GPT-6 Reaches Every ChatGPT Account
On October 8, 2026 OpenAI completed a global rollout of GPT-6 across ChatGPT, replacing the GPT-5.6 Sol and GPT-5.6 Luna builds that had served the product until then. The release is less about raw parameter scale than about a new presentation layer: answers are no longer confined to plain text, and API pricing dropped permanently by half at the same moment.
Two Models, Three Tiers, Half the Price
Paid subscribers run GPT-6 Sol while free and Go users receive GPT-6 Luna. These are separate models tuned for everyday conversation rather than one set of weights stretched across every plan. Plus, Pro, Business and Enterprise accounts began receiving the update on October 7, with free users following the next day.
The split follows the GPT-6 family announcement of September 22, which folded several capabilities from the flagship GPT-6 Astra into the mid-tier and lightweight builds and cut prices at the same time. Sol moved from $4 to $2 per million input tokens and from $20 to $10 per million output tokens. Luna dropped from $0.20 to $0.10 per million input tokens and from $1.20 to $0.50 per million output tokens. OpenAI now operates a clean three-level lineup: Astra at the top, Sol in the middle, Luna at the entry point.
Sol also carried the sharper benchmark story. On AutomationBench at the highest reasoning setting it scored 33.2 percent, above the 26.9 percent posted by Claude Opus 5 under maximum-effort conditions, at roughly one ninth of the cost per task. On the Agents Last Exam evaluation it reached 56.4 percent.
Intelligent UI Turns Answers into Things You Can Use
The substantive change sits in what OpenAI calls the Intelligent UI layer. Instead of writing one continuous block of text, GPT-6 first classifies the question and then chooses how to render the result. The model was trained deliberately to mix text, visual elements and interactive components, picking the combination that fits the request.
Comparison questions come back as side by side layouts. Conceptual explanations arrive as interactive diagrams. Simple queries still return plain sentences. A question about how a wing generates lift can produce a series of illustrations covering airflow, pressure differential and lift formation. Practical requests return recipe cards, bicycle drivetrain diagrams, multi day hiking maps or a personal savings calculator. Short prompts can also generate a working calculator, a bill splitting tool, or a small playable game inside the chat window.
Product manager Aarush Selvan framed the goal as finishing tasks rather than producing paragraphs, whether the user is planning meals, organizing a trip or learning something new.
The implementation rests on a natively streamable component library and a compiler that processes interface elements while GPT-6 is still generating. Widgets therefore appear progressively rather than after the full response lands. The layer is also restrained: it stays out of the way for text-only answers, and users can dial the amount of visual content up or down, or turn it off to keep the traditional text experience.
Thinking Out Loud Reduces Perceived Wait
Reasoning models have always made users wait through a complete internal chain before the first token appeared. GPT-6 instead streams while reasoning continues in the background, then revises and extends earlier content as the thinking deepens.
OpenAI measured a 44 percent reduction in average time to first response on search-dependent questions when comparing GPT-6 Instant with GPT-5.6 Instant. On internal evaluations of high value everyday agent tasks, GPT-6 Extra High responded in roughly the same time as GPT-5.6 Medium while scoring above GPT-5.6 Extra High.
Search Judgment, Safety Floor, and Scope Limits
Under the hood the model is better at deciding when a live search is actually required and at grounding answers in the results it retrieves, which raises factual reliability.
Security work inherited from GPT-6 Astra tightens defenses against cyber, biological and violent misuse. The deployment safety report places both Sol and Luna at the High capability threshold in cybersecurity and biochemistry without reaching Critical, and below High for AI self improvement. Against multi turn jailbreak attempts, both versions held better than GPT-5.6 Sol at every attack budget. On instruction hierarchy tests, Sol and Luna resisted system prompt injections at 99.99 percent and 99.79 percent respectively. OpenAI reports factual accuracy ahead of the corresponding GPT-5.6 releases across nearly all metrics, including medical, legal and financial questions where errors carry real cost. On the under 18 safety evaluation, Luna's emotional reliance score improved from 0.927 to 0.734.
Two boundaries matter. This rollout covers the Chat experience; the models behind Work and Codex stay on their previously released versions. And the wider signal is that competition among frontier models is shifting from answer quality toward how efficiently an answer can actually be acted on.
