GLM 5.3 Flash API Pricing: Cost Per Million Tokens Explained
GLM 5.3 Flash costs 15 cents per million input tokens, 50 cents output, and 3 cents cached input, undercutting Opus by a wide margin.
What does GLM 5.3 Flash cost per million tokens?
GLM 5.3 Flash, ZAI’s newly revealed model (formerly the mystery “Ox Alpha” stealth release), is priced at 15 cents per million input tokens and 50 cents per million output tokens, with cached input priced at just 3 cents per million tokens. That’s roughly the cost of a rounding error compared to frontier closed models like Claude Opus, yet ZAI’s own benchmarks put GLM 5.3 Flash within a few points of Opus 4.8 on agentic and coding tasks. The pricing is not a discount gimmick either. It’s tied directly to how the model is built and where it runs.
TL;DR
- GLM 5.3 Flash’s API pricing is 15 cents per million input tokens, 50 cents per million output tokens, and 3 cents per million cached input tokens, making it one of the cheapest capable models on the market.
- The model is a mixture-of-experts architecture with 320 billion total parameters but only 18 billion active per token, which is why inference stays cheap despite the model’s large footprint.
- On independent testing (KingBench), GLM 5.3 Flash scored 78.75%, landing just below Claude Opus 4.8 and above Opus 5, Kimi K3, and DeepSeek V4 Pro, all while costing a fraction of their API rates.
- ZAI released the full weights under an MIT license on Hugging Face, meaning anyone can self-host the model with no licensing fees at all, effectively driving the marginal cost to whatever compute you already own.
- Part of the low pricing comes from serving infrastructure: ZAI confirmed the entire stealth preview, all 44 trillion tokens of it, ran on Chinese AI chips using a custom SGLang-based inference engine that reportedly delivered a roughly 3x improvement in end-to-end serving performance.
- The pricing puts GLM 5.3 Flash in direct competition with both API-based frontier models and the growing category of high-memory local hardware, since its low active-parameter count makes it fast even on consumer-grade unified-memory machines.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How does GLM 5.3 Flash’s pricing compare to Claude Opus and other frontier models?
The gap is large. Frontier closed models from Anthropic and OpenAI typically charge several dollars per million tokens for both input and output, especially for their top-tier reasoning models. GLM 5.3 Flash’s 15 cents input and 50 cents output pricing is an order of magnitude cheaper, and the cached input rate of 3 cents per million tokens pushes the cost down even further for workloads that reuse context repeatedly, such as long agentic sessions or repeated tool-calling loops.
What makes this notable isn’t just that it’s cheap. It’s that ZAI’s own benchmark claims put GLM 5.3 Flash “approaching Claude Opus 4.8 on coding and agentic benchmarks,” and independent testing backs that up. On one creator’s KingBench evaluation, run through the official API after the model’s public reveal, GLM 5.3 Flash scored 63 out of 80 (78.75%). That places it just under Opus 4.8 (80%) and above Opus 5 (77.5%), Kimi K3 (77.5%), DeepSeek V4 Pro (76.25%), and its own predecessor GLM 5.2 (75%). The larger GLM 5.3 model still topped that same chart at 91.25%, so Flash isn’t the best model in the family, but it’s competitive with Opus-tier models at a small fraction of the price.
Why is GLM 5.3 Flash so cheap to run?
Two things explain it: the model’s architecture and the hardware it runs on.
GLM 5.3 Flash is a mixture-of-experts (MoE) model with 320 billion total parameters, but only 18 billion of those are active for any given token. In MoE architectures, total parameters determine how much memory the model needs, while active parameters determine how much compute is spent per token. That means GLM 5.3 Flash carries the knowledge capacity of a much larger model while running (and costing) closer to an 18-billion-parameter model in practice. This is the core reason ZAI can price it so aggressively without losing money on every request.
The second factor is serving infrastructure. ZAI confirmed that the entire stealth preview period, which served 44 trillion tokens to over 500,000 unique users across more than 13 million sessions, ran on Chinese AI chips using a custom inference engine built on SGLang. ZAI says this setup delivered roughly a 3x improvement in end-to-end serving performance. Running at that scale without relying on Nvidia hardware appears to be a meaningful part of how ZAI can sustain sub-dollar pricing, and it echoes a pattern also seen with DeepSeek, where domestic chip infrastructure correlates with aggressive API pricing.
Is GLM 5.3 Flash worth it for developers on cost alone?
For most token-heavy workloads, the math is hard to argue with. At 15 cents per million input tokens, a workload that would cost tens of dollars on a frontier closed model could cost under a dollar on GLM 5.3 Flash, and the model still lands competitively on agentic and coding benchmarks like Terminal-Bench 2.1 and DeepSWE. ZAI’s benchmark chart shows GLM 5.3 Flash scoring 84.3 on Terminal-Bench 2.1, close to Opus 4.8’s 85, and 63.4 on DeepSWE, up sharply from GLM 5.2’s 46.2.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
There’s a caveat worth noting. Independent retesting after the official API launch showed some regression compared to the earlier stealth-period scores, particularly on one-shot visual generation tasks (things like SVG generation or 3D web builds). Agentic and reasoning tasks, including a hard math problem and a full local fine-tuning pipeline, stayed consistent at top scores. Whether the drop is normal run-to-run variance, a difference between the preview and final serving setup, or a different checkpoint isn’t confirmed, but it’s a reason to benchmark your own specific use case rather than relying purely on published numbers.
What does the MIT license mean for the total cost of running GLM 5.3 Flash?
The API pricing is only half the cost story. ZAI released the full model weights on Hugging Face under an MIT license, which is a permissive open-source license with no usage restrictions. That means the true “cost” of GLM 5.3 Flash for a developer isn’t fixed at 15 cents per million tokens at all. If you have the hardware, you can run it yourself for the cost of electricity and compute, with no per-token fee to ZAI whatsoever.
This matters because of the model’s architecture. Since only 18 billion parameters are active per token, the model can run at usable speeds on high-memory consumer and workstation hardware, even though the full 320-billion-parameter model needs to be loaded into memory. At 4-bit quantization, the model reportedly needs around 180 gigs of memory, with lower-bit quantization projects already underway that could bring that down toward the 100-gig range. That threshold lines up closely with newly released high-memory hardware built explicitly for local AI workloads, turning GLM 5.3 Flash into a genuine “buy hardware once, run forever” option rather than a recurring API cost.
Frequently Asked Questions
How much does GLM 5.3 Flash cost per million tokens?
Input tokens cost 15 cents per million, output tokens cost 50 cents per million, and cached input tokens cost 3 cents per million through ZAI’s official API.
Is GLM 5.3 Flash cheaper than Claude Opus?
Yes, by a wide margin. Frontier closed models like Claude Opus typically charge several dollars per million tokens, while GLM 5.3 Flash’s pricing sits at a small fraction of that, despite scoring close to Opus-tier models on several independent and published benchmarks.
Can I run GLM 5.3 Flash without paying ZAI’s API fees?
Yes. The full weights are released under an MIT license on Hugging Face, so you can self-host the model on your own hardware with no licensing cost, though you’ll need enough memory to load its 320 billion total parameters (around 180 gigs at 4-bit quantization).
Why is GLM 5.3 Flash priced so much lower than other frontier-adjacent models?
Its mixture-of-experts design activates only 18 billion of its 320 billion parameters per token, keeping compute costs low, and ZAI has confirmed it serves the model on Chinese AI chips with a custom SGLang-based inference engine that improved serving efficiency, both of which support the aggressive pricing.
Does cheaper pricing mean lower quality?
Not necessarily. Independent benchmark testing places GLM 5.3 Flash between the two most recent Claude Opus models on a mixed reasoning, coding, and agentic test suite, though scores on some one-shot visual generation tasks were less consistent than the model’s earlier stealth-period results.