Dots vs GPT-5.6 Sol: A Subscription and a Rate Card From the Same Company

The company sells two things that both look like "an Open​AI model" from the outside. Dots is the always-on agent announced at DevDay on September 29, 2026: each dot gets its own cloud computer and browser, reaches more than 4,000 apps through the company's connector ecosystem, lives inside ChatGPT, Slack and Teams, and runs on GPT-6 Astra, which is the model the company names in its own launch materials. GPT-5.6 Sol is a metered API model, shipping since July 9, 2026: a 1,050,000-token context window, up to 128,000 tokens of output per response, text, image and file input with text out, tool calling, structured output and a configurable reasoning effort dial. The first is a subscription whose allowance nobody has published; the second is a rate card with four numbers on it. That difference, not a benchmark gap, is what this comparison is actually about.

So the useful question is not which one is smarter. It is which one you can plan around, and which one quietly decides how much of your work is possible at all.

The one thing that separates them

A dot has no token meter. You cannot see how many tokens it burned, because that is not the unit it is sold in. GPT-5.6 Sol has nothing but a token meter — every capability it has is priced per million tokens, and its one hard architectural limit, the 1,050,000-token window, is also the limit on how much of a problem you can hand it in one go.

That asymmetry explains almost every practical difference between them. When you cannot meter something, you cannot forecast it, and you cannot move it. When you can meter something, you can compare it to a cheaper sibling and switch, because the only thing that changes is an identifier string.

What GPT-5.6 Sol publishes

• Context and output — 1,050,000 tokens in, 128,000 tokens out, which is enough for a full repository or a year of correspondence in a single call

• Input surface — text, image and file in, text out, with vision, tool calling, JSON and reasoning all listed as first-class capabilities

• Price — $4.00 per million input tokens and $20.00 per million output tokens below a 272,000-token prompt, stepping to $8.00 and $30.00 above that line; cached reads cost $0.40 per million, cache writes $5.00

• Independent evidence — an Artificial Analysis Coding Index of 77.4, third of 138 models measured, and an Intelligence Index of 47, thirteenth of 145; 94.1 on GPQA Diamond and 84 on long-context recall, all third-party figures carried with their source in our catalogue rather than vendor numbers

• Observed behaviour in our own traffic — a 1.72-second median time to first token over a seven-day window, about 171 output tokens per second, and 88.99 million tokens routed in that week

Those numbers are not a sales pitch; they are a specification. You can price a job on this model before you run it, and you can check the answer afterwards, because every unit is counted somewhere.

What Dots does not publish

OpenAI has said exactly one thing about what a dot costs: the first one is included with Pro and Business Premium, with extended limits in the first month and no drawdown of ChatGPT usage limits for conversations with your dot. Everything else is unquantified. There is no allowance figure, no price for a second dot, no published rate for faster or larger capacity, and no per-task cost anywhere in the launch materials. The $500-per-month Pro 500 tier is a usage allowance plus the Ultrafast mode, not a per-dot rate, so even the most expensive plan on OpenAI's sheet does not convert into a cost per unit of agent work.

Read that as a description of a product category rather than a gap in the documentation. A dot is priced like a seat, not like a workload. Seats are easy to budget and impossible to optimise; that is the whole trade.

Two bills, one of which you can forecast

Put the two cost models next to each other and the difference is not size, it is shape. GPT-5.6 Sol's cost is a straight line through the origin that gets steeper past 272,000 input tokens. A dot's cost is a flat line that tells you nothing about how far along it you are — until it changes.

That second property is the one that surprises teams. A capacity change, a speed tier change or a change to what the first dot includes arrives as a support email and a new price, with no equivalent to a rate-card revision you can diff. A token price change is a number that moves; a subscription change is a policy that moves. Both are legitimate, and only one of them is something you can model in a spreadsheet.

Where the two shapes belong in one pipeline

A dot earns its place on work that has no natural unit: chasing a decision across three apps over four days, keeping a project moving when nobody is watching it. GPT-5.6 Sol earns its place on work that does. Extracting structured records from a long archive, reviewing a large repository in one pass, generating a long document from a long source — those tasks have a beginning, an end and a token count, and the 1.05-million-token window is the reason some of them are possible at all rather than a convenience for the ones that already were.

OrcaRouter sits entirely on the metered side of that line. The model is routed as openai/gpt-5.6-sol at OpenAI's list price passed through with no markup on top, in the same catalogue as 200-plus other models behind one key, with automatic failover across upstream providers and a routing DSL for sending an individual request to the model that should answer it. What that does not buy you is anything inside a dot: there is no Dots API, no model identifier to point a request at, and no way to substitute the intelligence running underneath one. The catalogue reaches the rate card and stops there.

Which purchase is a mistake

Buying a dot expecting to optimise it is a mistake, because the surface you would optimise is not exposed. Buying GPT-5.6 Sol expecting it to chase a project for you is the same mistake in the other direction — a model that answers when called is not a worker that keeps going when you stop calling.

The teams that get this right treat the two as different line items with different verbs attached: a subscription for the work that has to keep happening, and a metered model for the work that has to be counted. The subscription is the one you cancel; the meter is the one you forecast. Knowing which is which before you commit is most of the decision.