Dots vs DeepSeek V4 Pro: What Your Money Buys in Each Case

Put Dots and DeepSeek V4 Pro on the same page and the comparison looks absurd for about ten seconds, which is exactly how long it takes to notice that they answer the same question from opposite ends. Dots is the company's always-on agent, announced at DevDay on September 29, 2026: a persistent worker with its own cloud computer and browser, connectors to more than 4,000 apps, reachable inside ChatGPT, Slack and Teams, running on GPT-6 Astra, and included with Pro and Business Premium at no extra charge. DeepSeek V4 Pro is a downloadable frontier text model — 1.6 trillion total parameters with about 49 billion active per token, a 1,048,576-token context window, up to 384,000 tokens of output — that you can run yourself or call at $0.66 per million input tokens and $1.98 per million output tokens through our catalogue. One is a worker you cannot measure; the other is a commodity you can.

That is the whole argument, and it survives contact with the details. What follows is the detail: what each one actually publishes about itself, where the two price curves come from, and the specific job shapes where picking one over the other is a mistake rather than a preference.

What each one publishes

• Identity — a hosted agent surface with no model identifier of its own (Dots) against a named, downloadable model you can serve on your own hardware (DeepSeek V4 Pro)

• Underlying model — GPT-6 Astra, named by OpenAI in its launch materials as the model behind dots (vendor-stated) against DeepSeek V4 Pro itself, weights published by DeepSeek

• Architecture — not disclosed for dots beyond the model it runs on, against a sparsely activated mixture of experts at roughly 1.6 trillion total and 49 billion active parameters

• Context and output — undocumented response limits for a dot, against 1,048,576 tokens of context and 384,000 tokens of maximum output per request

• Modality — a dot works across apps and a browser, against text in and text out only, with no image, audio or video input

• Price — the first dot is included with Pro and Business Premium, allowance unpublished, against $0.66 per million input and $1.98 per million output as currently listed in our catalogue

• Evidence — vendor demos and capability statements, no independent benchmark for dots, against a catalogue-listed 96.2 on τ²-Bench and an Artificial Analysis Intelligence Index of 36 with the model placed 39th on that board

The cheapest frontier model is the one nobody argues about

DeepSeek V4 Pro has been generally available since April 24, 2026, and its selling point is not a benchmark headline — it is that a model of this class costs less than a tenth of what a closed frontier model charges, and you can stop paying entirely by downloading it. That second clause is the one with teeth. If DeepSeek stopped serving tomorrow, the weights would not disappear; they would keep running on whatever hardware you had already bought, at whatever quantization fitted. That is a different kind of guarantee from a subscription, and for a team building a product on top of an agent, it is the difference between a dependency and a component.

The published numbers sit where you would expect a well-engineered open model to sit rather than at the top of the board: 92.8 on GPQA Diamond and 80.3 on long-context recall are competitive; the Artificial Analysis Intelligence Index of 36 with a 39th-place ranking is mid-field, and the agentic tool-use headline of 96.2 on τ²-Bench is the figure DeepSeek leads with — all of it listed in our catalogue against its sources, none of it measured by us. The honest read is that this model wins on cost per unit of work and on deployability, and does not win on raw reasoning against the current frontier.

Two price curves, one of which moves without warning

A dot's price is a subscription line and nothing more: you pay for Pro or Business Premium, you get one dot, and OpenAI has not told anyone where the allowance ends. There is no allowance figure, no price for a second dot, no published rate for extra speed or capacity, and no per-task cost. CNBC's keynote coverage has finance chief Sarah Friar pricing the $500 Pro 500 tier as a usage allowance plus the new Ultrafast mode rather than as a per-dot rate, which means the most expensive plan on OpenAI's sheet still does not convert into a cost per unit of agent work.

DeepSeek V4 Pro's price is a rate card, and a rate card can be audited in a way a subscription cannot. This one has a wrinkle worth knowing before you budget: our catalogue carries two figures for the same model — the overview text describes $0.44 per million input and $0.87 per million output, while the pricing record lists $0.66 and $1.98 — and it also carries timed windows that double the listed rate during two periods each day, 01:00–04:00 and 06:00–10:00 UTC. That is not a reason to avoid the model; it is a reason to read the rate card rather than the summary, and a reminder that a number in prose and a number in a pricing table can drift apart even on a page that is trying to be accurate.

The same distinction decides what happens to your bill when a vendor moves. A rate cut from DeepSeek is live on OrcaRouter the day it lands, because the catalogue passes provider list price through with 0% markup and adds nothing on top. A change to a dot's allowance arrives as a support email. Both are fair; only one of them is forecastable.

Where the two belong in one pipeline

The mistake is treating this as a bake-off. A dot is the right shape for work that is currently being dropped: a project that needs chasing across apps on a schedule, where the value is that somebody-or-something keeps moving. DeepSeek V4 Pro is the right shape for work that has a unit: extracting structured records from a million-token archive, generating a long document from a long source, running the same classification over a queue, or fanning a hard problem across many cheap attempts.

Those shapes are complementary, and the plumbing between them is one endpoint. OrcaRouter serves DeepSeek V4 Pro as deepseek/deepseek-v4-pro at provider list price with no markup, in the same catalogue as 200-plus other models behind a single key, with automatic failover across upstream providers and a routing DSL for sending an individual request to the model that should handle it. The layer above it — the agent that decides what to send — is where you choose between a dot you rent and a loop you own. Nothing we offer reaches inside a dot: there is no dots API, no model identifier, no way to swap the model running it. If you want the reasoning under your control, you build the loop and point it at a metered model.

The question that decides it

Ask who holds the work when it goes wrong. With a dot, OpenAI holds the computer, the browser, the connectors and the allowance, and your remedy is to stop using it. With DeepSeek V4 Pro, you hold a copy of the model, a rate card and the option to move it somewhere else. Neither answer is better in general — renting is how most teams should start — but only one of them lets you write the failure mode down in advance, and for anything you intend to depend on for a year, that is usually the deciding fact.