Zhipu AI pricing: API tiers and per-token cost classes

How Zhipu AI pricing is structured

The BigModel platform meters every API call at the token level, with separate input and output rates per model tier and a free trial credit applied at sign-up.

The zhipu ai pricing model on the BigModel platform is usage-based and billed per token. Every request to a GLM model consumes input tokens (the prompt plus any system message) and output tokens (the generated response). These two token classes are billed at different rates; output is typically priced higher because generation requires more compute than prefill. The billing granularity is fine enough that even very small requests appear on the usage dashboard within minutes of being made.

The cost varies across the model catalog. The small-tier variants carry the lowest rate and suit high-volume workloads where quality requirements are moderate — classification, entity extraction, short summarisation. The mid-tier covers the majority of production use cases at a rate that most teams find acceptable for everyday assistant and document-processing workloads. The flagship tier, covering the latest GLM-4.5+ generation, carries a higher rate that is justified for tasks where quality gains translate directly to user or product outcomes.

Beyond the standard chat-completions billing, the platform has separate rates for the code-specialised and multimodal variants. The code variants are typically priced at a rate close to the mid-tier chat models; multimodal calls that include image tokens are priced at a different schedule that accounts for the vision processing step.

Free trial credit and account activation

Every new BigModel account receives a free trial credit that covers meaningful evaluation volume without requiring a payment method upfront.

The free trial credit is applied automatically when an account is fully verified — email confirmation plus phone SMS. No payment method is needed to activate it. The credit amount is denominated in CNY tokens and is sufficient for a thorough evaluation run across the model catalog: testing the small, mid, and flagship tiers with the same set of prompts, running a few hundred requests through the code variants, and exploring the multimodal surface with a batch of sample images.

Once the trial credit is exhausted, the console surfaces a payment prompt. The platform accepts major international credit cards — Visa, Mastercard, and similar networks — as well as domestic Chinese payment methods. The billing cycle is monthly, and the dashboard shows a running tally of token spend in real time so a team can see exactly when the trial is approaching exhaustion and plan a payment method before it runs out.

Cost classes by model tier

Five model tiers on the BigModel platform each carry distinct input and output token cost classes.

The table below maps each model tier to its relative cost class. The figures here are reference classes rather than live prices — the BigModel platform console shows the current numeric rates, which the Zhipu AI team adjusts periodically. The relative ordering (small cheapest, multimodal highest per image token) has been consistent across recent pricing revisions.

Zhipu AI pricing — model tier vs. input and output token cost classes

Model tier	Input cost class	Output cost class
Small (GLM-4-flash, GLM-4-air)	Lowest — suitable for high-volume, latency-sensitive calls	Low — still below mid-tier output even at volume
Mid (GLM-4, GLM-4-plus)	Moderate — the practical optimum for most production workloads	Moderate — covers the bulk of assistant and document tasks
Flagship (GLM-4.5+)	Higher — justified where quality uplift is measurable in outcomes	Highest in the text-only range — significant for long generations
Code (GLM-Coder variants)	Close to mid-tier — code-specialised but not flagship priced	Close to mid-tier — output length for code completions tends to be shorter
Multimodal (GLM-4V)	Image tokens billed separately at a higher rate than text	Text output billed at mid-tier equivalent; image token ingestion is the main cost driver

Payment methods and billing currency

The BigModel platform bills in CNY at the underlying tier and converts for international cards at the gateway.

Domestic Chinese accounts pay in CNY directly. International accounts pay through a gateway that converts from CNY at the prevailing rate at the time of billing. For procurement teams that need a hard currency budget, the practical approach is to pre-load a CNY credit balance in an amount that represents your expected monthly spend, rather than relying on per-month gateway conversion. Most teams that adopt this approach find their finance team's internal cost-center reconciliation aligns more cleanly with a pre-loaded balance than with a variable monthly card charge in a foreign currency.

The platform does not currently offer contractual USD or EUR invoicing for standard accounts. Enterprise accounts with high volume can negotiate billing terms through the BigModel team directly. For evaluation purposes and moderate-volume production, the gateway conversion path is the standard route and works reliably for most international cards. Budget guidance from ai.gov on responsible AI procurement is a useful starting point for teams working through an internal approval process for ongoing API spend.

Estimating your monthly spend

A simple three-variable formula covers the spend estimate for any GLM API workload.

The three variables are: average request volume per day, average token count per request (input plus output combined), and the per-token rate for your model tier. Multiply them together and scale to a month. For a team sending 500 requests per day with an average of 800 tokens per request on the mid-tier, the daily token volume is 400,000 tokens, which at the mid-tier cost class resolves to a predictable monthly figure that most small teams find budget-friendly. The flagship tier at the same volume would be roughly two to three times higher.

The most common calibration mistake is underestimating output length. Prompts that produce long-form outputs — summaries of long documents, multi-step reasoning traces, or code generation with explanation — can consume three to five times more output tokens than the prompt itself. Running a batch of 100 representative requests through the API and checking the token counts in the BigModel dashboard is a faster path to an accurate estimate than modelling from word counts alone.