29, is still "not yet available in Chat." A Plus subscriber's Chat model and Codex model are now different generations of the same name.

Two OpenAI help articles have not caught up. Article 20001354, "GPT-6 and other models in ChatGPT," still says "Free and Go use GPT-5.6 Luna" and that Think on those tiers "uses GPT-5.6 Luna, not GPT-5.6 Sol." Article 20001275 still says GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna "are not available in regular ChatGPT conversations." The announcement, the release note, OpenAI's X post, and the system card all say Free and Go get GPT-6 Luna and paid tiers get GPT-6 Sol in Chat. We print the four sources that agree and footnote the two that do not.

The Speed Claim, Quoted Exactly

OpenAI's post contains one performance number, and it is narrower than most of the coverage: "For questions that need web search, GPT-6 Instant starts answering 44% sooner, on average, than GPT-5.6 Instant." Web-search questions, Instant level, time to start answering. The Decoder reported it as a general 44% cut in wait times from interleaved thinking; that is not what the sentence says.

The interleaving is real and separate. The release note: "GPT-6 can also begin answering while it continues to think or use tools, then add more findings without another prompt. You can start reading before the full answer is ready." The post adds a second claim with no number attached: "In an internal evaluation of high-value, everyday agentic tasks, GPT-6 Extra High is able to begin answering in the same amount of time as GPT-5.6 Medium while achieving a better overall score than GPT-5.6 Extra High." The evaluation is unnamed and no score is printed.

There is no benchmark table in the post. The only other figures are "1.2 billion" weekly users, stated twice.

The Benchmarks, Sorted by Who Reported Them

Independent. Artificial Analysis has not scored the October builds; its entries are the September API models, on Intelligence Index v4.3.2, which has not changed since September 19. GPT-6 Sol (max) 48; GPT-6 Luna (max) 38; GPT-6 Sol with reasoning off 29; GPT-6 Luna with reasoning off 18. For the Instant experience the non-reasoning rows are the honest comparison: AA measures GPT-6 Sol at 95.9 tokens per second with 0.87 seconds to first token with reasoning off, and Luna at 113 tokens per second and 0.95 seconds. For context, the models the Chat tab does not get: GPT-6.1 Sol (max) 52, GPT-6 Astra (max) 53, and Google's Gemini 4 Argon (high) 53 since September 30.

Vendor. None for capability. The October system card is a safety document, and that is where the numbers are.

Three honest observations:

• The safety card shows regressions, and says so. Against the August GPT-5.6 Sol and Luna builds, the October GPT-6 builds show statistically significant declines on self-harm, gore, sexual content, under-18 emotional reliance, and vision extremism in OpenAI's production benchmarks; on emotional reliance for minors the Sol score falls from 0.921 to 0.770 and Luna from 0.927 to 0.734, where higher is better. The tables have six columns (two GPT-5.5 Instant updates, the two August GPT-5.6 builds, the two October GPT-6 builds); compare Sol with Sol and Luna with Luna. Multi-turn jailbreak point estimates sit slightly below the September builds. Prompt-injection robustness is 97.13% for Sol and 95.80% for Luna; environmental-barrier circumvention 28% for Sol and 15.9% for Luna. Factuality improved except for claim-level errors on fact-heavy prompts.
• Preparedness is unchanged, and it is High. The card rates the October builds High in cybersecurity and High in biological and chemical capability, not High in AI self-improvement, with the same safeguards as GPT-5.6 Sol and Luna.
• The API does not get Intelligent UI. The chat-latest alias points at "the latest Instant model currently used in ChatGPT," but its snapshot "will be regularly updated" with no pinning, it carries a 400K context and a $5 / $0.50 cached / $30 price per million, and OpenAI's own page says "We recommend leveraging GPT-6 Astra for production API usage." The interface layer is a ChatGPT product, not a model feature.

What Intelligent UI Is, and Where It Stops

OpenAI describes the mechanism in one paragraph: GPT-6 writes a response that mixes prose with interface elements, and a compiler renders it progressively from a component library in the client. The examples in the post are calculators, bill splitters, comparison tables, diagrams, and small games.

The limits, from the help center: it works on the web and in "supported updated apps," so update the mobile app; the older macOS and Windows desktop apps do not support it; a "Layout and visuals" setting reduces visuals but there is no text-only mode; there is no separate usage quota for it; and it is off at the Pro reasoning level, because Pro runs GPT-6 Astra and "doesn't support Intelligent UI." Enterprise admins get no workspace-level switch for it that we could find; GPT-6 in Chat follows the workspace's existing GPT-6 Sol access setting, and GPT-6 Astra stays off by default.

Aarush Selvan, an OpenAI product manager, told TechCrunch that "ChatGPT has predominantly been a text-based interface" and that for everyday tasks "the most helpful answers aren't just text." That is the whole pitch, and it is a reasonable one. Nobody has tested it independently yet, and we have no timing of the 44% claim.

What It Costs, and What the Caps Are

Nothing changes on the API side for Sol and Luna: GPT-6 Sol stays $2 / $0.20 cached / $10 per million and GPT-6 Luna $0.10 / $0.01 / $0.50, with GPT-6.1 Sol at $2 / $0.10 / $10 and Astra at $10 / $1 / $50, all for prompts up to 272K input tokens. GPT-5.6 Sol is still sold at its $4/$20 promotional rate "at least through November 21, 2026."

On the consumer side, the only numeric caps OpenAI publishes are for Business: 15 GPT-6 Pro messages per month on Standard and 50 per week on Premium, shared with GPT-5.6 Sol Pro. For Plus, the three Pro tiers, and Enterprise, OpenAI's own words are "Your allowance does not correspond to a fixed number of messages or tasks," so those caps print as dashes. One change worth knowing if you are buying Pro 200 today: new subscriptions carry "a lower usage allowance" than grandfathered ones, and grandfathering applies only to accounts active from September 22 through 10 a.m. Pacific on September 29, and only through October 29.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use GPT-6 in Chat if you are a Plus, Free, or Go user who stayed on GPT-5.6 because that was all the Chat tab offered. You now get a GPT-6 model at no change in price, a visual answer format for everyday questions, and a faster start on searches by OpenAI's own measure. Update the mobile app first.

Skip the assumption that Chat is the frontier. If you pay for Plus and run real work, the Work tab and Codex have GPT-6.1 Sol and Extra High; Chat has GPT-6 Sol at High. If your use touches the categories the system card flags, the October builds regressed against August's on OpenAI's own benchmarks. And if you build on the API, nothing here applies to you: there is no Intelligent UI endpoint and chat-latest moves under your feet.

The frontier is elsewhere: Claude Opus 5.5 (max) leads the v4.3.2 index at 58, Claude Sonnet 5.5 at 56, and GPT-6 Astra, Claude Fable 5.1, and Gemini 4 Argon at 53. GPT-6 Sol at 48 is a good everyday model, which is what OpenAI calls it: "tuned for everyday conversation." Our model comparison has the current standings side by side.

Update, October 16, 2026

• October 8: Free and Go got GPT-6 Luna as scheduled. Thesys reverse-engineered Intelligent UI on the web: about 70 components in ChatGPT's client registry with 39 observed in responses, a markup format mixing Markdown, JSX-like tags, and JavaScript that a server compiles and streams as patches at a median 230 ms, and four documented limits: interface state resets on reload, the model cannot see interface state, logic errors fail silently, and layouts are fixed. The same day OpenAI switched on GPT-6.1 Sol Ultrafast in Codex and Work for Pro 500 and eligible Enterprise and Edu plans.
• October 9: Codex gained "Composer predictions" in beta, a Tab-to-accept suggestion for your next message, for personal Pro users on GPT-6 Astra or GPT-6.1 Sol threads.
• October 14: ZDNet ran five prompts through Intelligent UI with no failures and used a generated bill splitter in the iPhone app, which settles the mobile question OpenAI never spelled out. OpenAI retired GPT-5.5 from ChatGPT, Work, and Codex on all plans; the API is unaffected. Eight states sued OpenAI in their own courts over product-safety claims, per Engadget.
• October 15: The FTC subpoenaed OpenAI and Anthropic for documents on security incidents and safeguards, first reported by Axios and confirmed by the agency to CBS News. ChatGPT Voice moved to GPT-Live-1.1. The plan page now lists Free at a 27K Instant context, Go and Plus at 54K Instant and 256K reasoning, Pro at 128K and 400K.
• October 16: Both stale help articles still say GPT-5.6 Luna. Artificial Analysis has no entry for the October builds, and nobody has independently timed the 44% claim.

Sources

• OpenAI: GPT-6 and Intelligent UI for everyone
• OpenAI help center: GPT-6 and other models in ChatGPT
• OpenAI help center: ChatGPT release notes
• OpenAI help center: ChatGPT Business models and limits
• GPT-6 Sol and GPT-6 Luna: October 2026 update (system card)
• OpenAI API pricing page
• OpenAI model docs: chat-latest
• Artificial Analysis: GPT-6 Sol
• Thesys: How ChatGPT's Intelligent UI works
• TechCrunch: ChatGPT is getting a lot more visual

*Not sure whether the Chat tab, the Work tab, or the API is where your work belongs? Take the free 2-minute quiz and get matched. Then read the other model that shipped today at a tenth of the price and with a surcharge nobody led with: Claude Haiku 5.5.*