GLM-5.3 Pricing: An Operating Budget for API and Coding Plans

Updated October 22, 2026 · first published October 22, 2026

GLM-5.3’s API list price is $1.40 per million uncached input tokens and $4.40 per million output tokens; cached input is $0.26 per million. Those rates make a single request look inexpensive, but a coding agent repeats context, generates reasoning and code, calls tools, and sometimes retries. Budget the workload and the path to an accepted change, not a headline token price.

This guide is an operating-budget reference for the API and subscription. It is separate from the Ox Alpha article, which covers a temporary free listing and its rate-cliff risk. Prices and documentation checked 22 October 2026. Prices and plan terms can change; use the live Z.ai pricing table when approving a budget.

What does GLM-5.3 cost per token?

Z.ai lists GLM-5.3 at $1.40/M input, $0.26/M cached input, and $4.40/M output. GLM-5.3-Flash is a separate model and rate card: $0.15/M input, $0.03/M cached input, and $0.50/M output. The pricing page currently labels cached-input storage “Limited-time Free” for both. Treat storage as a temporary zero-priced line, not a permanent entitlement, and recheck before forecasting. These prices are provider-specific list rates; another host’s bill can include account plans, usage units, minimums, or infrastructure charges.

The official GLM-5.3 model guide says reasoning is always enabled, with low, high and max effort settings; the default is max. There is no reasoning-off setting. Reasoning is not a separate free allowance in this budget. Include all reported output tokens in your estimate, including reasoning tokens when the usage response reports them as output. The API’s actual usage fields and invoice should settle the accounting treatment for your endpoint. Z.ai’s model guide exposes reasoning content in streamed responses, while its pricing table has one output rate; do not model visible final text alone as the whole generation.

What does a representative request cost?

Here is a reproducible token mix for one agent turn: 100,000 input tokens total, of which 80,000 qualify as cached and 20,000 are uncached, plus 10,000 output tokens. Formula: (uncached input × input rate) + (cached input × cached-input rate) + (output × output rate), with token counts divided by one million.

At 1,000 such turns in a month, inference is $92.80 on GLM-5.3 or $10.40 on Flash. If an accepted task averages two identical turns after a repair or review pass, that becomes $0.1856 or $0.0208 per successful task, and $185.60 or $20.80 for 1,000 accepted tasks. These are illustrative calculations, not a performance claim: the model, context size, cache hit rate, output length, and probability of acceptance all change the result.

For a retry-aware unit cost, use total billed inference and tool charges divided by completed, accepted tasks. If 1,000 tasks are accepted after 1,400 calls, and each call has the same mix above, GLM-5.3 costs $130.72 total, or about $0.131 per accepted task. A failed task still consumed tokens; dividing by started tasks or successful API calls hides that loss.

How should a team budget a Coding Plan?

Z.ai sells the GLM Coding Plan as a subscription for supported coding agents. The official plan overview says pricing starts at $18/month and gives Lite, Pro, and Max weekly credit allowances of 10,000, 60,000, and 140,000, alongside 5-hour limits of 2,000, 12,000, and 28,000. Credits are metered using model-specific input, cached-input, and output multipliers; quota resets on different schedules, and the documentation describes a 50% off-peak model-credit rate during stated hours. A subscription is therefore a capped credit and access arrangement, not “unlimited API tokens.” Z.ai also ties concurrency to tier and resource availability, restricts use to supported tools, and prohibits account sharing. Check current checkout pricing, quota rules, eligible models, and supported agent tools before procurement.

Keep subscription spend separate from metered API spend. Z.ai documents separate endpoints for the Coding Plan (/api/coding/paas/v4) and general API balance or resource packages (/api/paas/v4); the endpoints are not interchangeable. A Coding Plan request should not be valued by multiplying tokens by the public API list rates, and a general API request should not be assumed covered by the subscription. Record which credential, endpoint, plan, and quota pool actually served each request. See Z.ai’s Coding Plan credit rules and endpoint setup guide.

For procurement, estimate both a normal month and a burst month. Apply the measured weekly calls to the plan’s 5-hour and weekly pools, then price any overflow on the actual route it will use. Include seats that go unused, peak-hour versus off-peak timing, and tool quotas. Do not convert credits to a universal token allowance: Z.ai’s own estimates vary with cache-hit rate and model mix. Revisit the calculation after a model switch or plan revision.

What costs sit outside the model invoice?

Open weights do not make an operating service free. Self-hosting shifts the bill to accelerators, memory, storage, networking, orchestration, monitoring, security work, and staff time for deployment and upgrades. Hosted inference also leaves costs outside tokens: agent runtime or compute, tool and search calls, code execution, artifact storage, observability, egress, and human review. Z.ai’s pricing page, for example, separately lists Web Search at $0.01/use. Put each item in the ledger, even when it is included in another subscription or currently promotional.

A useful monthly ledger records date, environment, team, feature, task ID, provider, endpoint, exact model ID, plan or API credential class, input tokens split into cached and uncached, output and reasoning tokens, tool calls, retries, accepted outcome, and invoice amount/currency. Save cache-hit rate and the rate-card version too. This lets finance reconcile invoices and lets engineering compare prompt changes without treating every token as equivalent.

How should GLM-5.3 be compared across hosts?

Compare the service contract and the same workload, not just the model name. Cloudflare Workers AI currently lists GLM-5.3 at the same token prices as Z.ai’s API table, but bills through Neurons and requires Workers Paid or prepaid AI Gateway credits for this model; its pricing page describes a $0.011 per 1,000-Neuron rate and a daily free allocation, while certain models require paid billing. Confirm the active billing path and whether the included allocation applies before counting it as savings. See the Cloudflare GLM-5.3 page and Workers AI pricing rules.

Then replay a representative coding workload and track cost per accepted task, latency, review time, tool spend, and failure rate. A lower token bill can lose if it needs extra generations or human correction; a higher rate can win if it materially reduces total work. The durable budget is the one based on measured outcomes and a named provider route.

Related

• Ox Alpha was GLM-5.3: the cost of a free window
• Cost per successful task
• Why provider prices are not directly comparable

Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →

Back to finopsllm.com