JA EN
shiichan

GLM-5.3 Arrives on Workers AI: Coding Benchmarks Jump Up to 6x at the Same Price

Cloudflare Changelog ★★★★★ View Markdown
#cloudflare#ai#serverless

Hi, I'm Shii-chan!

Cloudflare Changelog developers.cloudflare.com

Today's news is about a new model joining Workers AI, and the numbers are honestly kind of wild. Let's dig in!

What was announced?

According to the Cloudflare Changelog, a new model, @cf/zai-org/glm-5.3, is now available on Workers AI. It's Z.ai's flagship agentic coding model, built for long-running, tool-driven development workflows rather than single-turn chat. It has a 1 M token context window and supports reasoning plus function calling.

Why it matters

Here's the interesting part: GLM-5.3 uses the exact same base model as GLM-5.2. Every gain comes purely from post-training. Even so, Z.ai reports a 50% improvement over GLM-5.2 on its in-house Z.ai Code Bench, and calls GLM-5.3 the most capable open-weights model for coding. Getting this much extra performance out of the same foundation shows just how much the post-training work paid off.

What changes

The big deal is that pricing stays identical to GLM-5.2 while performance jumps significantly. SWE-Marathon scores roughly double, and Terminal Bench 3.0 jumps more than 6x. That means the same budget now buys you an agent that can handle longer, more complex coding tasks on Workers AI. GLM-5.3 also takes the top spot in Z.ai's comparisons for CyberGym (vulnerability discovery) and AutomationBench (long-horizon automation), so it looks useful well beyond pure coding tasks too.

Dive Deep

Here's how the public benchmark scores moved from GLM-5.2 to GLM-5.3:

  • Terminal Bench 2.1: 81.0 → 88.2
  • Terminal Bench 3.0: 4.6 → 28.3 (open-source state of the art)
  • DeepSWE: 46.2 → 66.9
  • FrontierSWE: 67.5 → 78.1
  • SWE-Marathon: 19.4 → 42.5

GLM-5.3 also topped Z.ai's comparisons on:

  • CyberGym (vulnerability discovery): 84.5
  • AutomationBench (long-horizon automation): 48.2

Pricing is unchanged from GLM-5.2:

  • Input: $1.40 per M tokens
  • Cached input: $0.26 per M tokens
  • Output: $4.40 per M tokens

You can use it through the Workers AI binding (env.AI.run()), the REST API, the OpenAI-compatible endpoint, or AI Gateway. It requires a Workers Paid plan or prepaid AI Gateway credits.

Wrap-up

  • GLM-5.3, Z.ai's flagship agentic coding model, is now on Workers AI
  • It shares GLM-5.2's base model, with all gains coming from post-training
  • Terminal Bench 3.0 jumped more than 6x, and SWE-Marathon roughly doubled
  • Pricing matches GLM-5.2 ($1.40/M input, $4.40/M output)
  • Requires a Workers Paid plan or prepaid AI Gateway credits

This update is worth checking out if you're building long-running autonomous coding agents on Workers AI or want better performance without a higher bill.

Related posts