What the October figures show

Vercel's AI Gateway moves over 100 trillion tokens a month, and its November Production Index reads that traffic for signals about how enterprises and startups pick models. Four shifts stand out for October 2026: decision models reached a quarter of requests at negligible cost, voice use widened, Anthropic closed the team-share gap with OpenAI while winning new arrivals, and the bulk of language-model traffic now goes to very recent releases.

  • Decision models took roughly one in four gateway requests six weeks after launch, while consuming half a cent of every dollar spent. Nearly a third of paid teams used one, rising to seven in ten among teams above 10 billion tokens a month.
  • Voice adoption climbed to about 11% of paid teams, from 4% in July.
  • OpenAI and Anthropic now serve about the same share of paid teams; new October teams chose Anthropic nearly two to one. OpenAI's share of gateway spend has grown by two-thirds since December.
  • Over half of language-model requests went to models under two months old, up from under 30%. Within a month of a typical model's peak, three in four of its teams have moved elsewhere.

Decision models: 26% of requests, 0.53% of spend

Production workloads are separating generation from judgment. A decision model classifies, scores or selects among options rather than writing text, at roughly one seventieth the per-request cost of a language model. Fifteen such models from eight vendors have been on AI Gateway since mid-September.

In October they carried about a quarter of requests but only 0.53% of total spend — a consequence of low token prices and very short outputs. Nearly all of that volume went to TypeSafe AI's Jev, which drew almost four times Anthropic's total request count; teams spent $64 on Anthropic models for every $1 they spent on Jev. Around 11% of paid teams used a decision model without sending any request to a text-generating model at all.

Adoption scales with usage: just over one in five teams below 10 million tokens used one, nearly half in the 100 million to 1 billion band, and seven in ten above 10 billion. Bigger teams naturally run more model types — the top 1% by spend use a median of 29 models, against one for the bottom tenth — and the more requests a team sends, the more it saves by offloading simple calls to a cheaper model.

Voice grows 2.6x since July

Roughly one in nine paid teams used a voice model in October, against one in 24 in July — a 161% increase. The growth comes from both directions: teams that paid in both months went from 5% to 14% using voice, and the share sending at least 100 voice requests a month rose fivefold. Retention improved too, with six in ten of September's voice teams returning in October, versus fewer than half in July.

A spoken interaction involves up to three components: speech-to-text, a language model for the reply, and text-to-speech. Speech-to-text usage has nearly tripled since July and text-to-speech more than tripled. How teams combine them flipped between July and October: about a quarter used a single all-in-one model and one in six assembled the pieces themselves in July, and by October those proportions had reversed. For most teams voice is an added interface on an existing app or agent, with their preferred language model remaining in the middle.

Anthropic ahead on new teams; OpenAI gains on spend share

Per 100 paid teams, OpenAI went from 70 in December to 55 in October and Anthropic from 48 to 54. Among teams present since December, Anthropic's reach rose from about half to two-thirds and OpenAI's from roughly three in four to four in five, but Anthropic's advantage among incoming teams was decisive: it was the first-choice lab for nearly twice as many October newcomers.

With the largest teams the labs compete on volume rather than presence — more than nine in ten of the top 1% by spend use both. OpenAI now captures close to 15 cents of each gateway dollar, up from about 9 in December, while Anthropic's share halved to about 32 cents.

A six-week half-life for models

More than half of October's language-model requests went to models less than 60 days old, a level that had never exceeded three in ten before September. Within a family, a new version ships every 61 days at the median, and a typical model peaks within a month of launch; four weeks after that peak, three in four of its teams have stopped using it, with or without a successor. Most remain on the gateway, running something else.

Among teams processing at least 10 million tokens monthly, about half changed their most-used model in October. Four in five of those switches went to a newer release and six in ten to a different lab. Older releases persist as secondary models: seven in ten of these teams still called a model at least ten months old, though it handled under 3% of the typical team's tokens. The takeaway for labs is that each release has to re-earn the business, since the effective half-life of a model is about six weeks.

Model and modality notes

  • Opus 5.5 led spend in its first full month, taking about one in seven dollars. Four of the six most-used models are Anthropic's: Sonnet 5.5 (26% of paid teams), Haiku 5.5 (22%), Haiku 4.5 (20%) and Opus 5.5 (18%), alongside Jev (28%) and GPT-6 Luna (19%).
  • Image and video set records, with the largest teams adopting fastest: a third of the top 1% by spend now use an image model (a quarter in July) and one in six a video model (one in ten).
  • GPT-Image-2.5 Sunburst topped image spend in its first full month, ahead of Flare and GPT-Image-2. Seedance 2.5 from ByteDance led video spend for a third straight month, followed by Google's Veo 3.1 and Kuaishou's Kling 3.0.

Methodology

Figures are anonymized, aggregate gateway traffic through October 2026. Spend is estimated at published list prices. A paid team sent at least one paid request during the month, and team size is total monthly tokens. Decision models follow the gateway's own classification. Model versions are counted separately and aged from first appearance on the gateway, whose history starts October 20, 2025. Switching rates cover paid teams with at least 10 million tokens in both September and October, with "most-used" measured by tokens. Voice data begins June 22, 2026; "assembled the pieces" means speech-to-text, a language model and text-to-speech all used on the same day. A new team's first-choice lab is the one receiving the most language-model requests in its first month. Prior months may be revised as methodology changes.