GPT-6 Comes to ChatGPT: Sol, Luna, and the Route OpenAI Already Deprecated

On 7 October 2026 GPT-6 Sol became the default answer engine inside ChatGPT for anyone paying for a seat, GPT-6 Astra kept the Pro effort level, and GPT-6 Luna was named for the free tiers in OpenAI's own announcement — while the help centre for the same product says Free and Go get GPT-5.6 Luna instead. The models themselves shipped on 22 September 2026, nearly three weeks before this, so nothing about their weights or their rate cards changed this week. What changed is where they are reachable and who they answer. The complication is the other Sol: GPT-6.1 Sol, released 29 September, now sits where GPT-6 Sol used to in OpenAI's flagship pricing table, and the independent evaluator that benchmarks GPT-6 Sol has already marked it deprecated in favour of it.

If you only want the practical line: on a paid ChatGPT tier the model you are talking to is the same GPT-6 Sol you can call over the API, and the more interesting story is the two ways OpenAI has already started stepping away from it.

What actually changed on 7 October

Through September, GPT-6 was an API generation with three members and no consumer surface of its own. GPT-6 Astra sat at the top at $10.00 / $50.00, GPT-6 Sol in the middle at $2.00 / $10.00, and GPT-6 Luna at the bottom at $0.10 / $0.50. On 7 October OpenAI put that generation in front of the more than 1.2 billion people who use ChatGPT each week under a capability it calls Intelligent UI, and on 8 October it widened the rollout to the Free and Go tiers.

Intelligent UI is a response format, not a new model. OpenAI describes it as "a library of native, streamable components, along with a compiler that processes the interface as the model generates it," so an answer can arrive as a chart, a form, a clickable diagram, a calculator or a game rather than a wall of prose. The examples OpenAI gives are a calculator for exploring how savings grow, a bill splitter for dinner with friends, and a game you can play inside the conversation. Because the interface is compiled as the model generates, the response can start appearing before the model has finished thinking — which is the other half of the 7 October change, and the one that is easier to measure.

Both of the performance claims OpenAI attaches to that are vendor-reported and were produced internally. In OpenAI's evaluation of "high-value, everyday agentic tasks," GPT-6 at Extra High begins answering in the same time as GPT-5.6 Medium while scoring better overall than GPT-5.6 Extra High. For questions needing web search, GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant. Neither figure has an independent counterpart yet, and both are comparisons against the previous generation rather than against the model OpenAI is now selling above it.

The rollout reached Plus, Pro, Business and Enterprise in the Chat tab on 7 October, with Free and Go following from 8 October, and enterprise availability gated by workplace admin settings. The models powering Work and Codex were explicitly left unchanged.

Which model is actually answering you

The tier rule is worth pinning down, because two OpenAI surfaces describe it differently on the same day.

• ChatGPT Instant, Medium, High and Extra High, on Plus, Pro, Business and Enterprise: GPT-6 Sol
• ChatGPT at the Pro effort level: GPT-6 Pro, powered by GPT-6 Astra
• Free and Go, per OpenAI's announcement: GPT-6 Luna
• Free and Go, per OpenAI's help centre for GPT-6 in ChatGPT: GPT-5.6 Luna

The announcement says plainly that "GPT-6 in ChatGPT is powered by GPT-6 Sol for Plus, Pro, Business, and Enterprise tiers, and GPT-6 Luna for Free and Go tiers." The help centre page, read the same day, says that under Instant — "fast responses for everyday questions" — Free and Go use GPT-5.6 Luna, and adds that free and Go users receive GPT-5.6 Luna and do not have access to GPT-5.6 Sol. Those are not compatible statements about the same tier. One of them is stale or one of them is wrong, and OpenAI has not said which. If your work depends on which cheap model sits behind the free tier, treat the help centre entry as the more specific page and the announcement as the broader one, and re-check before you build against either.

The second thing to notice is that "GPT-6" was never a single model, and the paid tier does not get a special checkpoint. The GPT-6 Sol answering a ChatGPT question on a Plus seat is the same GPT-6 Sol on the API rate card below. The consumer surface is a distribution channel, not a different SKU.

The independent read, and a deprecation flag

Artificial Analysis benchmarks these models without taking money from the vendors, and its GPT-6 Sol page currently carries a deprecation notice. The page states that "OpenAI has launched a newer release, GPT-6.1 Sol," and suggests considering it instead, and it notes that only the default 10k-input-token workload continues to be benchmarked while results for other workloads are frozen as historical.

On the numbers as they stand: GPT-6 Sol scores 48 on the Artificial Analysis Intelligence Index in the max reasoning configuration, ranked 18th of 212 models on that board, with a Coding Index of 60 and an Agentic Index of 52.5, at $1.54 blended per million tokens. GPT-6.1 Sol scores 51.8 on the same index at $1.47 blended — a higher score for less money on the evaluator's blended basis. GPT-6 Astra scores 52.7 at $7.70 blended, and GPT-6 Luna 38.1 at $0.077. Those are independent figures rather than vendor charts, and they are the honest basis for comparing the two Sol models, because OpenAI has never published the two against each other.

Two caveats. First, an index score is not a workload: GPT-6 Sol is a reasoning model with a configurable reasoning.effort running from none to max, and the figure above is one configuration of it, not a range. If a decision rests on a two- or three-point gap, benchmark on your own traffic. Second, deprecated at an independent evaluator means the measurements have stopped being refreshed; it does not mean the vendor has retired the model. The two events are separate, and this week only the first has happened.

What did happen on the vendor side is quieter. OpenAI's current pricing table lists GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna in its flagship rows. It does not list GPT-6 Sol, which is what that table does with a superseded generation. The model is still callable and still documented on its own page, which meters it at $2.00 per million input tokens, $0.20 cached input, $2.50 for cache writes and $10.00 output, with a 1,050,000-token context window and a 128,000-token output ceiling — and which carries a line pointing readers to GPT-6.1 Sol as the newer Sol model.

The two Sol rate cards, side by side

Because the two models are priced the same at the headline rate, the difference shows up in the tiers underneath it.

• Standard input: GPT-6 Sol $2.00 per million, GPT-6.1 Sol $2.00 per million
• Standard output: GPT-6 Sol $10.00 per million, GPT-6.1 Sol $10.00 per million
• Cached input: GPT-6 Sol $0.20 per million, GPT-6.1 Sol $0.10 per million
• Cache writes: GPT-6 Sol $2.50 per million, GPT-6.1 Sol $2.50 per million
• Above 272K input tokens: both reprice the whole request to $4.00 / $15.00
• Cached input above 272K: GPT-6 Sol $0.40 per million, GPT-6.1 Sol $0.20 per million
• Context and output: both 1,050,000 tokens in, 128,000 tokens out
• Reasoning effort: both expose a configurable reasoning.effort, from none to max

The long-context rule is the same one OpenAI uses across this generation: prompts above 272,000 input tokens are charged at 2x the input and cache rates and 1.5x the output rate for the full request, not just the portion above the line. The practical read is that GPT-6.1 Sol is the same list price with half-price cached input at both tiers, a better independent score and a benchmark refresh that is still running. The reason to pin GPT-6 Sol anyway is that it is the model your evaluation baseline was built against — and that is a statement about your regression suite rather than about the models.

Where this leaves anyone building on it

Three consequences follow. First, there is no model id for the generation on its own: on OrcaRouter the four live routes are openai/gpt-6-sol, openai/gpt-6-luna, openai/gpt-6-astra and openai/gpt-6.1-sol, each with its own rate card and its own role, and no openai/gpt-6 to call. If your code says "GPT-6," it is saying four different things depending on where it runs.

Second, the consumer surface moves faster than the API. Intelligent UI arrived in the Chat tab on a schedule that had nothing to do with a model release, and it will change again on a similar schedule. Anything you build for production that depends on the generation should depend on a pinned model id and a rate card you have read, not on which capability ChatGPT is advertising this month.

Third, when an evaluator and a vendor both start marking a model superseded within weeks of each other, the window for a clean migration is open and it will not stay open indefinitely.

Routing across a generation that is mid-transition

The awkward part of a transition like this is that the old model still works and the new one may not be better on your traffic yet. That is the case OrcaRouter is built for: one API in front of 200+ models, with the provider's list price passed through at 0% markup, so a vendor price change lands here the same day rather than at your next billing reconciliation, plus automatic failover across providers if a route degrades. All four GPT-6 models are live routes in the same catalogue, called through the OpenAI-compatible chat completions and Responses endpoints, so moving traffic from one Sol to the other is a model-id change rather than a rewrite. The routing DSL lets you split by request size or by share, which is the usual shape of this migration: send a slice of traffic to GPT-6.1 Sol, compare on your own evaluations, widen the slice, and keep GPT-6 Sol as the fallback until the numbers say otherwise.

One thing OrcaRouter does not do is hand you Intelligent UI. That capability belongs to ChatGPT's own surface — it is a rendering pipeline in the client, not a model, and nothing in an API response resells it.

Bottom line

The 7 October change is a distribution event, not a model release. GPT-6 Sol became the default answer engine for paid ChatGPT seats and Intelligent UI gave it somewhere new to be seen; the models themselves date from 22 September 2026. What a builder should act on is that OpenAI has already dropped GPT-6 Sol from its flagship pricing table, an independent evaluator has already marked it deprecated, and the successor costs the same per token with better cache economics and a still-running benchmark refresh — while ChatGPT keeps serving the older model to a billion-plus weekly users. Pin GPT-6 Sol only if a frozen evaluation baseline is worth more than that delta; otherwise GPT-6.1 Sol is the same price with an active benchmark and better cache economics, and the migration is a model-id change.