Dots vs GLM-5.2: The Two Ends of the Agent Supply Chain

Dots and GLM-5.2 sit at opposite ends of the same supply chain, and neither one is the other's rival. Dots is the agent surface the company announced at DevDay on September 29, 2026 — an always-on worker with its own cloud computer and browser, connections to more than 4,000 apps, reachable in ChatGPT, Slack and Teams, powered by GPT-6 Astra and included with Pro and Business Premium at no extra cost. GLM-5.2 is Z.ai's open-weight flagship text model, released June 16, 2026, with a one-million-token context window, up to 128,000 tokens of output and a published rate of $1.40 per million input tokens and $4.40 per million output tokens. One is the layer you rent and cannot see inside; the other is the layer you can download, inspect and price to the token. If you are deciding how to run agents this quarter, that is the pair of decisions in front of you.

What each one is, in one pass

• What it is — a hosted always-on agent with no model identifier of its own (Dots) against a downloadable text model with published weights (GLM-5.2)

• Who makes it — OpenAI, announced at DevDay on 2026-09-29, against Z.ai (Zhipu AI), released 2026-06-16

• Where it runs — OpenAI's cloud computer and browser, isolated from your machine unless you link them, against wherever you deploy it: your own hardware, a cloud of your choosing, or our endpoint

• Model underneath — GPT-6 Astra, named by OpenAI in its launch materials (vendor-stated), against GLM-5.2 itself, a mixture-of-experts model that Z.ai has released with open weights

• Modality — messages, app data and a browser at one end, against text in and text out only, with no image or audio input

• Capacity — undocumented response limits for a dot, against 1,000,000 tokens of context and 128,000 tokens of maximum output

• Price — the first dot included with Pro or Business Premium, allowance unpublished, against $1.40 per million input and $4.40 per million output as listed in our catalogue

• Evidence — vendor demos, no independent benchmark for dots, against an Artificial Analysis Intelligence Index of 33.7 and a vendor-reported 99.1 on τ²-Bench, both carried with their sources in our catalogue

The open-weights half, and the day-one catch

GLM-5.2's proposition is the familiar open-weight one, and it is stronger than usual here because the model is cheap by any measure and the weights mean the deployment decision is yours. Z.ai's own published results lead with agentic tool use — 99.1 on τ²-Bench, 76.8 on MCP-Atlas and 74.4 on FrontierSWE — and those are vendor figures, which is worth saying plainly because they are the numbers that make the model look like a bargain. Independent measurement is more measured: an Artificial Analysis Intelligence Index of 33.7 puts it mid-field on that board, a τ-banking slice of 34.6 is modest, and 78.3 on long-context recall sits behind the current leaders. Our catalogue carries 123.2 million tokens of traffic on this model over the trailing seven days, which is the clearest sign available that people are actually running it.

Here is the catch, and it is the reason this article exists rather than a straight spec comparison. Artificial Analysis marks GLM-5.2 as deprecated in favour of a newer release, GLM-5.3, and its live page says so at the top: the model is still routable and still generally available, but the vendor's roadmap has moved on and the index figures attached to GLM-5.2 are a snapshot rather than a standing measurement. Its other independent figures stay static for the same reason — AA's notice explains that after deprecation it continues to benchmark only the default workload for the model, and results for everything else become historical.

Two lessons follow from that, and both apply well beyond this model. First, a price and a licence are durable facts about an open-weight release; a leaderboard position is a moving one, and a model can be cheap, downloadable and nonetheless superseded. Second, the successor is already here: Z.ai has shipped GLM-5.3 as its current flagship, priced below GLM-5.2 at $1.26 input and $3.96 output per million with an independent Intelligence Index of 44.8 against GLM-5.2's 33.7. Anyone building now should price the newer model first and treat GLM-5.2 as the cheaper floor under it, not as the frontier of the family.

What a dot costs, and why neither of you can budget for it

OpenAI has published one price for dots — the first one is included with Pro and Business Premium, with extended limits in the first month and conversations that do not draw down ChatGPT usage limits. No allowance figure, no price for a second dot, no rate for extra speed or capacity, no cost per finished task, and no enterprise price for the specialist dots currently in internal testing. CNBC's keynote coverage has finance chief Sarah Friar pricing the new $500 Pro 500 tier as a usage allowance plus the Ultrafast mode rather than as a per-dot rate; the most expensive plan on OpenAI's sheet does not convert into a cost per unit of agent work either.

GLM-5.2 has the opposite problem, which is the better problem: too many numbers, all of them checkable. The rate card is $1.40 and $4.40 per million tokens with cached reads at $0.26, and the model is downloadable, so the marginal cost of a request you run yourself is hardware and electricity rather than a vendor's invoice. That is the trade in one sentence. A dot gives you a worker and no meter; GLM-5.2 gives you a meter and no worker.

Putting a metered model under an agent you rent

The useful way to hold both is to give each one the half of the job it does well. A dot handles the coordination: watching a project, chasing the people and systems involved, coming back with something finished. GLM-5.2 — or the successor you have priced against it — handles the reasoning you need to own: the extraction pass over a million-token archive, the long generation, the classification queue, the part where a wrong answer costs more than the tokens it saved.

That split needs the reasoning half to be portable, which is a routing problem rather than a model problem. OrcaRouter serves GLM-5.2 as z-ai/glm-5.2 at provider list price passed through with 0% markup — nothing added on top, so a Z.ai rate cut reaches your account the same day — in a catalogue of 200-plus models behind a single key, with the failover and routing tools that let a job move to the newer or cheaper model without a code change. What it does not offer is a dot: there is no dots endpoint, no model identifier for the agent, and no way to swap the intelligence running inside it. The boundary is real, and pretending otherwise would be the one dishonest slide in this comparison.

The recommendation, stated once

If you want the work to happen without you driving it, buy the dot — it costs nothing beyond a subscription you may already hold, and its limitation is that you cannot yet prove what it saved you. If you want to know what your reasoning costs and to be able to move it later, put GLM-5.2 or its successor behind your own endpoint and meter it. The teams getting this right in late 2026 are doing both, and the discipline that makes it work is refusing to let the rented layer own anything they would be unable to replace.