GLM-5.3-FlashX: GLM-5.3 at ~200 tok/s, $0.37/$1.25 per million
A faster tier of the MIT-open GLM-5.3-Flash, now on Vercel AI Gateway and OpenRouter. Plus: Copilot Code Review goes GA and Vercel's one-tag WebMCP.
Copy markdown
Same model, ~2x the speed
GLM-5.3-FlashX is the existing GLM-5.3-Flash served faster: a dedicated SGLang engine with encode–prefill–decode disaggregation pushes it to a vendor-quoted ~200 tok/s versus ~98 for base Flash. Pricing is $0.37 in / $1.25 out per million tokens ($0.075 cached) — about 2.5x base — with the same 1M-token context. Nothing about the weights changed; you're paying for latency.
The brain it's running
The underlying GLM-5.3-Flash (320B total, 18B active) is MIT-licensed open weights on Hugging Face since Aug 26. Zhipu reports 42 on the Artificial Analysis Intelligence Index and 63.4 on DeepSWE v1.1, with its own code bench near Claude Opus 4.8. Translation: a cheap, fast coding-agent backend you can rent by the token or self-host.
Where to call it today
The FlashX tier is live on Vercel's AI Gateway (~200 tok/s, with built-in spend tracking and request tracing) and on OpenRouter; the model id is glm-5.3-flashx. Swap it into an existing coding or agent loop with one string change, or pull the base weights from Hugging Face if you'd rather run it yourself.
Copilot Code Review hits GA
GitHub shipped Copilot Code Review to general availability: clearer review progress, automatic resolution of comments once they're addressed, and smart batch-suggestion commit messages. If Copilot already reviews your PRs, this is now the default flow rather than a preview toggle.
Vercel's one-tag WebMCP
Vercel's mcp-handler added experimental WebMCP support: expose your MCP tools to in-browser agents with a single script tag. That's a weekend build — make your web app's own actions directly callable by an agent running in the visitor's browser, no separate server round-trip.