Among the figures circulating after Reuters’ Anthropic exclusive is a seemingly small, strategically huge fraction: GPT-6 Astra accounts for about 13% of enterprise AI spend Ramp tracks, versus roughly 8% attributable to Claude Fable. It is not a census of the world; it is a sensor from corporate cards and expense software biased toward certain segments. Even so, for pre-IPO investors, the sensor is enough to trip alarms.
Ramp also publishes a September AI Index describing “cracks in the AI thesis”: enterprises route tokens away from the most expensive frontier models, frontier-token share falls from August peaks, and effective price per million tokens drops sharply. In that climate, a new model combining high capability with better economics — Astra — has a tailwind.
The apparent paradox is that Anthropic still appears in other Ramp cuts with strong account penetration (in some snapshots, more businesses use Claude than OpenAI). Account penetration and spend share are not the same: you can sit in more companies with light seats and still lose dollars when intensive workloads migrate to a cheaper-per-task rival.
That explains the nervousness among some prospective Anthropic investors described by Reuters. Enterprise leadership was narrated as a stable moat; spend data suggests a permeable moat once OpenAI recovered product narrative with Astra. Three percentage points are not a defeat; the slope can be if it persists.
Methodologically, caution is warranted. Ramp observes its customers, not the full Fortune 500 or the public sector. Large annual contracts negotiated off-card may be underrepresented. Models hosted in clouds with aggregated billing can confuse attribution. The numbers are directional, not verdicts.
Even with caveats, the industry will use them because they are timely and quantified. OpenAI sales decks can already contrast “13 vs 8” beside Artificial Analysis tables. Anthropic’s must answer with logo retention, document-workflow depth, and total TCO proof — including prompt engineering and avoided failures — not token price alone.
For CFOs, Ramp’s September index offers another reading: spend per employee among top spenders compressed, a sign of post-euphoria optimization. In that environment, every share point is fought with calculators, not keynotes. Astra arrives just as corporate buying grows up.
There are product implications too. If spend concentrates in coding agents and automation, labs will win share where they reduce integration friction (IDE, tickets, CRM). An intelligence tie does not move budget; a reliable connector and a predictable invoice do.
Startups selling atop third-party APIs feel the move twice: their own COGS shifts when the underlying model’s price shifts, and their customers ask why they have not migrated to the month’s cost leader. Ramp share volatility transmits into application-layer margins.
Anthropic, if it launches a new pre-IPO model, will need to show not only that it regains index points but that it regains spend points. A launch that impresses social media and leaves the Ramp invoice untouched would be an incomplete victory for the listing thesis.
OpenAI, symmetrically, must ensure share growth is not paired with incidents that scare the same buyers. The 13% is a fragile asset if an agent exclusive pushes risk committees to freeze expansions.
Neural Pulse advises editors to treat the figure as what it is: an early indicator of budget reallocation, not a moral ranking of labs. The useful question is which workloads moved and under what controls.
Taken together, Ramp turns a war of narratives into a war of percentages. Astra has printed the first digit Anthropic did not want to see on the eve of its IPO. The rest of autumn will tell whether that 13% was a novelty spike or the start of a durable redistribution.
We will keep cross-checking AI Index updates against cloud-provider data and adoption surveys. Until then, 13 versus 8 is the numeric postcard of September 2026’s competitive moment.
