Dots vs Claude Opus 5: A Rented Worker Against a Metered Model

Dots and Claude Opus 5 are not competitors, and the fastest way to see that is to try to write either one's price next to the other's. Dots is the company's always-on agent, announced at DevDay on September 29, 2026: it gets its own cloud computer and browser, reaches more than 4,000 apps through the company's plugin ecosystem, runs on GPT-6 Astra, and is included with Pro and Business Premium at no extra cost. Claude Opus 5 is Anthro​pic's flagship reasoning model, released July 24, 2026, billed at $5.00 per million input tokens and $25.00 per million output tokens with a one-million-token context window and up to 128,000 tokens of output. One is a worker you rent by the month; the other is a model you meter by the token. The comparison only becomes useful once you stop asking which is better and start asking what you would own afterwards.

The one-line answer

If your problem is a long-running task that lives inside other people's software — chasing a bug across Slack threads and ticket systems, assembling a deliverable out of app data — a dot is the product built for it and Claude Opus 5 is a component that might run inside it. If your problem is a single hard call you need to make correctly, on input you already have, on a budget you have to forecast, Claude Opus 5 is the thing you call and a dot is an expensive way to reach the same intelligence. The overlap is real but thin: both are useful to the same team, at different layers.

Two things that are not the same kind of thing

• What you buy — an always-on agent with a cloud computer, a browser and app connectors (Dots) against a text-in, text-out reasoning model with tool and image input (Claude Opus 5)

• How it is priced — included with a Pro or Business Premium subscription, allowance unpublished (Dots) against $5.00 per million input and $25.00 per million output, cached reads at $0.50 (Claude Opus 5)

• What the price tracks — nothing published, so cost does not scale visibly with work done (Dots) against tokens, so cost rises with every document you send and every token it writes (Claude Opus 5)

• Where the work happens — on OpenAI's cloud computer, isolated from yours unless you link them, with a browser you can watch (Dots) against wherever you run it: Anthropic's API, a cloud of your choosing, or the same endpoint you already call for everything else (Claude Opus 5)

• Who can inspect the result — you, in an Activity View that shows background progress (Dots) against you, in the response, the tool calls and the reasoning trace your client records (Claude Opus 5)

• Published evidence — vendor capability statements and demos, no independent benchmark (Dots) against an Artificial Analysis Intelligence Index of 50.8 and an Agentic Index placement that independent trackers re-measure every few weeks (Claude Opus 5)

What Claude Opus 5 costs, in units you can meter

The reason Claude Opus 5 is easy to write about is that every claim about it has a denominator. A one-million-token context accepts an entire codebase or a decade of contracts in one call; up to 128,000 output tokens means one response can be longer than most reports. Those two numbers set the ceiling of a single request, and the rate card sets what it costs: at $5.00 and $25.00 per million, a 300,000-token prompt with a 20,000-token answer comes to roughly $2.00, and cached input at $0.50 per million drops the repeat cost of the same long preamble by ninety percent.

That arithmetic is the whole reason teams keep a metered model in the stack even after they buy an agent. An agent's allowance is a ceiling you discover by hitting it; a token meter is a curve you can plot before you ship. On OrcaRouter the model is routed as anthropic/claude-opus-5 at Anthropic's list rate with 0% markup — the provider's price, passed through, nothing added — and it sits in the same catalogue as 200-plus other models behind a single key, which means the model you use to sanity-check a plan and the model you route in production are reachable without a second contract.

What a dot costs, and why the number is missing

OpenAI has published exactly one price for dots: the first one is included with Pro and Business Premium, with extended limits in the first month and, in the company's words, 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, no per-task cost, and no enterprise price for the specialist dots that are currently in internal testing. CNBC's account of the keynote has finance chief Sarah Friar pricing the $500 Pro 500 tier as a usage allowance plus the new Ultrafast mode, not as a per-dot rate — so even the most expensive plan on the sheet does not convert into a cost per unit of agent work.

That is not a scandal; it is the shape of a subscription product in its first month. It does mean the honest comparison is asymmetric. You can price a job on Claude Opus 5 to the cent before you run it. You can price a dot only by comparing your monthly bill against the work you believe it did, which is how most people price their own time and almost nobody prices an API.

The workflow where the two actually meet

The place these two systems touch is the point where an agent hand-off meets a metered call. A dot can carry a project forward while you do something else — that is its entire pitch — but the reasoning inside any custom loop you own is a token bill you control, and the most common shape in production today is the one that keeps those layers separate: a hosted agent for the connective work, and a metered model behind your own endpoint for the decisions you need to audit, retry, or run against a different vendor next quarter.

OrcaRouter sits on the metered side of that boundary. One key reaches Claude Opus 5 and 200-plus other models at provider list price with no markup, so a rate cut from Anthropic lands on your bill the same day rather than at the next contract renewal; automatic failover and the routing DSL exist so that a long-running job started on a frontier model can degrade to a cheaper sibling mid-flight instead of failing outright. None of that reaches inside a dot — there is no API for one, no model identifier to route to, and no way to swap the intelligence inside it. Anyone offering you a "Dots endpoint" is selling something else.

How to decide this week

Buy the dot if you have a class of work that is currently being dropped — things that need to happen across several apps, on a schedule, without you opening a chat window. It costs nothing beyond the subscription you may already pay, and the honest caveat is that you cannot yet measure whether it saves you money; you can only observe whether the work gets done.

Keep Claude Opus 5 metered if the work has a deadline and a budget attached, or if the output has to be reproducible and auditable. Route it through an endpoint you control, keep the rate card in front of you, and treat the agent layer above it as replaceable — because in twelve months the layer that looks permanent today is the one most likely to have been renamed, and the token meter underneath will still be the thing you can forecast.