Live pricing on OneInfer
Pulled from the OneInfer catalog at request time — rates can change. Capture the timestamp with any benchmark or production decision.
| Model | Provider | Input $/1M tokens | Output $/1M tokens |
|---|---|---|---|
| GLM-5.3 | zai | $140.00 | $440.00 |
| GLM-5.3 | zai | $15.00 | $50.00 |
| GLM-5.3 | zai | $37.00 | $125.00 |
| GLM-5.3 | bytedance | $15.00 | $50.00 |
| GLM-5.3 | akashml | $19.00 | $440.00 |
| GLM-5.3 | novita | $140.00 | $440.00 |
| GLM-5.3 | novita | $15.00 | $50.00 |
| GLM-5.3 | together_ai | $15.00 | $50.00 |
| GLM-5.3 | deepinfra | $15.00 | $50.00 |
| Qwen | deepinfra | $38.00 | $40.00 |
| Qwen | deepinfra | $9.00 | $58.00 |
| Qwen | deepinfra | $38.00 | $155.00 |
| Qwen | deepinfra | $15.00 | $150.00 |
| Qwen | deepinfra | $22.00 | $96.80 |
| Qwen | deepinfra | $16.50 | $66.00 |
| Qwen | deepinfra | $1.10 | — |
| Qwen | deepinfra | $2.20 | — |
| Qwen | deepinfra | $1.10 | — |
| Qwen | deepinfra | $1.10 | — |
| Qwen | deepinfra | $2.75 | — |
| Qwen | deepinfra | $5.50 | — |
| Qwen | cerebras | $99.00 | $149.00 |
| Qwen | deepinfra | $13.20 | $26.40 |
| Qwen | deepinfra | $13.20 | $55.00 |
| Qwen | deepinfra | $8.80 | $30.80 |
| Qwen | novita | $40.00 | $320.00 |
| Qwen | novita | $30.00 | $240.00 |
| Qwen | novita | $25.00 | $200.00 |
| Qwen | novita | $60.00 | $360.00 |
| Qwen | thinkingmachines | $165.00 | $750.00 |
| Qwen | deepinfra | $11.00 | $16.50 |
| Qwen | novita | $60.00 | $360.00 |
| Qwen | thinkingmachines | $186.00 | $559.50 |
| Qwen | groq | $60.00 | $300.00 |
| Qwen | novita | $24.80 | $148.50 |
| Qwen | thinkingmachines | $45.00 | $133.50 |
| Qwen | akashml | $10.00 | $90.00 |
| Qwen | novita | $125.00 | $375.00 |
| Qwen | novita | $200.00 | $600.00 |
| Qwen | together_ai | $200.00 | $600.00 |
| Qwen | akashml | $22.50 | $198.00 |
| Qwen | novita | $15.00 | $47.00 |
| Qwen | deepinfra | $15.00 | $47.00 |
| Qwen | novita | $200.00 | $600.00 |
| Qwen | deepinfra | $200.00 | $600.00 |
| Qwen | oneinfer | $15.00 | $47.00 |
| Qwen | oneinfer | $15.00 | $47.00 |
| Qwen | novita | $NaN | $NaN |
| Qwen | fal | $NaN | $NaN |
| Qwen | fal | $NaN | $NaN |
| Qwen | fal | $NaN | $NaN |
| Qwen | oneinfer | $NaN | $NaN |
| Qwen | oneinfer | $NaN | $NaN |
| Qwen | deepinfra | $NaN | $NaN |
| Qwen | deepinfra | $NaN | $NaN |
| Qwen | deepinfra | $NaN | $NaN |
| Qwen | deepinfra | $NaN | $NaN |
| Qwen | oneinfer | $NaN | $NaN |
Decision snapshot
| Decision factor | GLM-5.3 | Qwen3.8-Max |
|---|---|---|
| API model ID | glm-5.3 | qwen3.8-max |
| Independent intelligence score | 60 | 58 |
| Input price / 1M tokens | $1.40 | $2.00 |
| Cached input / 1M tokens | $0.26 | $0.25 implicit; $0.17 explicit read |
| Output price / 1M tokens | $4.40 | $6.00 |
| Context window | 1M tokens | 1M tokens |
| Maximum output | 128K tokens | 131K tokens |
| Input modalities | Text | Text, image, and video |
| Managed endpoint weights | Weights announced for release | Max service is proprietary; related base weights available |
How we keep this comparison honest
- Pin the rival to qwen3.8-max. Results for Qwen3.8 27B, Flash, older Qwen-Max releases, or community routes are not substituted.
- Use Artificial Analysis Intelligence Index v4.1.1 for both models in the independent table.
- Keep Z.ai's same-table coding and agent results labelled as vendor-reported evidence.
- Separate the managed Qwen3.8-Max endpoint from Qwen3.8-2.4T-A95B weights: Qwen documents Max-only additions such as vision, non-thinking mode, one-million-token defaults, and built-in tools.
- Record cache mode and region when pricing Qwen because implicit cache, explicit cache creation, and explicit cache reads have different rates.
Independent performance and efficiency
Artificial Analysis measured both reasoning models with Intelligence Index v4.1.1 on their first-party APIs. These figures are a dated snapshot and should be reproduced with your prompts and deployment region.
| Artificial Analysis metric | GLM-5.3 (max) | Qwen3.8-Max |
|---|---|---|
| Intelligence Index v4.1.1 | 60 | 58 |
| Cost per Intelligence Index task | $0.68 | $0.91 |
| Output speed | 84.1 tokens/s | 20.9 tokens/s |
| Total evaluation output tokens | 170M | 150M |
What the independent results mean
- GLM-5.3 leads measured general intelligence by 2 points, so the quality gap is narrow enough to justify a workload-specific bake-off.
- GLM generated output about four times faster, the largest practical separation in this comparison.
- GLM cost about 25% less per evaluated task despite Qwen using about 12% fewer output tokens.
- Qwen remains competitive on intelligence while adding visual input and non-thinking operation that base GLM-5.3 does not offer.
Matched coding and agent benchmarks reported by Z.ai
Z.ai lists GLM-5.3 and Qwen3.8-Max in the same release matrix. The rows are useful for task-level direction, but they remain vendor-reported and do not establish production reliability.
| Benchmark | GLM-5.3 | Qwen3.8-Max |
|---|---|---|
| Terminal-Bench 2.1 | 88.2 | 86.6 |
| DeepSWE v1.1 | 66.9 | 56.6 |
| NL2Repo | 58.0 | 55.9 |
| CyberGym | 84.5 | 78.5 |
| Toolathlon Verified | 73.0 | 72.5 |
| AutomationBench v1.0.6 | 48.2 | 39.8 |
| Agents' Last Exam (ALE-CLI) | 28.5 | 27.0 |
| HLE with tools | 62.5 | 56.2 |
Ready to test the workflow?
Capabilities and managed tools
| Capability | GLM-5.3 | Qwen3.8-Max |
|---|---|---|
| Reasoning control | Low, high, and max; always enabled | Thinking or non-thinking mode |
| Multimodal understanding | Text only | Text, image, and video |
| Function calling and structured output | Supported | Supported |
| Built-in hosted tools | Provider-dependent | Code interpreter, web extractor, web search, text-to-image search, and image-to-image search |
| Context and output | 1M context; 128K output | 1M context; 131K output; up to 262K reasoning |
| Prefix completion | Verify provider support | Supported through Partial Mode |
| Fine-tuning | Verify provider support | Qwen Cloud documents support |
Qwen caching and regional pricing
Qwen Cloud lists $2 input and $6 output per million tokens for qwen3.8-max, with implicit cached input at $0.25, explicit cache creation at $2.50, and explicit cache reads at $0.17. Alibaba Model Studio also publishes region-specific CNY rates. Calculate costs using the deployment region, cache strategy, and expected prompt reuse rather than treating one cache number as universal.
Managed Max is not identical to the downloadable checkpoint
Qwen3.8-Max is based on the downloadable Qwen3.8-2.4T-A95B checkpoint, a 2.4T-parameter MoE with 95B active parameters. Qwen explicitly identifies vision input, non-thinking support, a one-million-token default, and official built-in tools as additions to the managed Max service. Self-hosting the related checkpoint therefore does not guarantee endpoint parity, and its enormous size requires substantial distributed serving infrastructure. Review the Qwen3.8 license and reproduce the exact features you need before deployment.
Which model should you choose?
| Workload or priority | Recommended starting point | Why |
|---|---|---|
| Highest independent intelligence score | GLM-5.3 | It leads by 2 points in the current matched evaluation. |
| Fast interactive coding agents | GLM-5.3 | It generated output about four times faster. |
| Lower measured task cost | GLM-5.3 | It cost $0.68 versus $0.91 per independent evaluation task. |
| Image or video understanding | Qwen3.8-Max | Qwen accepts both; base GLM-5.3 is text-only. |
| Routine tasks without reasoning | Qwen3.8-Max | Qwen can disable thinking; GLM reasoning is always enabled. |
| Managed search and code tools | Qwen3.8-Max | Its endpoint documents a broad built-in tool catalog. |
| Downloadable related weights | Qwen3.8 base checkpoint | Weights are available, but do not assume full Max endpoint parity. |
| Repository coding and long-horizon agents | Test both | GLM leads most matched rows, but private task success and latency should decide. |