GLM 5.3 Flash Free vs Paid: Current Routes, API Price, and Cost per Accepted Task

The four official GLM-5.3-Flash routes

The routes use different endpoints and commercial rules. The ZCode connection guide assigns Coding Plan traffic to https://api.z.ai/api/coding/paas/v4. Prepaid balance and pay-as-you-go calls use the general endpoint, https://api.z.ai/api/paas/v4. Z.ai says the coding endpoint is limited to coding scenarios and cannot replace the general endpoint.

What the free route includes

ZCode documents a free daily trial quota for new users who connect a Z.ai account. The global Z.ai section does not restate the exact GLM-5.3-Flash allowance or trial duration in its prose, so the balance shown in ZCode should decide what you can use. The same page gives exact terms for the BigModel connection used in China: the trial lasts five days and supplies 5 million GLM-5.3-Flash tokens per day. Do not copy that China-specific allowance into a global Z.ai budget without checking the live provider panel. ZCode documents both connection paths and their trial wording.

Z.ai advertised a paid-plan Coding Plan campaign for September 3–20, 2026. Its published terms give GLM-5.3-Flash use through ZCode zero quota consumption from 23:00 to 09:00 Singapore time during the campaign, while other supported agents get double the standard quota in that window. It still requires a paid plan, and an exhausted limit can block participation. The campaign is at the end of its advertised dates as this article is updated; do not treat it as an ongoing free tier or include it in a future budget without checking the official terms and your live account.

Self-hosting removes the Z.ai per-token API charge. Z.ai publishes the GLM-5.3-Flash weights under the MIT license and documents local serving with frameworks including vLLM and SGLang. The model has 320 billion total parameters with 18 billion active, so the remaining bill comes from infrastructure and operations. The official model card establishes the license and deployment options. You still need a measured serving cost for your own hardware and workload.

What the paid routes buy

Coding Plan

Coding Plan quota covers individual coding work in approved tools. Z.ai's subscription terms exclude websites, bots, SaaS products, and other general-purpose API traffic unless Z.ai has agreed otherwise in writing.

The official ZCode page currently displays promotional monthly prices of $12.60 for Lite, $56 for Pro, and $117.60 for Max, alongside reference prices of $18, $80, and $168. It describes Pro as six times Lite usage and Max as fourteen times Lite usage. The page does not show an end date for those lower prices and says checkout on Z.ai has the final details, so use the amount on the order page for a purchase decision. See the current public plan cards.

Coding Plan measures quota in credits, with Lite at 10,000 weekly credits, Pro at 60,000, and Max at 140,000. Separate five-hour limits also apply. For GLM-5.3-Flash, the current formula multiplies new input tokens by 2.3, cached input tokens by 0.56, and output tokens by 8, then divides by 10,000. Off-peak model use consumes half the standard credits. These details make a simple “requests per month” promise unreliable because a short cached edit and a long agent run consume very different amounts. Z.ai publishes the current credit formula and limits.

Pay-as-you-go API

The general API fits an application backend, an evaluation harness, or any workflow that needs per-request billing records outside supported coding tools. Use model code glm-5.3-flash at the general Z.ai API endpoint. The current official price table lists:

“Limited-time free” applies only to cached-input storage in that table. It does not make GLM-5.3-Flash input or output free. Third-party gateways can set different prices, routing, and free-credit rules, so a gateway invoice should not be described as the direct Z.ai API rate.

GLM-5.3-Flash defaults to max reasoning effort and supports low, high, and max; thinking cannot be disabled. The model documentation recommends max. That setting carries no cost guarantee, so record the effort level with every evaluation because it can change token use and acceptance. Z.ai also recommends clear_thinking: false for preserved reasoning in agent work, while its general API defaults to clearing prior reasoning. Retaining earlier reasoning can move more history into later requests, including cached input. The chat-completion reference documents these parameters.

Calculate the API cost of one attempt

The API response reports prompt_tokens, completion_tokens, and prompt_tokens_details.cached_tokens. Subtract cached input from total prompt input before applying the new-input rate:

new_input_tokens = prompt_tokens - cached_tokens

attempt_cost = (
    new_input_tokens * 0.15
  + cached_tokens * 0.03
  + completion_tokens * 0.50
) / 1,000,000

Use the provider's billing readback when it differs from a reconstructed estimate. Keep the request ID, model ID, reasoning effort, token fields, timestamp, retry number, and outcome on the same ledger row. A response that is syntactically valid but fails your acceptance check belongs in the cost total.

Calculate cost per accepted task

Define acceptance before the first run. For a code change, acceptance might require the requested behavior, a passing test suite, no modification to protected files, and reviewer approval. For extraction, it might require schema validation and a manual check of sampled records. An API response is not automatically an accepted task.

API cost per accepted task =
  total GLM API charges across all attempts / accepted tasks

All-in cost per accepted task =
  (GLM charges + validation compute + other tools + review labor)
  / accepted tasks

Include retries, rejected outputs, abandoned agent turns, and billable failures in the numerator. When zero tasks pass, report the total spend and NO_ACCEPTED_TASK; division by zero does not produce a meaningful unit cost.

Consider a transparent batch example. Across all attempts, a team records 1.2 million new input tokens, 4.8 million cached input tokens, and 600,000 output tokens. The direct Z.ai API calculation is:

(1.2 * $0.15) + (4.8 * $0.03) + (0.6 * $0.50)
= $0.180 + $0.144 + $0.300
= $0.624 total API cost

If 48 tasks pass the frozen acceptance check, the API cost per accepted task is $0.624 / 48, or $0.013. If only 36 pass, it rises to about $0.0173. The token bill remains small in both cases. If review and validation for the batch consume two staff hours at an internal rate of $30 per hour, the all-in cost at 48 accepted tasks becomes ($0.624 + $60) / 48, or about $1.26. Those labor figures are illustrative. Replace them with your own recorded time and rate.

For Coding Plan, replace API charges with the portion of the subscription fee allocated to the workload:

plan cost per accepted task =
  allocated subscription fee / accepted tasks

Use the full invoice when the plan serves only this workflow. If several workflows share the subscription, choose one allocation rule, such as credits consumed, and apply it every month. Unused quota still has a cost. The free trial has zero subscription cost, but validation time, retries, and migration work remain in the all-in figure.

Which GLM-5.3-Flash route should you use?

Start the ZCode trial on a disposable public or synthetic repository. Give it tasks that resemble your work without exposing private code, secrets, or Git history. Save the tasks and acceptance results, then decide whether to review the current client and your organization's data requirements for private work. The workspace-upload findings above concern 3.12.3; the reported 3.14.0 change and the company's deletion claim have not been independently tested by Grow AI Skills.

Coding Plan fits work that happens interactively in supported coding agents and stays within the five-hour and weekly limits. Check the live plan price, supported-tool list, and reset windows before subscribing. Do not design a production service around plan credentials.

Use the general API when software calls the model, when you need per-request usage records, or when tasks must run outside Z.ai's approved coding tools. Run the same acceptance checks during evaluation and production so the cost denominator stays comparable.

Self-hosting becomes credible when deployment control, data location, or sustained utilization can justify operating a 320B-parameter mixture-of-experts model. Compare hardware rental, idle capacity, engineering time, and accepted throughput against the managed API. A zero-dollar token line by itself is not a cost comparison.

The broader Grow AI Skills coding-model comparison reports one third-party solve-rate result for GLM-5.3-Flash, but that source did not disclose its realized cost per correct result. Do not combine that solve rate with today's list price and call the result measured. Run the model on your own frozen tasks, export the token ledger and acceptance decisions, then divide the complete spend by the tasks that actually passed.