Vercel's AI Gateway routed enough tokens in June 2026 to show a clear market signal: open-weight models hit 29% of total token volume on under 4% of spend, nearly tripling their share since April. DeepSeek alone accounts for 22.6% of tokens, ranking third behind Anthropic and Google. Z.ai's GLM 5.2 launched June 16 and grew daily token volume 50x by month end, reaching as high as 7th place on single days. The price per token stayed flat, not because the market got cheaper, but because cheap open-weight volume and rising closed-weight frontier prices offset each other exactly.
Anthropic still dominates where it counts. It captured 61% of gateway spend on 32% of tokens, and held 72% or more of spend in coding agents, back-office agents, and app generation. OpenAI's token share fell from 12.5% to 10.3% while its spend share rose from 13.3% to 16.1%, a 50% cost-per-token increase relative to the market in one month. That divergence, less volume but more expensive work, is invisible without production routing data. Claude Fable 5 launched June 9, reached 22% of Opus 4.8 request volume in four days, then went offline after a US export-control directive took effect June 12. It stayed down until July 1.
The modality breakdown is the section worth reading in full. No single lab leads everything. OpenAI's GPT Image took 53% of image generation and 52% of image spend. ByteDance's Seedance led video spend at 49% on only a third of videos generated. xAI's Grok Imagine generated 42% of videos on 19% of spend, mirroring DeepSeek's volume-discount position in text. Chinese labs together took roughly two-thirds of video dollars. The report also breaks out B2B versus B2C routing, back-office agent cost-per-token data, and Google's concentration in consumer workloads at under 2% of coding agent tokens.
[READ ORIGINAL →]