Compact MoE sibling of GLM-4.5: 106B total / 12B active. Same hybrid-reasoning and tool-calling stack tuned for high-throughput, low-cost inference. 128K context.
GLM 4.5 Air is a text-generation language model developed by Z.ai. It offers a context window of 128,000 tokens and can generate up to 96,000 tokens in a single response. The model is optimized for…
GLM 4.5 Air specializes in text generation with a strong emphasis on reasoning, particularly mathematical problem-solving as evidenced by its 96.5 score on MATH-500. It can handle complex multi-step instructions, generate coherent long-form text up to 96,000 tokens, and maintain context across 128,000 tokens. Capabilities include answering factual questions, summarizing lengthy documents, translating text between languages, performing logical deductions, and writing code. The model is designed to follow detailed prompts and produce structured outputs. Its large context window allows it to work with entire books, extended reports, or long conversation logs. However, it is a text-only model and cannot process images or other media. For tasks that do not require reasoning or long outputs, a smaller or cheaper model might be sufficient.
The best use cases for GLM 4.5 Air involve tasks that benefit from its large context window and high output limit. Examples include: analyzing and summarizing long academic papers, generating detailed technical documentation, solving complex mathematical problems step-by-step, creating comprehensive study guides, and processing extensive user logs or chat histories. The model also performs well on coding tasks that require understanding long code files or generating large codebases. Because of its pricing structure – $0.20 input and $1.10 output per million tokens – it is cost-effective for scenarios where input is cheaper than output. Applications that need to output many tokens, such as writing long-form content or generating multiple reasoning steps, can be economical compared to models with higher output token costs.
While GLM 4.5 Air offers strong reasoning and a large context, it may be overkill for simpler tasks. Consider using a cheaper, smaller model when the task does not require its full context window or output limit. For example, if you need quick classification, simple translation, or short answer generation, a model with lower token costs would be more economical. Also, if your application does not involve mathematical reasoning or long-form generation, the premium for GLM 4.5 Air’s capabilities may not be justified. The model’s output cost ($1.10 per 1M tokens) is higher than its input cost, so tasks that generate a lot of output (e.g., long summaries from short inputs) could be more expensive than alternative models with lower output costs. Always evaluate the trade-off between capability and cost for your specific use case.
The MATH-500 benchmark evaluates a model’s ability to solve mathematical problems across various difficulty levels, including algebra, geometry, number theory, and more. A score of 96.5 indicates that GLM 4.5 Air correctly answered 96.5% of the problems in the test set. This suggests strong mathematical reasoning capability, comparable to or exceeding other models in its class. It does not guarantee perfect performance on all math problems, especially those outside the distribution of the benchmark. Users should interpret this score as an indicator of the model’s proficiency in symbolic reasoning and step-by-step problem-solving. The benchmark does not measure other important skills like creativity, common sense, or factuality. For non-mathematical tasks, other benchmarks would provide a more relevant comparison.
Specific latency data for GLM 4.5 Air on OrcaRouter is not provided. In general, response speed depends on factors such as the length of input and output tokens, server load, and network conditions. Models with larger context windows and output limits may exhibit longer processing times when generating very long responses. Because GLM 4.5 Air can output up to 96,000 tokens, generating the maximum output will take considerably longer than short responses. OrcaRouter’s API infrastructure is designed to minimize overhead, but actual speed will vary. For applications where low latency is critical, consider using smaller models or shorter output lengths. The model’s performance on MATH-500 suggests efficient reasoning, but real-time applications should be tested under expected load.
Strengths: High mathematical reasoning capability (MATH-500 score 96.5). Large 128K context window allows processing of extensive texts. Maximum output of 96,000 tokens enables generation of full-length documents. Zero-markup pricing on OrcaRouter makes costs transparent. Limitations: Text-only modality; cannot process images, audio, or video. The high output cost ($1.10 per 1M tokens) may be prohibitive for applications that generate very long responses frequently. Benchmark scores for other domains (e.g., general knowledge, code generation) are not provided, so its overall versatility is unknown. Like all language models, it may produce incorrect or biased outputs. It does not have internet access or real-time knowledge by default. Users should validate outputs for critical applications.
Pricing for GLM 4.5 Air is billed at the provider’s rate with zero markup on OrcaRouter. The cost is $0.20 per 1 million input tokens and $1.10 per 1 million output tokens. Input tokens include all text in the prompt (system, user, and assistant messages up to the last response). Output tokens are the generated text. There are no additional fees or platform surcharges. You pay exactly the provider rate. This transparent pricing model allows you to predict costs based on token usage. Billing is typically based on the number of tokens consumed in each API call. Caching policies may apply on OrcaRouter; check the platform documentation for details on whether repeated calls with identical inputs are discounted.
The primary trade-off is between capability and cost. GLM 4.5 Air offers high output limits and strong reasoning, but its output token cost ($1.10 per 1M) is relatively high. For tasks that generate many output tokens from short inputs, the cost can accumulate quickly. Conversely, tasks with large inputs but short outputs benefit from the lower input cost ($0.20 per 1M). The zero-markup pricing on OrcaRouter means you are not paying extra beyond the provider rate, but you still need to manage token usage. If your application primarily requires compact responses, a model with lower output cost might be more economical. For applications requiring long outputs or heavy reasoning, GLM 4.5 Air may be cost-effective despite the higher output cost because of its performance.
OrcaRouter may implement caching policies that reduce the cost for repeated identical input tokens. Specific discount details for GLM 4.5 Air are not provided. Typically, caching discounts apply to prompt tokens that have been processed before, lowering the effective input cost. Users should refer to OrcaRouter’s documentation or support to confirm current caching practices. Since the base input cost is already low at $0.20 per 1M tokens, caching could further reduce costs for applications with repetitive prompts. Output tokens are generally not cached because they vary per call. Always verify the latest billing terms directly with OrcaRouter to understand any available discounts or promotions.
To use GLM 4.5 Air, send HTTP requests to OrcaRouter’s OpenAI-compatible API endpoint: https://api.orcarouter.ai/v1. Include a valid API key in the Authorization header. Specify the model as "z-ai/glm-4.5-air" in the request body. The API supports standard OpenAI chat completion parameters: messages (array of objects with role and content), temperature, max_tokens, top_p, stop, frequency_penalty, presence_penalty, and others. For example, set "max_tokens" to up to 96000 to use the full output capacity. The API returns a JSON response with the generated completion. Streaming is supported by setting "stream": true. Make sure your client library uses the correct base URL and model name. OrcaRouter’s API is compatible with OpenAI’s client SDKs, so migration is straightforward.
GLM 4.5 Air supports a range of parameters through OrcaRouter’s OpenAI-compatible API. Required: model ("z-ai/glm-4.5-air") and messages. Optional parameters include: temperature (0.0 to 2.0, default 1.0) to control randomness; top_p (0.0 to 1.0) for nucleus sampling; max_tokens (up to 96000) to limit output length; stop (list of sequences to halt generation); frequency_penalty and presence_penalty (both -2.0 to 2.0) to penalize token repetition; and stream (boolean) for real-time token delivery. The context window is 128000 tokens, so ensure total tokens in messages plus the generated output do not exceed that limit; otherwise, the request will be truncated or rejected. OrcaRouter may also support additional parameters like logit_bias or user; check documentation. Always refer to the latest API reference for exact details.
Migrating to GLM 4.5 Air on OrcaRouter is simple if you already use an OpenAI-compatible API. Change the base URL to https://api.orcarouter.ai/v1, replace the model name with "z-ai/glm-4.5-air", and use your OrcaRouter API key. No other changes to request structure are required if you use standard parameters. The response format is identical to OpenAI’s chat completions. If you are migrating from a non-OpenAI platform, you will need to adapt your code to use the chat completions format. OrcaRouter also supports function calling and tool use, though not all models do; check if GLM 4.5 Air supports these. Test with small requests first to validate behavior and costs. OrcaRouter provides credit-based billing, so ensure you have sufficient balance before migration.
Within OrcaRouter’s catalog, GLM 4.5 Air stands out for its combination of a large context window (128K), high output limit (96K), and strong mathematical reasoning (MATH-500 96.5). Compared to smaller models, it offers deeper reasoning but at a higher cost per output token. Compared to larger or frontier models, it may lack general knowledge breadth or multimodal capabilities, but it is more cost-effective for text-only, reasoning-heavy tasks. The zero-markup pricing makes it competitive against models with similar capabilities that might include platform fees. For applications that don’t require math or long outputs, cheaper alternatives exist. For tasks needing multi-modal input, other models with image processing would be better. Overall, it occupies a niche as a dedicated reasoning engine with generous token limits.
GLM 4.5 Air is a variant of Z.ai’s GLM-4 family. While specific comparisons are not provided, the "Air" designation typically suggests a more lightweight or cost-optimized version compared to the base GLM-4 model. It likely sacrifices some performance for lower latency or cost, though the MATH-500 score of 96.5 indicates it retains strong reasoning. The context window (128K) and output limit (96K) are generous, possibly larger than earlier GLM-4 iterations. The pricing ($0.20/$1.10 per 1M tokens) is competitive. Without direct benchmark comparisons, users should test both models on their specific tasks. The main differences may be in speed, efficiency, or slightly different training data. OrcaRouter may offer other GLM-4 models with different pricing; compare token costs and performance to choose the best fit.
GLM 4.5 Air is a proprietary model from Z.ai, not open-weight. Compared to open-weight models like those from the Llama or Mistral families, it offers the advantage of being hosted and managed by OrcaRouter with no self-hosting overhead. Its pricing is per-token, whereas open models require compute infrastructure costs. The MATH-500 score is high, but open models may have different strengths (e.g., broader knowledge). The context window (128K) is large but some open models offer similar or larger contexts. The output limit of 96K tokens is unusually high compared to most open models, which typically cap at 4K-32K. For users who need very long generations without managing infrastructure, GLM 4.5 Air is convenient. For those who require customizability or data sovereignty, open-weight models may be preferred.
GLM 4.5 Air is text-only, so it cannot process images, audio, or video. If your application requires understanding visual content (e.g., analyzing charts, reading handwriting, interpreting photos), you would need a multimodal model such as GPT-4V or Claude 3. Similarly, it cannot generate images or speech. For tasks that combine text and image reasoning, a multimodal model would be essential. GLM 4.5 Air’s strength lies purely in textual reasoning and generation. Users should assess whether their use case genuinely needs multimodal input or if text-only is sufficient. If text-only is adequate, GLM 4.5 Air may be more cost-effective for reasoning-heavy tasks than multimodal models, which often charge higher token rates and may include vision capabilities that are unused.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key="$ORCAROUTER_API_KEY",
)
response = client.chat.completions.create(
model="z-ai/glm-4.5-air",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)do_sampleinclude_reasoningmax_tokensreasoningrequest_idresponse_formatstopstreamtemperaturethinkingtool_choicetool_streamtoolstop_puser_id| Input / 1M tokens | $0.200 |
| Output / 1M tokens | $1.10 |
| Cache read / 1M | $0.030 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
GET /api/public/models/z-ai/glm-4.5-airOpen @misc{orcarouter_glm_4_5_air,
title = {GLM 4.5 Air API},
author = {Z.ai},
year = {2025},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/z-ai/glm-4.5-air}
}Z.ai. (2025). GLM 4.5 Air API. OrcaRouter. https://www.orcarouter.ai/models/z-ai/glm-4.5-air