One API key, two GLM models: what routing easy steps to GLM-5.3-Flash does to the bill (prices checked 2026-10-27)

IteraCompute published a cost analysis showing that routing easy agent-loop steps to GLM-5.3-Flash, priced at roughly a tenth of GLM-5.3, cuts bills substantially when both models sit behind one OpenAI-compatible API key. Under a 70/30 token split with a 10:1 input-to-output ratio, the provider's list prices put the mixed workload at $40.80 per million tokens at a 0% cache-hit rate versus $110.80 for all-GLM-5.3 on the same endpoint, and $20.47 versus $55.00 at 90% cache hits. The write-up, which discloses it comes from one provider rather than a neutral ranking, also compares Z.ai official, Sail Research and Near AI endpoints, noting that splitting across two providers is cheapest overall.

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. 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: python 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. | 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. 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 | If you want to check our prices for all models on one page, they're on the IteraCompute pricing page https://iteracompute.com/pricing/?utm source=dev&utm medium=social&utm campaign=one api mix 202610&utm content=20261027 dev glm53mix 01 . 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.