GLM 5.3 FlashX: same model, 2.5x the price for pure speed
Z.ai's speed-tuned GLM costs 2.5x the base for the same benchmarks — plus it hits Vercel's gateway, WebMCP reaches browser agents, and v0 gets private npm.
Copy markdown
200 tokens/second — on a good day
GLM 5.3 FlashX is the speed-tuned variant of GLM 5.3 Flash: 320B total / 18B active params, 1M context, native multimodal. Z.ai claims up to 200 tok/s versus the ~98 tok/s independently measured on the base model — but that peak assumes short outputs, warm caches, and an uncongested cluster.
You pay 2.5x for zero new capability
FlashX runs $0.37 in / $1.25 out per million tokens against the base model's $0.15 / $0.50 — a flat 2.5x, and the multiplier bites hardest on output. Same architecture, same benchmarks (about rank 42 on Artificial Analysis's Intelligence Index). You're buying latency, not intelligence.
When FlashX actually earns its premium
Reach for it on latency-sensitive paths where a human or downstream service is waiting: chained agents, inline code completion, real-time chat. For batch or throughput-bound jobs, the cheaper base model is the better call.
Already routable on Vercel's AI Gateway
GLM 5.3 FlashX is live on Vercel AI Gateway with automatic fallbacks, spend tracking, and request traces behind a single API key — no separate Z.ai account to wire up before you test it against your own base model calls.
WebMCP: hand MCP tools to browser agents, no OAuth
Vercel's mcp-handler shipped experimental WebMCP support, letting your site expose MCP tools to browser-based agents without standing up a separate OAuth flow. A weekend unlock for making your own app agent-operable from the client.
Elsewhere: v0 can install your private npm packages
v0 now pulls private packages from npm and custom registries via env vars like NPM_TOKEN, so it can build against your internal design system and shared components instead of only public dependencies.