Cognition reports 4.8x token throughput on Nvidia Vera Rubin in its own coding-agent test
Cognition says its SWE-2 model produced up to 4.8 times more total token throughput on Nvidia's Vera Rubin NVL72 than on GB200, in a test on CoreWeave. The company-run result points to more agent capacity per system, but whether long coding tasks get cheaper depends on what the new systems cost per hour.
The Engineer · Build desk
What happened
- CoreWeave identified Cognition as the first customer running production workloads on Vera Rubin, after the two stood up its cluster in early September.
- The test used SWE-2 inference workloads built from a sample of tasks in Cognition's FrontierCode benchmark, measured against a GB200 NVL72 baseline.
- CoreWeave separately reported 3.8 times higher output-token throughput on reinforcement-learning workloads.
- The September 30 announcement followed Cognition's September 8 Series E, which it said raised more than $2 billion at a $48 billion valuation.
Why it matters
- Who has to decide decision Cognition has to choose whether the extra capacity buys more concurrent sessions, shorter agent cycles or heavier use of each deployment, and capacity pricing decides which of those pays off.
- Who's exposed exposure Devin training, reinforcement learning and production inference already run on thousands of CoreWeave GPUs, so how fast Cognition moves onto Vera Rubin depends on one provider's rollout.
- What it sets up precedent Vera Rubin's first production customer is a coding-agent company, so the first customer benchmark buyers can point to reflects a coding agent's token mix and not their own.
Next story · Product · Sep 30
Factory's CEO accuses an advisor turned Cognition executive of sharing confidential informationConfirmed2 publishers
'Up to' is the most honest phrase in the announcement. RuntimeWire calls the 4.8x an upper-end throughput claim for one SWE-2 test [13]. It describes the figures as vendor-associated results, not an independent replication [11].
I think the most useful line in the release is CoreWeave's statement that the gain came with no loss in generation speed [7]. A throughput gain bought by slowing each request would cost a coding agent time on every step. The agent makes repeated model calls as it plans, edits, tests and revises code [5]. CoreWeave's reinforcement-learning multiple counts output tokens only, so it measures something different from the total-token 4.8x [8][1].
A coding agent that runs through many steps can consume far more inference than a single prompt-and-answer exchange [6]. The business paying for that inference is also growing. By the company's own measure, Cognition's run-rate revenue rose from $492 million in May to almost $900 million [18], roughly 1.8 times [20].
For the multiple to carry over to another operator, several things would need to hold. That operator's token mix would need to resemble SWE-2 on Cognition's task sample [9]. Nvidia describes those tasks as real-world software-engineering work [10]. Its GB200 baseline would need to match the one Cognition's engineers measured against [9]. And its workload would need to land near the top of the range [13]. The published materials do not include absolute throughput, system pricing, energy use or enough configuration detail to reproduce the comparison [12]. None of those conditions can be checked.
Cost per token is a system's hourly price divided by the tokens it produces in that hour [19]. At the full 4.8x, a Vera Rubin NVL72 matches a GB200 NVL72 on cost per token if it costs 4.8 times as much per hour, and beats it at any lower price [19]. At a smaller multiple, the break-even price falls in step [19]. Token counts also stop short of completed work. By RuntimeWire's reading, the result does not establish 4.8 times more completed coding tasks or lower total cost across workloads [16]. Whether an agent's output stays useful as throughput rises is a separate question [14].
What to watch
- CoreWeave's hourly pricing for Vera Rubin NVL72 against GB200 NVL72, which sets whether 4.8x becomes a lower cost per token.
- An independent benchmark of Vera Rubin against GB200 on agent workloads that publishes absolute throughput and full configuration.
- Whether Cognition reports completed tasks or cost per task on Vera Rubin, in place of token throughput alone.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption30
- Hype gap+35
- Incentives85
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Cognition says its SWE-2 model produced up to 4.8 times more total token throughput on Nvidia's Vera Rubin NVL72 than on a GB200 NVL72 system; the test ran on CoreWeave.
- [2]
CoreWeave identified Cognition as the first customer running production workloads on Vera Rubin.
- [3]
The announcement came on September 30th.
- [4]
Cognition worked with CoreWeave to stand up its Vera Rubin cluster in early September, according to CoreWeave.
- [5]
A coding agent can make repeated model calls as it plans, edits, tests and revises code.
- [6]
An agent that runs through many steps can consume far more inference than a single prompt-and-answer exchange.
- [7]
CoreWeave said the throughput improvement came with no loss in generation speed.
- [8]
CoreWeave's release reported 3.8 times higher output-token throughput on reinforcement-learning workloads.
- [9]
Cognition's engineers compared Vera Rubin and GB200 NVL72 using SWE-2 inference workloads built from a sample of tasks in Cognition's FrontierCode benchmark, with GB200 NVL72 as the baseline.
- [10]
Nvidia's account describes the workload as real-world software-engineering tasks.
- [11]
The results are vendor-associated, not an independent replication.
- [12]
The published materials do not provide absolute throughput figures, system pricing, energy use or enough configuration detail to reproduce the comparison.
- [13]
The 4.8x result is an upper-end throughput claim for this SWE-2 test.
- [14]
More tokens per unit of time could allow Cognition to run more sessions on a given deployment, shorten some agent cycles or use available compute more intensively; which benefit dominates depends on how capacity is priced and whether the agent's work remains useful as throughput rises.
- [15]
CoreWeave said Cognition expanded to thousands of GPUs on its platform in less than nine months, using it for Devin training, reinforcement learning and production inference.
- [16]
The 4.8x result does not establish a 4.8x increase in completed coding tasks, nor does it show lower total cost across workloads.
- [17]
On September 8th Cognition announced a Series E that it said raised more than $2 billion at a $48 billion valuation, led by Andreessen Horowitz and Accel.
- [18]
Cognition said run-rate revenue had grown from $492 million in May to almost $900 million; the figure is company-reported and the calculation was not defined in detail.
- [19]
Cost per token is hourly system price divided by tokens per hour. At the 4.8x upper bound, a Vera Rubin NVL72 has equal cost per token to a GB200 NVL72 when its hourly price is 4.8 times GB200's, and lower cost per token at any lower hourly price; at a smaller throughput multiple, the break-even price ratio falls to that multiple.
- [20]
Cognition's reported run-rate revenue rose roughly 1.8 times between May and its September 8 announcement.
Sources
2 independent publishers whose own reporting we read for this story.
- nvidianews.nvidia.comFrom Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI
1 article · September 30, 2026
- runtimewire.comCognition puts Nvidia's Vera Rubin to work, reports up to 4.8x token throughput
1 article · September 30, 2026