The two products supplied 107.2 percent of net AI-session growth in Graphite's six-product sample from April to September. A separate coding benchmark shows why enterprises should buy agents for measured, accepted work rather than assume unattended labor substitution.
Graphite estimates 76 billion worldwide AI-product web and mobile sessions in September 2026, 19 percent above January's 64 billion. The composition of the rebound tells us more. From April to September, ChatGPT and Claude together added 10.386 billion sessions, exceeding the 9.690 billion net increase across all six products in Graphite's dataset.
CTOL's reading is that measured AI traffic growth is becoming more concentrated, while reliable unattended coding remains a distinct technical challenge. A separate SWE Cycle study reports FullCycle solve rates of just 2.15 percent to 13.50 percent across six coding agent models on a demanding end-to-end benchmark. These are not comparable growth measures: one tracks product activity, the other tests whether an agent can complete a technical workflow without human scaffolding.
For investors, the distinction separates observable distribution strength from assumptions about future labor substitution. For engineering leaders, it supports a more immediate decision: deploy agents first in scoped workflows with review and acceptance checks, and authorize broader unattended execution only when completed work costs less than the existing process at comparable quality.
The traffic rebound has two principal contributors
Graphite's State of GenAI study, published November 9, 2026, estimates worldwide sessions across ChatGPT, Claude, Gemini, Grok, Perplexity and DeepSeek. Its September figure represents 37 percent growth from September 2025 and 19 percent growth from January 2026.
The April-to-September product data identify where the more recent increase came from.
| Product | April 2026 | September 2026 | Change |
|---|---|---|---|
| ChatGPT | 53.752 | 61.351 | +7.599 |
| Claude | 2.649 | 5.436 | +2.787 |
| Gemini | 3.548 | 3.718 | +0.170 |
| DeepSeek | 3.422 | 2.972 | -0.450 |
| Grok | 1.597 | 1.380 | -0.217 |
| Perplexity | 0.868 | 0.669 | -0.199 |
| Six-product total | 65.836 | 75.526 | +9.690 |
Source: CTOL calculations from Graphite's monthly product estimates. Figures are billions of estimated web visits plus mobile app sessions. The September total rounds to Graphite's reported 76 billion. Sessions are not deduplicated users, paid seats or revenue; desktop applications are excluded.
ChatGPT supplied 7.599 billion additional sessions, equivalent to 78.4 percent of the six-product net increase. Claude added another 2.787 billion, or 28.8 percent. Together, their contribution was 107.2 percent of net growth because Gemini, DeepSeek, Grok and Perplexity collectively lost 0.696 billion sessions.
The concentration measure tells the same story. ChatGPT and Claude's combined share of this six-product traffic universe rose from 85.7 percent in April to 88.4 percent in September. ChatGPT alone represented approximately 81.2 percent of September sessions.
The distribution finding matters. It suggests that the incremental measured activity favored two services rather than lifting every participant equally. It does not establish paid enterprise conversion or economic returns, but it identifies where the observed usage advantage has strengthened.
For investors examining consumer-facing AI platforms, that is the signal to carry into assessments of distribution and user engagement. It should not be converted directly into a claim about autonomous-agent productivity.
The harder coding test exposes a completion problem
The SWE Cycle study, released in a May 2026 version-one paper, evaluates 489 filtered instances across six coding agent models.
Its FullCycle protocol requires an agent to reconstruct an environment, implement a requested change and generate verification tests in a bare repository without human scaffolding. A solve is recorded only when the study's end-to-end binary success criterion is met.
Across the evaluated models, FullCycle solve rates range from 2.15 percent to 13.50 percent. Equivalently, 86.50 percent to 97.85 percent of evaluated attempts did not qualify as successful FullCycle solves under the protocol.
That is the important result for buyers considering unattended repository changes. The agent must do more than produce code: it must establish the working environment, make the change and provide verification that satisfies the evaluator.
The paper also reports isolated-task results ranging from 28.83 percent to 78.12 percent. Those come from separate environment, implementation and test-generation tasks. They are not matched observations from which to calculate a measured decline when tasks are combined.
The economic consequence is not that every unsuccessful benchmark attempt becomes an equally expensive production failure. Rather, an unattended-completion product must overcome the whole workflow, including the steps around code generation. A strong result on a narrower component does not remove that burden.
There is one material limitation in applying this finding: SWE Cycle tests unscaffolded work in bare repositories, while production teams can use established environments, reusable tools, narrower tasks and human checkpoints. Those differences can change both completion rates and costs.
The limitation narrows the conclusion, but it does not make the observed FullCycle result irrelevant. It makes a controlled production trial, rather than an isolated benchmark score, the appropriate basis for spending on autonomy.
Enterprise value is determined per accepted change
The two studies describe different commercial exposures. Graphite provides evidence about where AI-product activity is growing. SWE Cycle provides evidence about the technical difficulty of completing a particular class of autonomous coding task.
Neither directly measures enterprise return on investment. But buyers do not have to stop at that absence of data. They can define the economic test an agent must pass.
For a coding deployment, the relevant unit is fully loaded cost per accepted change:
Cost per accepted change = (agent and tool expenditure + human review and repair + testing, security and integration costs) / accepted production changes
An accepted change must satisfy the same quality and security requirements used for the existing engineering process. A comparison also needs a similar task cohort and an equivalent accounting of human effort, including failed attempts and recovery work.
This makes the trade-off measurable.
An agent that accelerates the initial implementation but requires extensive debugging, review and integration may still have value as an assistant. It has not necessarily reduced the total cost of delivering accepted code. Conversely, a deployment with strong tools and human checkpoints may be economical even if it does not resemble the unattended conditions in SWE Cycle.
That distinction changes procurement. An engineering organization can justify limited agent adoption through measured improvements in reviewed workflows without assuming that entire engineering tasks can already be delegated unattended.
For suppliers selling autonomous execution, the commercial burden is higher. They need evidence of completed production work and the cost of obtaining it, not simply growth in usage or isolated-task capability.
The benchmark cannot calculate those costs. Its low FullCycle solve rates nevertheless argue against making unassisted completion the default assumption for unfamiliar-repository work.
What would change the investment and procurement call
As of November 9, 2026, CTOL's view is that AI distribution strength is observable and concentrated, while unattended FullCycle-style coding remains a reliability risk rather than a demonstrated source of labor substitution.
Confidence is high in the within-sample traffic decomposition and moderate in the enterprise deployment implication. The former follows directly from comparable Graphite observations; the latter depends on how demanding benchmark conditions transfer into production workflows.
Set February 9, 2027 as the next decision checkpoint, rather than assuming a new study will be published on that date. The evidence that would change the procurement default is a comparable production coding cohort showing accepted changes at equivalent quality, with disclosed human intervention, security overhead and fully loaded cost below a human-only baseline.
A later SWE Cycle evaluation using the same protocol could separately establish technical progress if solve rates improve materially. Another increase in consumer sessions would test demand and distribution, not unattended execution economics.
Until that production evidence exists, the defensible position is to recognize the distribution gains, deploy agents where reviewed output demonstrably saves resources, and resist treating gross usage growth as a substitute for accepted work.