GLM-5.3-Flash Slashes Agentic Inference Costs by up to 50x
Zhipu AI revealed that its unannounced Ox Alpha model on OpenRouter is GLM-5.3-Flash. It delivers frontier-grade coding performance at $0.15 per million input tokens and $0.50 per million output tokens under an MIT license.
Impact: High
Why it matters
You can replace expensive frontier models in long-running agent loops to cut repository-scale token expenses by over 90%.
TL;DR
- 01GLM-5.3-Flash is available under an MIT license with a 1M token context window and 18B active parameters.
- 02Pricing is $0.15/1M input and $0.50/1M output, up to 50x cheaper than premier frontier models.
- 03Benchmark scores on Terminal-Bench 2.0 and DeepSWE rival leading proprietary models.
Key facts
- Total / Active Parameters
- 320B total / 18B active
- Context Window
- 1,000,000 tokens
- Input / Output Pricing
- $0.15 / $0.50 per 1M tokens
- License
- MIT
- Terminal-Bench 2.0 Margin
- -0.7 vs Claude Opus 4.8
Open Architecture and High-Volume Economics
Zhipu AI's GLM-5.3-Flash uses a Mixture-of-Experts architecture with 320B total parameters and 18B active parameters per token. Released under an MIT license, it supports a 1M-token context window along with text, image, and video modalities. During an initial anonymous deployment under the alias Ox Alpha on OpenRouter, the model processed approximately 23 trillion tokens in less than six days.
Pricing Disruption for Agentic Loops
Standard list pricing places GLM-5.3-Flash at $0.15 per million input tokens and $0.50 per million output tokens. Compared to GPT-5.6 Terra, this represents roughly a 13× reduction in input costs and a 24× reduction in output costs. Against Claude Opus 5, GLM-5.3-Flash is approximately 33× cheaper on input and 50× cheaper on output. For automated coding workflows where agents run dozens of file-read, test, and repair cycles, these unit economics significantly reduce operating overhead.
Benchmark Parity
Evaluation results show tight margins against frontier systems:
- Terminal-Bench 2.0: Trailed Claude Opus 4.8 by 0.7 points and GPT-5.6 Terra by 1.5 points.
- DeepSWE: Scored closely behind top frontier models on software engineering tasks.
- AutomationBench & GDPVal-AA v2: Outperformed Claude Opus 4.8 and GPT-5.6 Terra.
Try it in 2 minutes
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "zhipu/glm-5.3-flash",
"messages": [{"role": "user", "content": "Refactor this agent loop."}]
}'bash
✓ When to use
- High-volume coding agents that make hundreds of API calls for testing, linting, and multi-file edits.
- Long-context document ingestion and multimodal processing pipelines requiring an MIT open-weight license.
✕ When NOT to use
- When attempting zero-setup local deployment on standard consumer laptops without sufficient memory.
- When tasks strictly depend on specialized proprietary ecosystem integrations unique to closed frontier APIs.
What to do today
- Route high-iteration background agent tasks to GLM-5.3-Flash on OpenRouter or dedicated cloud endpoints.
- Benchmark your existing agent test suite against GLM-5.3-Flash to measure task success rates versus frontier models.
Sources