Vercel AI Gateway adds GLM-5.2 Fast at 150-250 tok/s

• Vercel added a serverless zai/glm-5.2-fast endpoint to AI Gateway, and wafer_ai's launch post says Wafer is the serving partner.
• The launch post from wafer_ai priced the endpoint at $3.00 per 1M input tokens, $10.25 per 1M output tokens, plus $0.50 cached input, with a 1M-token context window.
• Throughput is the headline: wafer_ai's benchmark chart claimed 150 to 250 tok/s, while vercel_dev's announcement described it as 2x faster token throughput than other providers in internal benchmarks.
• The rollout also expands AI Gateway's model mix, because jediahkatz's support note said this was the first GLM model the gateway team had supported.

You can jump straight to the AI Gateway model link, inspect wafer_ai's throughput chart, and see rauchg's reply confirm that Wafer is the inference partner behind this specific integration.

The practical update is simple: GLM-5.2 Fast is now another hosted option inside Vercel's existing gateway, under the slug zai/glm-5.2-fast, according to vercel_dev's announcement and wafer_ai's launch post.

When asked who was behind the deployment, rauchg's reply said Vercel works with multiple inference partners and that Wafer is the partner in this case.

The performance claim came with a concrete chart. wafer_ai's benchmark post advertised 150 to 250 tok/s, and the attached screenshot showed Wafer peaking at 245.5 TPS on AI Gateway traffic, ahead of Baseten at 121.0, Fireworks at 80.5, and Z.ai at 53.0.

That is narrower than a broad "fast model" claim, and more useful. The comparison is specifically about provider throughput for GLM traffic routed through AI Gateway, not a model-quality benchmark.

The commercial terms were in the launch tweet itself:

• Input: $3.00 per 1M tokens, per wafer_ai's pricing post
• Output: $10.25 per 1M tokens, per wafer_ai's pricing post
• Cached input: $0.50 per 1M tokens, per wafer_ai's pricing post
• Context window: 1M tokens, per wafer_ai's pricing post

For model context, eliebakouch's CursorBench comment called GLM 5.2 roughly as cost efficient as Opus 4.8 on CursorBench, while natolambert's post argued that performance in that range puts pressure on frontier-model margins.

The small but notable platform detail came after launch: jediahkatz's support note said this was AI Gateway's first GLM integration and asked users to report odd behavior.

Wafer also used the thread to position the underlying model as more than a speed demo. In wafer_ai's DeepSWE reply, the company said GLM 5.2 ranks at or near the top on coding benchmarks and is the number one open-source model on DeepSWE.