Loading today’s stories
Invest1 distinct publisher3 min readPublished
The 4.4x per-token premium in Vercel's gateway data only computes at the friendliest end of its own range, and the per-seat figures suggest buyers may be paying for fewer tokens rather than dearer ones.
The Investor · Invest desk
Compiled by The InvestorSomething wrong?How this is made
Work the 4.4x backwards [2] and it needs a friendly assumption at every step. Take the top of Anthropic's spend range against the bottom of its volume range, 65% of the money on 30% of the tokens, then assume OpenAI absorbs every token Anthropic does not (all 70% of them, for its 35% of spend), and the premium computes at 4.33 [1]. Swap the ends, 61% on 32% against 35% on 68%, and it is 3.70 [2]. Midpoints give 4.01 [3]. The load-bearing assumption is a two-name token market, and while Anthropic plus OpenAI genuinely is 96 to 100% of the spend [4], spend share is the wrong instrument for locating cheap tokens: hand third parties a fifth of gateway volume for 4% of the money and Anthropic's measured premium falls to about 2.84 [5].
The seat numbers point somewhere else entirely. Roughly $420 against $310 [8] is a 1.35x premium per user [6], and 1.35 divided by the 4.01 price gap leaves an Anthropic seat buying about a third of the tokens an OpenAI seat buys [7]. That is either a small deployment paying dearly or, or rather the more interesting version, a workload that finishes in fewer tokens. The described routing pattern supports the second reading, with the hardest multi-step agentic work going to Claude and high-volume simple work going to whatever is cheaper [10].
The procurement question that settles this is price per finished task, and nobody publishes the denominator. My view, probably wrong in the particulars: this looks closer to a measured trade than a default, because the premium concentrates in coding and agentic applications at 54% against 21% [7], a 2.6x edge [9] in the one category where buyers already instrument their own retries and tool calls. The counter-thesis is respectable. Two thirds of development teams run multiple providers [10], which means the comparison is made per task by engineers rather than per contract by procurement, and a line item defended one route at a time can move in an afternoon without anyone renegotiating anything.
What would break the reading is the measurement spread. A Menlo Ventures survey had Anthropic at 40% of enterprise LLM spend in December 2025 [5] against the gateway's 61% floor [1], 21 points apart on the same question [8], with Ramp's 34.4-44% band straddling both [6]; the climb from 12% in 2023 [4] is not in dispute, the level is. If the stronger OpenAI growth Ramp saw in some Q3 2026 cohorts [11] widens while the gateway share holds, the honest conclusion is that Vercel measures a developer niche and calls it enterprise. That matters more to one seller than the other: Anthropic draws roughly 80% of revenue from enterprise and API [9], while OpenAI has ChatGPT, ChatGPT Enterprise and an Azure channel sitting outside any gateway's field of view [12].
Ranked by verification strength, evidence, and original report placement.
Vercel AI Gateway data shows Anthropic taking a 61-65% share of business AI API spending on just 30-32% of total token volume.
Anthropic is capturing north of 60% of business AI API spending while OpenAI sits at around 35%.
In 2023 Anthropic held a modest 12% of enterprise LLM spend, with OpenAI at around 50%.
A Menlo Ventures survey in December 2025 showed Anthropic at 40% of enterprise LLM expenditures and OpenAI down to 27%.
Ramp sales data from the same period showed Anthropic capturing 34.4-44% of business AI spending versus OpenAI's 32.3-40%.
Some Ramp data showed stronger growth for OpenAI in certain business cohorts during Q3 2026.
Distinct publishers with included, body-backed reporting in this cluster.
cryptobriefing.com
1 article · August 29, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
science
Text watermarks land on 2 December. The detection they imply does not.1 distinct publisher
science
Claude's watermark is a compliance artefact, not a cheating detector1 distinct publisher
build
Lovable's next product is your app's tool surface, served by a hosted MCP server1 distinct publisher
invest
Scalable Capital puts ChatGPT, Claude and Grok inside the European order ticket2 distinct publishers
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One retelling, no primaries
Every figure in this story arrives second-hand through Crypto Briefing, and the three measurement outfits it credits — Vercel's gateway, a Menlo Ventures survey, Ramp's card data — are named but never linked, dated precisely, or described. The numbers doing the heaviest work are the ones with no provider attached at all: 54% of coding, $420 versus $310 per user, an 80% enterprise revenue mix, 67% of teams on multiple models. And the two series that are attributed disagree by 21 points about Anthropic's share.
Real money, borrowed meters
This is not vapour: the shares come from things enterprises actually did — API calls routed, cards charged, a survey answered — and the multi-provider routing pattern is consistent across all of them. But each meter only sees its own pipe. A gateway measures the traffic that chose to pass through it; a spend-management platform sees what its own customers expense. Real behaviour, unknown denominator, and no way from here to tell how much of the market either instrument covers.
The premium is smaller than advertised
4.4x only appears if you pair 65% of spend with 30% of tokens and hand OpenAI every token Anthropic does not have. Run the same published ranges the other way and it is 3.70x; at the midpoints, 4.01x. Give other providers a fifth of gateway tokens and 4% of spend and it lands near 2.84x. Then set the per-seat figures beside it: $420 against $310 is 1.35x, which reads less like buyers paying several times more per token and more like buyers purchasing about a third as many of them. The headline word 'quietly' is doing work the arithmetic does not support.
Vendors holding the tape measure
Each number reaches the reader through an organisation with something to promote around it — a gateway service that bills for routed traffic, a spend platform whose selling point is visibility into exactly this line item, a venture firm whose research is also its brand — and none of the three appears here with methodology attached. The two companies whose revenue is being characterised say nothing; there is no release, no filing, no comment from Anthropic or OpenAI to check the shares against. Crypto Briefing itself discloses no position, and the only correction against its own framing, the Q3 2026 Ramp cohorts, is left to the last line.
Direction plausible, magnitudes not
That Anthropic earns disproportionate enterprise API dollars relative to its token volume is asserted by three independent instruments and fits the routing behaviour described, so the direction is credible. Everything quantitative is soft: a single outlet, unlinked sources, unattributed key figures, a premium multiple that changes materially with assumptions the story never states, and a headline share that no primary publication here confirms.