This is IteraCompute, so treat this as a disclosed write-up from one provider, not a neutral ranking. Every price below is a list price in USD per 1M tokens, checked on 2026-10-27 from the providers' pricing pages or OpenRouter's endpoint list. Prices change often, so check them again before you commit.
The idea: stop sending every step to the big model
Most agent loops are made of a few hard steps (planning, writing the actual patch, the final answer) and many easy ones (classifying the request, summarising a file or tool output, extracting fields, checking formats, writing commit messages). If every step goes to GLM-5.3, you pay GLM-5.3 prices for the easy work too.
GLM-5.3-Flash is roughly a tenth of the price of GLM-5.3. If the endpoint is OpenAI-compatible and both models sit behind the same key, routing is one field:
from openai import OpenAI
client = OpenAI(base_url="https://api.iteracompute.com/v1", api_key="YOUR_KEY")
EASY = {"classify", "summarize_tool_output", "extract", "commit_message"}
def run(step: str, messages: list[dict]) -> str:
model = "z-ai/glm-5.3-flash" if step in EASY else "z-ai/glm-5.3"
r = client.chat.completions.create(model=model, messages=messages)
return r.choices[0].message.content
The same pattern works with any OpenAI-compatible provider that serves both models. What changes is the bill.
Prices used (USD per 1M tokens, 2026-10-27)
| Model | Provider | Input | Cached input | Output |
|---|---|---|---|---|
| GLM-5.3 | Z.ai (official) | 1.40 | 0.26 | 4.40 |
| GLM-5.3 | IteraCompute | 0.82 | 0.20 | 2.88 |
| GLM-5.3 | Sail Research (fp8, via OpenRouter) | 0.50 | 0.08 | 2.00 |
| GLM-5.3-Flash | Z.ai (official) | 0.15 | 0.03 | 0.50 |
| GLM-5.3-Flash | IteraCompute | 0.08 | 0.023 | 0.28 |
| GLM-5.3-Flash | Near AI (fp8, via OpenRouter) | 0.08 | 0.012 | 0.50 |
| DeepSeek V4.1 Flash | IteraCompute | 0.14 | 0.006 | 0.55 |
Sail Research and Near AI are the cheapest standard (no promo) endpoints that declare fp8 for each model. One exception: at 0% cache hit, AkashML (fp8, 0.19 / 0.19 / 4.40) is a little cheaper than Sail for GLM-5.3, which would make that row $28.00 instead of $30.10 in the first table.
Monthly bill: 100M input + 10M output tokens
Workload assumption: 70% of tokens go to the cheap model and 30% to GLM-5.3, with the same 10:1 input-to-output ratio on both. Cache-hit rate is the share of input tokens billed at the cached price.
| Setup | 0% | 50% | 80% | 90% |
|---|---|---|---|---|
| All GLM-5.3, Z.ai official | $184.00 | $127.00 | $92.80 | $81.40 |
| All GLM-5.3, IteraCompute | $110.80 | $79.80 | $61.20 | $55.00 |
| All GLM-5.3, Sail Research fp8 | $70.00 | $49.00 | $36.40 | $32.20 |
| 70/30 mix, Z.ai official (one key) | $69.20 | $47.90 | $35.12 | $30.86 |
| 70/30 mix, IteraCompute (one key) | $40.80 | $29.51 | $22.73 | $20.47 |
| 70/30 mix with DeepSeek V4.1 Flash as the cheap model, IteraCompute | $46.89 | $32.90 | $24.51 | $21.71 |
| 70/30 mix, Sail (GLM-5.3) + Near AI (Flash), two providers | $30.10 | $21.42 | $16.21 | $14.48 |
Same workload, 50/50 split:
| Setup | 0% | 50% | 80% | 90% |
|---|---|---|---|---|
| 50/50 mix, Z.ai official | $102.00 | $70.50 | $51.60 | $45.30 |
| 50/50 mix, IteraCompute | $60.80 | $43.88 | $33.72 | $30.34 |
| 50/50 mix, Sail + Near AI | $41.50 | $29.30 | $21.98 | $19.54 |
What the numbers say
- Routing saves more than switching hosts. On IteraCompute, the 70/30 mix costs about 63% less than sending everything to GLM-5.3 there. It is also 36–42% cheaper than sending everything to the cheapest fp8 GLM-5.3 host. At 50/50 that gap shrinks to 6–13%, so how much traffic you can move to Flash matters most.
- Against the official API, the same mix on one key is 33–41% cheaper on IteraCompute at both splits.
- We are not the cheapest way to run this mix. If you buy GLM-5.3 from Sail and Flash from Near AI, our bill is 38–46% higher at 70/30 and 50–60% higher at 50/50. You get there through OpenRouter with provider pinning (one key, plus OpenRouter's fees) or by opening accounts with both providers.
- The cheap model choice matters too. DeepSeek V4.1 Flash has a much lower cached price (0.006), but its input and output prices are higher, so in this workload GLM-5.3-Flash still comes out cheaper at every cache-hit rate.
Listed separately: promos and undeclared precision
- Time-limited promos: GMICloud runs about 50% off both models (GLM-5.3 0.70 / 0.13 / 2.20, Flash 0.075 / 0.015 / 0.25). The 70/30 mix there costs $34.60 / $23.95 / $17.56 / $15.43, cheaper than ours while the promo lasts.
- Endpoints that don't declare precision (for example DekaLLM or NanoGPT for GLM-5.3, LLM Gateway or ArgoLink for Flash) can bring the 70/30 mix down to about $24.26 / $20.41 / $15.29 / $13.52. fp4 endpoints are cheaper still. Neither is a like-for-like fp8 comparison, so they are left out of the tables.
Caveats before you copy this
- Token counts are an assumption. GLM-5.3-Flash always runs with reasoning on, so measure output tokens per step on your own traffic before trusting the split.
- Route by step type, not by guessing after the fact. Keep a fallback: if a Flash step fails validation, retry it on GLM-5.3 and count those retries in the bill.
- Keep the stable part of the prompt (system prompt, tool definitions, repo context) at the front so it hits the cache. At 80–90% cache hit, the cached-input price is a large part of the bill.
- Log model, input, cached and output tokens per step. Then you can recompute this table with your real numbers.
If you want to check our prices for all models on one page, they're on the IteraCompute pricing page. Model IDs are z-ai/glm-5.3, z-ai/glm-5.3-flash and deepseek/deepseek-v4.1-flash on an OpenAI-compatible endpoint.
Top comments (0)