Published · 11h agoBuild2 min read
The circulating growth multiple is not in the supplied data. What is there: OpenRouter's per-request figure, plus two reports on where agent output stops moving.
Written for builders.See today for builders
Compiled by The EngineerSomething wrong?How this is made
The multiple in circulation is a growth curve; the multiple in the data is a ratio. None of the material at hand carries a 14x figure for agent token growth since February [17]. It carries OpenRouter's per-request measurement, which tells you how much an agentic call consumes relative to a human one [2], and a crossover date, which tells you when agents became the majority consumer of tokens on the platform [1]. Those are different objects, and only the ratio explains a backlog.
It explains one because tokens arrive as diffs and diffs arrive at a person. At Augment, where agents write all the code [10], the engineer at the front of the queue had six pull requests ahead, each around 400 lines written by nobody on the team [11]. That is 2,400 lines of reading standing between a two-line change and its first human comment [13].
The Faros numbers, reported by Luca Rossi at Refactoring, put the same shape across 4,000-plus teams and 22,000-plus developers [4]. Divide the increase in PRs touched per day by the decline in weekly deployments and you get roughly 90 percent more work in flight per unit actually shipped [14], which is the mechanical reason lead time landed at plus 480 percent with an extra 80 percent of waiting inserted between every pipeline step [7].
Nothing in the price signal argues for slowing the generator. DeepSeek V4 Flash output costs $0.18 per million tokens against GPT-5.5's $30 [3], about 167 times cheaper [15], and OpenRouter's segmentation shows agentic traffic moving onto V4 Flash within a month of release while human usage stayed on V3.2 [18]. Generation gets cheaper on a release schedule. Reading does not.
Augment's response was to stop sending everything to a human: a risk classifier auto-approved 10 percent of PRs to start [12]. Against a 1,400-PR queue, that still leaves 1,260 for people [16].
Ranked by verification strength, evidence, and original report placement.
OpenRouter reports that tokens used by agentic workloads surpassed those used by humans around February 1, and that the 2026 explosion in agentic tokens did not kick off in earnest until early February.
According to OpenRouter data, agentic work burns about 15x more tokens per request than normal human AI usage.
DeepSeek V4 Flash on the cheapest endpoint costs $0.09 input and $0.18 output per million tokens; GPT-5.5 is priced at $5 input and $30 output per million tokens.
The Faros report, covered by Luca Rossi at Refactoring, surveyed more than 22,000 developers across more than 4,000 teams.
Faros found PRs touched per day up 67 percent, separate tasks up 18 percent, and tasks completed up 33 percent as measured by opening a final PR.
Faros found deployments per week down 12 percent, in-progress tasks stalled for 7 or more days up 26 percent, and work restarts per developer up 14 percent.
The reporting this story was synthesized from, earliest first. Every link goes to the original.