Cloudflare Workers AI GLM-5.3: The $1.40/M Coding Model That Just Hit Terminal Bench 3.0 SOTA

On August 28, 2026, Cloudflare quietly added Z.ai's GLM-5.3 to Workers AI under the model ID @cf/zai-org/glm-5.3. The interesting bit isn't that it's another Chinese frontier model on Cloudflare's edge — that story already played out with GLM-5.2 in June. The interesting bit is the price-to-performance ratio: GLM-5.3 costs the same as GLM-5.2 ($1.40 per million input tokens, $0.26 cached, $4.40 output) while roughly doubling GLM-5.2's score on long-horizon benchmarks like SWE-Marathon and improving Terminal Bench 3.0 by 6.1x (4.6 → 28.3, now open-source state of the art).

Cloudflare calls GLM-5.3 "the most capable open-weights model for coding" in the changelog body. The arXiv paper (2602.15763, "GLM-5: from Vibe Coding to Agentic Engineering") frames the model as the post-training-only successor to GLM-5.2 — same base weights, every gain from alignment work. That matters for capacity planning: if you already deployed GLM-5.2 on Workers AI, the migration is a one-line model-ID swap, and your unit economics stay flat.

This review covers: the exact Workers AI pricing and Neuron consumption (verified from /workers-ai/platform/pricing/), the paid-billing requirement (GLM-5.3 is on the frontier-model list — no free 10K Neurons/day), the current rate-limit picture (300 RPM default, frontier-RPM table not yet updated), the four API surfaces (binding, REST, OpenAI-compatible, AI Gateway), how GLM-5.3 stacks up against GLM-5.2 on the seven public benchmarks Z.ai reports, and the open-weights + self-hosting story under the new GLM-5.3 License.

Last verified: 2026-08-29. Pricing and benchmark figures pulled directly from the Cloudflare changelog and pricing page; if Cloudflare updates the frontier-rate-limit table to include GLM-5.3 in the next few days, I'll note it in the FAQ.

GLM-5.3 on Workers AI: pricing, Neurons, and the paid-tier requirement

Cloudflare's pricing page lists GLM-5.3 verbatim alongside GLM-5.2 and the rest of the frontier list. The exact row, lifted from /workers-ai/platform/pricing/:

@cf/zai-org/glm-5.3
  Input tokens          $1.400 per M    127,273 neurons per M
  Cached input tokens   $0.260 per M     23,636 neurons per M
  Output tokens         $4.400 per M    400,000 neurons per M

Three things to internalize:

• Same price as GLM-5.2. Z.ai and Cloudflare deliberately held the rates flat to make the upgrade a no-brainer for existing GLM-5.2 workloads. The benchmark improvements are pure upside on the same line item.
• Cache hits are roughly one-fifth the input price. $0.26/M cached vs $1.40/M full input — a 5.4x discount on repeated prefix. If you run an agent loop that re-feeds the system prompt + tool definitions every turn (typical Claude Code / Cursor pattern), this matters enormously. A 30K-token fixed prompt on every call costs $0.009 cached vs $0.042 full input — saves ~$0.033 per turn.
• Neuron consumption is 127,273 / 400,000 per 1K input/output. That's roughly twice what Qwen 3.8 27B (also on Workers AI) consumes per token. Heavier bills for the same token count, but GLM-5.3 needs fewer total tokens because the agentic tasks finish in fewer rounds.

The paid-tier requirement is the structural change for hobbyists. From the pricing page:

Some models require a paid billing method. This applies to @cf/moonshotai/kimi-k2.6, @cf/moonshotai/kimi-k2.7-code, @cf/zai-org/glm-5.2, @cf/zai-org/glm-5.3, @cf/zai-org/glm-5.3-flash, @cf/deepseek-ai/deepseek-v4-flash-0731, and Two paths: subscribe to Workers Paid ($5/month, includes 10M included Neurons + per-Neuron overage), or use prepaid AI Gateway credits without a Workers subscription. The AI Gateway path is interesting for teams that already use Cloudflare's gateway for OpenAI/Anthropic routing — same billing relationship, just add a model ID. The changelog notes that "Requests to frontier models that use prepaid credits receive higher rate limits," which is the explicit incentive.

Rate limits: 300 RPM default, frontier table pending

Cloudflare's rate-limits page lists GLM-5.3 under the paid-billing frontier umbrella but has not yet added it to the explicit "Frontier models" RPM table. That table currently lists only three models:

Model	Standard RPM	Prepaid AI Gateway RPM
@cf/zai-org/glm-5.2	20	50
@cf/moonshotai/kimi-k2.6	20	50
@cf/moonshotai/kimi-k2.7-code	20	50
@cf/zai-org/glm-5.3 (new, expected)	300 (Text Generation default)	tbd

In practice, GLM-5.3 currently inherits the 300 RPM Text Generation default — much higher than GLM-5.2's 20 RPM standard frontier ceiling. This is almost certainly a transitional state (Cloudflare usually caps frontier models at 20/50 RPM to protect GPU supply), so if you're building production traffic on GLM-5.3, plan for the table to be updated within a week or two. The 50 RPM prepaid-credits uplift pathway will likely apply when the table is updated.

For agent workloads with bursty call patterns, the 300 RPM default is a temporary gift — concurrent agent fleets can hammer GLM-5.3 hard today. For sustained batch jobs, queue with a token-bucket limiter.

Benchmarks: GLM-5.3 vs GLM-5.2, head-to-head

The headline claim from the changelog is the 50% Z.ai Code Bench improvement. The public benchmark deltas are more striking:

Benchmark	GLM-5.2	GLM-5.3	Delta
Terminal Bench 2.1	81.0	88.2	+7.2
Terminal Bench 3.0	4.6	28.3	+23.7 (OS SOTA)
DeepSWE	46.2	66.9	+20.7
FrontierSWE	67.5	78.1	+10.6
SWE-Marathon	19.4	42.5	+23.1 (2.2x)
CyberGym (vuln discovery)	—	84.5	top in Z.ai comparisons
AutomationBench	—	48.2	top in Z.ai comparisons

Three patterns to call out:

• Long-horizon work sees the biggest lift. SWE-Marathon (multi-hour agentic engineering tasks) more than doubled. Terminal Bench 3.0 (longer-horizon terminal tasks) jumped 6.1x. The post-training emphasis on agentic tool use shows up most where the model has to plan across many steps.
• Single-turn coding nudges up only modestly. Terminal Bench 2.1 went from 81.0 to 88.2 — a solid +9% gain, but not a revolution. If your workload is one-shot code generation (no tool loop, no multi-step planning), the upgrade is nice-to-have, not transformative.
• Vulnerability discovery and long-horizon automation are new SOTA claims for open weights — CyberGym 84.5 and AutomationBench 48.2 are benchmark wins for security-research and agent-automation use cases where open-weight models previously trailed closed APIs.

The "same base model as GLM-5.2, every gain from post-training" framing matters for one decision: if your GLM-5.2 deployment was bottlenecked on context understanding (rather than planning, tool use, or long-horizon coherence), GLM-5.3 won't help. If your deployment was bottlenecked on the agent getting stuck mid-task or losing track of the goal, GLM-5.3 should be a measurable improvement.

Calling GLM-5.3 from a Worker: four entry points

Cloudflare documents four API surfaces for Workers AI models, all supported on GLM-5.3.

1. Workers binding — env.AI.run()

// wrangler.toml bindings: [[ai_bindings]] binding = "AI"
// TypeScript Worker
export interface Env { AI: Ai }

export default {
  async fetch(req: Request, env: Env): Promise<Response> {
    const { prompt } = await req.json();

    const resp = await env.AI.run('@cf/zai-org/glm-5.3', {
      messages: [
        { role: 'system', content: 'You are a senior TypeScript engineer.' },
        { role: 'user',   content: prompt }
      ],
      max_tokens: 4096,
      temperature: 0.7,
      reasoning_effort: 'medium'  // 'low' | 'medium' | 'high'
    });

    return Response.json(resp);
  }
};

The reasoning_effort parameter (low / medium / high) is OpenAI-style — set high for the agentic benchmarks, low for simple Q&A.

2. REST API — account-scoped endpoint

curl -X POST "https://api.cloudflare.com/client/v4/accounts/$CF_ACCOUNT_ID/ai/run/@cf/zai-org/glm-5.3" \
  -H "Authorization: Bearer $CF_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "user", "content": "Refactor this Python function to use asyncio."}
    ],
    "max_tokens": 2048,
    "stream": true
  }'

Useful from CI runners, serverless functions outside the Workers runtime, or any environment where you can hold an API token.

3. OpenAI-compatible — drop-in for OpenAI SDKs

from openai import OpenAI

client = OpenAI(
    api_key="<your-cloudflare-api-token>",
    base_url="https://api.cloudflare.com/client/v4/accounts/<CF_ACCOUNT_ID>/ai/v1"
)

resp = client.chat.completions.create(
    model="@cf/zai-org/glm-5.3",
    messages=[
        {"role": "user", "content": "Write a TypeScript debounce function with tests."}
    ],
    max_tokens=2048,
    temperature=0.7,
)

print(resp.choices[0].message.content)

Works with any OpenAI SDK (Python, Node, Go, Rust) by swapping base_url. The model ID has to include the @cf/ prefix even on the OpenAI-compatible surface — that's the one gotcha.

4. AI Gateway — gateway-mediated routing + caching

// gateway.example.com routes to GLM-5.3 via Unified billing
curl -X POST "https://gateway.example.com/v1/chat/completions" \
  -H "Authorization: Bearer $GATEWAY_TOKEN" \
  -d '{
    "model": "@cf/zai-org/glm-5.3",
    "messages": [{"role": "user", "content": "Explain DSA attention."}]
  }'

AI Gateway is the prepaid-credits path mentioned earlier — set gateway billing to "Unified billing" and load credits, then route OpenAI/Anthropic/Workers AI models through one endpoint with caching, rate-limiting, and analytics on top.

Open weights: HF repo, license, self-hosting path

GLM-5.3 weights are published at huggingface.co/zai-org/GLM-5.3 under the custom GLM-5.3 License. The model is a MoE with DSA attention (GlmMoeDsaForCausalLM), 8 experts per token, ~754B total parameters stored as BF16 + FP8 sharded safetensors (~755 GB on disk across all shards).

License summary (full text at HF LICENSE file):

• Free for use, modification, merge, publication, distribution, sublicensing, and sale
• Fine-tuning and derivative works allowed
• No copyleft — derivative works can use any license
• One condition: Model-as-a-Service providers whose affiliate-group revenue exceeds US$10 billion in any 12-month period must complete a Z.AI security review before commercial use
• Provided "AS IS", no warranty
• Contact for the security review: glmlicense@z.ai

For typical SaaS startups and internal tools, the license is permissive in practice. The $10B-revenue threshold targets only the largest hyperscalers — every API reseller below that line can self-host GLM-5.3 without contacting Z.AI. The arXiv paper is 2602.15763 ("GLM-5: from Vibe Coding to Agentic Engineering") for the architecture details.

If you self-host, expect serious hardware requirements: 754B parameters at BF16 is ~1.5 TB GPU memory minimum for inference, or ~755 GB at FP8. An 8x H100 node (80 GB each, 640 GB total) won't fit BF16; you'd need an 8x B200 or 8x GB300 node for full BF16 inference, or accept FP8 quantization for an 8x H100 fit. For most teams, Workers AI's $1.40/M input is cheaper than the amortized GPU cost of self-hosting at this scale.

When to use GLM-5.3 vs the alternatives on Workers AI

The Workers AI China-model lineup now has three paid-tier options for coding/agent workloads:

Model	Input $/M	Output $/M	Best for
@cf/zai-org/glm-5.3	$1.40	$4.40	Long-horizon agentic engineering, multi-hour coding tasks, vulnerability research
@cf/zai-org/glm-5.3-flash	$0.15	$0.50	High-volume interactive workloads, multimodal input, cost-sensitive chat/agent loops
@cf/moonshotai/kimi-k2.7-code	$0.60	$2.50	Code-completion, code-review, code-search — cheaper than GLM-5.3 for sub-32K single-turn code tasks
@cf/qwen/qwen3.8-27b	$0.20	$0.80	Cheapest multimodal on Workers AI — for volume, not frontier benchmarks

The price pattern: GLM-5.3 is the most expensive of the four (10x the input cost of GLM-5.3-Flash, 7x Qwen 3.8 27B), justified by the benchmark deltas. If your workload is single-turn Q&A or simple chat, GLM-5.3-Flash or Qwen 3.8 27B will be a better cost/performance choice. If your workload is the agentic-coding / long-horizon-automation bucket that GLM-5.3's benchmarks target, the premium is worth it.

For the alternative path through Z.ai's own platform, the GLM Coding Plan subscription bundles GLM-5.3, GLM-5.3-Flash, GLM-5.2, and GLM-5-Turbo for IDE and agent use cases at a flat monthly rate. That's the right answer for IDE plugin developers who want frontier quality on a fixed budget; for API-based workloads, Workers AI's pay-per-token model beats any subscription once you're above ~5M tokens/month.

FAQ

Q: Is GLM-5.3 free on the Workers Free tier?

No. GLM-5.3 is on Cloudflare's paid-billing-only frontier list alongside GLM-5.2, GLM-5.3-Flash, Kimi K2.6, Kimi K2.7 Code, DeepSeek V4 Flash, and DeepSeek V4 Pro. To call it you need a Workers Paid plan ($5/month) or prepaid AI Gateway credits. The standard 10,000 free Neurons/day does not apply.

Q: How much does GLM-5.3 cost on Workers AI?

$1.40 per million input tokens, $0.26 per million cached input tokens, and $4.40 per million output tokens — same rates as GLM-5.2. Per 1,000 input tokens the model consumes 127,273 Neurons; per 1,000 output tokens, 400,000 Neurons. A typical 50K-input / 4K-output coding task runs around $0.088 per call.

Q: What is the model ID on Workers AI?

The official model ID is @cf/zai-org/glm-5.3 (note: namespace is zai-org, NOT zhipu). The sibling Flash variant is @cf/zai-org/glm-5.3-flash. Both share the same 1,048,576-token context window, reasoning, function calling, and structured outputs capabilities.

Q: What is the rate limit for GLM-5.3?

As of 2026-08-29, GLM-5.3 inherits the default Text Generation rate of 300 requests per minute per account. Cloudflare's explicit 'Frontier models' RPM table (currently listing GLM-5.2, Kimi K2.6, Kimi K2.7 Code at 20 RPM standard / 50 RPM with prepaid AI Gateway credits) has not been updated to include GLM-5.3 yet — expect that to be added in the days after launch.

Q: How does GLM-5.3 compare to GLM-5.2 on coding benchmarks?

Z.ai reports a 50% improvement on its in-house Z.ai Code Bench. On public benchmarks GLM-5.3 vs GLM-5.2 scores: Terminal Bench 2.1 88.2 vs 81.0, Terminal Bench 3.0 28.3 vs 4.6 (now open-source SOTA), DeepSWE 66.9 vs 46.2, FrontierSWE 78.1 vs 67.5, SWE-Marathon 42.5 vs 19.4 (more than 2x), CyberGym 84.5, AutomationBench 48.2. The same base model as GLM-5.2 — every gain comes from post-training.

Q: Can I self-host GLM-5.3?

Yes. GLM-5.3 weights are open and available at huggingface.co/zai-org/GLM-5.3 (754B parameters, BF16 + FP8 sharded, ~755 GB on disk). The model is a MoE with DSA attention (GlmMoeDsaForCausalLM, 8 experts per token). License is the custom GLM-5.3 License — free for use, modification, fine-tuning, and commercial deployment; Model-as-a-Service providers whose affiliate-group revenue exceeds US$10B in any 12-month period must complete a Z.AI security review first.

Q: How do I call GLM-5.3 from a Worker?

Four entry points: (1) Workers binding — env.AI.run('@cf/zai-org/glm-5.3', { messages: [...] }); (2) REST API — POST https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/@cf/zai-org/glm-5.3; (3) OpenAI-compatible — POST /v1/chat/completions against the Workers AI endpoint with model='@cf/zai-org/glm-5.3'; (4) AI Gateway — same model ID, with gateway billing set to 'Unified billing' for prepaid-credit access.

Q: What about China access?

Workers AI endpoints are global; the model runs in Cloudflare's GPU pools across 300+ cities and is reachable from mainland China through Cloudflare's standard anycast IPs. For higher throughput inside the GFW, Z.ai's own platform (z.ai) hosts the same GLM-5.3 family plus a GLM Coding Plan subscription that bundles GLM-5.3, GLM-5.3-Flash, GLM-5.2, and GLM-5-Turbo for IDE and agent use cases.

Bottom line

GLM-5.3 on Workers AI is the most capable open-weights coding model you can call today at $1.40/M input. The benchmark deltas against GLM-5.2 — Terminal Bench 3.0 6.1x, SWE-Marathon 2.2x, CyberGym 84.5 — are real and they came from post-training only, so the upgrade is a one-line model-ID swap for any existing GLM-5.2 deployment.

The two real gotchas are paid-billing requirement (no free 10K Neurons/day) and the rate-limit picture still in flux (300 RPM default today, almost certainly tightening once Cloudflare updates the frontier-RPM table). For teams already on Workers Paid or with AI Gateway credits loaded, this is a no-brainer upgrade. For teams using the free tier, the path is either upgrade to Paid ($5/month minimum) or stick with GLM-5.3-Flash / Qwen 3.8 27B which remain on the standard Text Generation 300 RPM with free-tier Neurons.

Last verified: 2026-08-29. Sources: Cloudflare changelog, Workers AI pricing, Workers AI limits, GLM-5.3 model page, HF model card, arXiv 2602.15763.