As of November 13, 2026, the page Vercel published on 9 Nov 2026 as the Vercel AI Gateway Production Index for November 2026, credited to Amelia Charles and Harpreet Arora, is a gateway-side observation worth reading line by line. This is a reading of an official traffic-and-spend report, not a model benchmark comparison: there are no scores, no rankings, no recommendations, no claim about which model is stronger, and no buying advice. Every figure below is quoted the way the official page writes it, and every reading stays inside the limits the report sets for itself.
What the report says about model churn
The line that draws the most attention is AI models now have a half-life of about six weeks. The half-life here is the report's own way of describing the pace of adoption, not a statement that a model stops working after six weeks. The words about and half-life both matter: one marks the figure as an approximation, the other frames it as a rate of replacement rather than a fixed expiry date. Folding the sentence into some other number of weeks or months is a conversion the page itself never makes, so it is kept verbatim here. A half-life describes how quickly a cohort is replaced; it says nothing about whether any single model can still be called, and reading it as a product lifetime would turn a traffic observation into a specification the report never offers.
Alongside it, the report gives several other qualified observations. A typical model peaks within a month of launch. Four weeks later, three in four of the teams using it at its peak aren't using it anymore. More than half of October's language-model requests went to models less than 60 days old. And a new version arrives every 61 days at the median. These sentences approach the same churn from four angles: when a model peaks, how many teams stay, how young the models serving requests are, and how often new versions appear.
They are four separate measures, not four phrasings of one number, and they should not be merged into a single average. Read together they describe a fast replacement cycle; read as one figure they would misrepresent what the page actually says. Keeping half-life of about six weeks as one observation, a typical model peaks within a month of launch as a second, and a new version arrives every 61 days at the median as a third is the discipline the report's own structure asks for. The median qualifier in particular belongs to a distribution, not to a guarantee about any single model. Nothing in these lines ranks models or says one is better than another; they are about timing, not capability, and about teams on one gateway, not about the whole market.
Reach across six models
The report's breadth figure covers six models, not four tiers, and it is not a count of requests. Every number is written in the reaching ... of paid teams sense, meaning the share of paid teams that a model reached, not the share of requests it served and certainly not a measure of capability. The official sentence reads Sonnet 5.5 (reaching 26% of paid teams), Haiku 5.5 (22%), Haiku 4.5 (20%), and Opus 5.5 (18%). Jev (28%) and GPT-6 Luna (19%) are the other two. Four of the six are Anthropic models; Jev and GPT-6 Luna are the remaining two, and neither can be dropped from the list.
| Model | Share of paid teams reached | Note |
|---|---|---|
| Sonnet 5.5 | 26% | Anthropic; the 26% reach in the title |
| Haiku 5.5 | 22% | Anthropic |
| Haiku 4.5 | 20% | Anthropic |
| Opus 5.5 | 18% | Anthropic; appears again in the spend section |
| Jev | 28% | One of the six; must not be omitted |
| GPT-6 Luna | 19% | One of the six; must not be omitted |
Reading this table as a verdict on which model is stronger would step outside the report's own terms. The 26%, 22%, 20%, 18%, 28%, and 19% figures describe how widely each model was adopted among paid teams; they say nothing about wins, losses, or quality, and they are not a recommendation to switch. The 26% in the title refers precisely to Sonnet 5.5 reaching 26% of paid teams, which is a reach figure and not the same thing as the 26% of all requests that the report gives elsewhere. Four Anthropic models appearing at once says something about the composition of traffic on this gateway; it is not an assessment of any provider, and the 28% next to Jev or the 19% next to GPT-6 Luna does not make either one better or worse. Separating reach from request volume, and both from capability, is the precondition for reading the spend figures below, because those use yet another denominator.
Spend and traffic are different lenses
The easiest figures to conflate are request share and spend share. The report says In September, about one in six language-model requests went to models less than two weeks old, describing how many September requests went to very new models. It also says Opus 5.5 was the top model by spend in its first full month. It took about one in seven dollars across all teams, describing where spend landed during the model's first full month on the gateway. about one in six and about one in seven dollars are both the official fractional wordings. The first is about the age of the models serving requests, the second about a share of spend, and the two do not share a denominator.
Both phrases carry about, and both are deliberately left as fractions rather than converted into percentages. That is not a stylistic choice alone: a conversion would state a precision the official sentence does not claim, and it would invite readers to compare two numbers that were never on the same scale. Quoting them verbatim keeps the request-side and spend-side observations distinct. The phrase top model by spend is likewise a statement about spend, not a capability ranking or a recommendation; it sits in a paragraph about where money went, not about which model performs best, and it should not be carried into an argument about model quality.
The report also lists 26% of all requests and 0.53% of total spend as two further official figures, one on the request side and one on the spend side. They are stated separately in the source, and they should be read separately rather than joined into a single causal story. The more important qualifier is the note that Spend is estimated at published list prices; actual bills may differ. Spend here is an estimate built from published list prices, not the final invoice each team pays, so treating one in seven dollars as an exact settled amount would misread it. There is no price-comparison table in this piece and no pricing verdict, because list price itself is not what a traffic report measures, and the report says nothing about what any team was actually billed.
Where the observations stop
The numbers only make sense inside their own scope. What the report offers is a single gateway's observation of traffic and spend. The line more than 100 trillion tokens a month gives the scale of the tokens passing through that gateway each month; it is a large sample, but it remains one gateway's record, not an industry-wide market share. The voice share rose to 11%, up from 4% in July, which is a shift inside that same gateway's traffic mix rather than a comparison of voice models across vendors. When new paid teams pick their first model, the report lists Anthropic 42.7%, OpenAI 24.0%, Google 16.5%, and DeepSeek 6.6%; those figures describe an onboarding choice on this platform, not a ranking of model quality and not a share of all developers everywhere.
The report is dated 9 Nov 2026, and its byline and figures belong to that page. This article's time anchor is As of November 13, 2026, and the two dates are kept apart: the first is a publication date on the source, the second is the point at which this reading was made. No later figure is added to the report, and nothing on the page is updated with outside numbers. The report itself limits spend to list-price estimates, limits breadth to reach among paid teams, and states request age, version cadence, and first-choice distribution in separate sentences. Together those limits draw a clear line around the whole document.
What sits inside that line is an observation of which models were being used, how quickly they were replaced, and how spend was distributed at list prices. What sits outside it is a model capability evaluation, a cross-platform industry census, and any advice about which model to pick. Reading the report as a statement about which model is stronger, or as a shortlist for a buying decision, goes beyond what the page says verbatim. The most direct way to check any number in this piece is to go back to the 9 Nov 2026 page and verify it there, keeping the official wording rather than a converted one.
How this differs from other comparisons on this site
Several existing reviews on this site work with different definitions and do not overlap with this one. The free-tier comparison covers what each provider gives away for free, the flagship comparison covers flagship positioning, the pricing comparison covers list price per million tokens, and the lightweight-agent comparison covers lightweight execution models. Those pieces are about free tiers, capability, unit price, or lightweight tiers, and their conclusions answer those questions.
This one is about the traffic and spend definitions inside an official Vercel report, and it is especially worth keeping apart from the free-tier review. One asks what can be used at no cost; the other asks which models reached paid teams and how spend was distributed at list prices. Different questions do not yield interchangeable conclusions, and a figure like a 26% reach among paid teams cannot be carried over into an argument about free access or about price per token. Keeping the angles separate is the whole point of a comparison that refuses to rank anything.
FAQ
Q1: How is this review different from the free-tier, pricing, flagship, and lightweight comparisons on this site? A1: Those cover free tiers, capability, unit price, and lightweight tiers. This piece only reads the traffic and spend figures that the official Vercel report gives verbatim; it does not redo a capability comparison and offers no buying advice.
Q2: Does a half-life of about six weeks mean a model only lasts six weeks? A2: No. The official line is AI models now have a half-life of about six weeks, which is an observation about the pace of adoption, with about marking it as an approximation rather than an expiry date. The same page also gives a typical model peaks within a month of launch and three in four of the teams using it at its peak aren't using it anymore as separate measures; none of them should be converted into some other number of weeks.
Q3: Does the six-model table show request volume or reach among paid teams? A3: It shows reaching ... of paid teams, that is, the share of paid teams each model reached, across six models: Sonnet 5.5 26%, Haiku 5.5 22%, Haiku 4.5 20%, Opus 5.5 18%, Jev 28%, and GPT-6 Luna 19%. It is neither request volume nor a capability score.
Q4: Can about one in six and about one in seven dollars be written as percentages? A4: They should not be. Both are the official fractional wordings from Vercel. The first concerns September requests to models less than two weeks old, the second concerns Opus 5.5's spend share in its first full month, and the denominators differ. Quoting them verbatim avoids a conversion the source does not make.
Q5: Can this report serve as a model capability evaluation or a buying guide? A5: No. It is a single gateway's observation of traffic and spend, not a model capability evaluation and not an industry-wide share. Spend also carries the note Spend is estimated at published list prices; actual bills may differ. Reading it as a verdict on which model is stronger, or as advice on which to choose, goes beyond what the page states verbatim.
Related reading
For two adjacent threads, the site also has an AI gateway comparison on gateway selection and a token cost comparison on how token cost is calculated; both use definitions different from the traffic and spend framing used here. To check anything in this piece, the most direct method is to go back to the 9 Nov 2026 page and verify the figures verbatim, rather than relying on a converted or second-hand number.