Contents
Z.ai has one of the cheapest paid tiers anywhere and genuinely permissive open weights. GLM-4.7-FlashX lists at $0.07 in, $0.01 cached and $0.40 out per 1M tokens as of August 2026, with two free Flash tiers below it, while the GLM-5.2 flagship sits at $1.40 and $4.40 with a million-token context and 128K output. The international endpoint runs from Singapore and the weights are MIT licensed. One honesty note that changes the picture from earlier versions of this page: when we re-checked in August 2026 we could not reach Z.ai's privacy policy at any of the paths we tried, all of which returned 404, so the no-storage claim that circulates for this provider is something we currently cannot verify. Add that to the absence of a published SLA and you have a provider that is excellent value and thin on paperwork.
Best for
Cost-sensitive automation from a provider with a strong published data posture, coding agents via the Claude-compatible endpoint, and self-hosted deployments under a genuinely permissive licence.
Not for
Businesses that need a contractual uptime SLA, or that cannot use a China-headquartered vendor regardless of where the servers sit.
Z.ai (GLM) API pricing per 1M tokens
Z.ai prices per million tokens with cached input at roughly a fifth of the miss price, and publishes a wide catalogue that includes vision models and a dedicated OCR model at $0.03 in and out, which is relevant if your automation reads scanned paperwork. Two things are absent from the pricing documentation we could read: context windows for any model, and any mention of batch or off-peak discounts. Z.ai also sells subscription coding plans with prompt quotas rather than per-token billing, which for continuous engineering work can beat metered pricing, but those plan prices are not on the pricing page we checked and we are not going to quote figures we cannot source. Capacity figures do exist on individual model pages: GLM-5.2 is documented at a 1M context window with up to 128K output. Prices below are list prices as of August 2026.
| Model | Context | Input / 1M | Output / 1M |
|---|---|---|---|
| GLM-5.2 (flagship)Cached input $0.26 per 1M | 1M in / 128K out | $1.40 | $4.40 |
| GLM-5.1Cached input $0.26; same rates as 5.2 | Not published on the pricing page | $1.40 | $4.40 |
| GLM-5Cached input $0.20 | Not published on the pricing page | $1.00 | $3.20 |
| GLM-5-TurboCached input $0.24 | Not published on the pricing page | $1.20 | $4.00 |
| GLM-4.7Cached input $0.11; GLM-4.6 and 4.5 list at the same rates | Not published on the pricing page | $0.60 | $2.20 |
| GLM-4.5-AirCached input $0.03; the cheap tier of the older line | Not published on the pricing page | $0.20 | $1.10 |
| GLM-4.7-FlashX (budget)Cached input $0.01 per 1M; cheapest paid tier in this comparison | Not published on the pricing page | $0.07 | $0.40 |
| GLM-4.7-FlashFree tier, subject to rate limits; GLM-4.5-Flash is also free | Not published on the pricing page | Free | Free |
| GLM-4.6V-FlashX (vision)Cached input $0.004; GLM-4.6V-Flash is free | Not published on the pricing page | $0.04 | $0.40 |
| GLM-OCRDocument OCR at a flat rate both ways | Not published on the pricing page | $0.03 | $0.03 |
List prices as of August 2026, taken from the provider's published pricing page on August 5, 2026. Model prices move fast - check the source links before you budget.
The cheapest credible tier in the comparison
At $0.07 per million input tokens, $0.01 cached and $0.40 output, GLM-4.7-FlashX as of August 2026 is the cheapest paid tier from any provider covered on this site, and there are two free Flash tiers below it. Put it on the two workloads we price every provider against. Workload A, an inbox agent running 1,500 times a month at about 3,000 input and 400 output tokens, costs about $0.56 a month, or $0.34 once the stable 2,400-token prefix hits cache. Workload B, extracting eight fields from 4,000 documents a month at 6,000 input and 300 output tokens each, costs about $2.16.
Move up the ladder and it stays reasonable. On GLM-4.7 at $0.60 and $2.20 the same two workloads are about $4.02 and $17.04. On the GLM-5.2 flagship at $1.40 and $4.40 they are about $8.94 and $38.88. And if Workload B is genuinely scanned paperwork, GLM-OCR at $0.03 per 1M in and out would run it for about $0.76 a month, which is the point at which the cost line stops being worth a meeting.
That is the level at which cost stops being a design constraint. You no longer need to be clever about which records to process, sample your data, or route only ambiguous cases to a model. You can just run everything through it and spend your engineering attention on the parts that matter, like error handling and human review gates.
The subscription coding plans are the other angle worth understanding. Rather than per-token billing they sell prompt quotas with flagship access included, which for continuous engineering automation can be substantially better value than metered billing. Third-party sources quote monthly figures but they are not on the official pricing page we checked in August 2026, so check the current plan pricing yourself before assuming. We are not going to print a number we cannot source.
Two absences to plan around. No batch discount and no off-peak discount appear anywhere in the pricing documentation we could read, so unlike OpenAI, Anthropic, Google and Moonshot there is no half-price overnight route. Cached input at roughly a fifth of the miss price is the only lever, which makes prompt structure the whole optimisation story: stable material at the front, volatile material at the end.
One structural note on the lineup: GLM-4.6 is the version many people know and it has been superseded, with GLM-5.2 the flagship as of August 2026 and GLM-5.1 listed at identical rates. The older 4.5 line, including a cheap Air tier at $0.20 and $1.10, is still on the price list. As with every fast-moving provider, keep the model name in configuration rather than hardcoded.
The privacy position, stated carefully
There is a lazy assumption in a lot of Western commentary that Chinese AI providers are uniformly worse on data handling than US ones. Across the four we cover, that is not what the published documents show: Alibaba publishes an explicit no-training pledge, and Z.ai has historically published a no-storage claim stronger than the 30-day abuse-monitoring retention that OpenAI, Anthropic and Google all apply by default. DeepSeek and Moonshot, by contrast, both document training on inputs as the default. The picture is mixed in both directions, which is the only honest way to describe it.
For Z.ai specifically we have to downgrade our own earlier reporting. Previous versions of this page cited a published privacy policy stating that API prompts are processed in real time and not stored. Re-checking in August 2026, every privacy policy path we tried on z.ai and docs.z.ai returned a 404, and the pricing documentation is silent on data handling. The claim may still be accurate and a site reorganisation is the likeliest explanation, but we cannot cite a document we cannot open, so we are treating it as unverified until we can.
The practical consequence is a procurement one. If you need to tell a client or an auditor what happens to their data at this vendor, you currently cannot point at a linkable page, and that is the same objection we raise about xAI. Ask Z.ai for the current policy in writing before you send anything sensitive. If they produce it and it says what it used to say, this is a strong data position at a very low price.
The everything else, for a China-headquartered vendor, includes considerations that have nothing to do with the policy text. Your clients may have views. Your insurer or your own contractual commitments to customers may constrain which jurisdictions your subprocessors can sit in. Procurement at a larger customer may simply have a list. Those are real business constraints and they are legitimate reasons to choose differently, quite separately from any technical assessment.
The gap we would weigh most heavily is the absence of a published SLA. A privacy policy tells you what happens to your data. An SLA tells you what happens when the service is down at the worst possible moment, and there is no contractual answer here. If you build something business-critical on Z.ai, build the failover to a second provider at the same time.
How it fits agent and coding workloads
Z.ai has clearly aimed the GLM line at coding and agent work. Function calling, MCP support and structured outputs are all there, and the Anthropic-compatible endpoint plus an official Claude Code setup helper make it a genuine drop-in for teams already running that toolchain. That compatibility is the most practically useful thing about the provider: it turns evaluating it into a fifteen-minute experiment rather than a project.
The vendor publishes strong benchmark figures on software engineering tasks, including direct comparisons that place GLM-5.2 within a percentage point of a named Claude model on one coding benchmark and ahead of named OpenAI models on others. We are not going to repeat those as established fact, because vendor-run benchmarks select for the vendor and we have not seen independent replication. Treat them as a statement of where the company is aiming rather than a measurement you can plan around.
The million-token context with up to 128K output on GLM-5.2 is a useful combination for document work, because a large output limit is rarer than a large input limit and it is what you need when the automation has to produce a long document rather than just read one. The same caution applies as everywhere: LOCA-bench, published in early 2026, found that most models including frontier and open-source ones degrade sharply as context grows, so treat the window as a ceiling and compact between steps. Z.ai's own claim that GLM-5.2 maintains more stable performance at ultra-long context is a vendor claim, and it is exactly the kind of claim your own twenty-case eval can settle.
Our practical recommendation: put GLM-4.7-FlashX in as the cheap first-pass tier of a two-model automation and see what it handles. At seven cents a million input tokens the experiment costs nothing, and if it clears your evaluation set on the routine steps, your flagship spend drops to the fraction of work that genuinely needs it. If you want that evaluation and the resulting build done for you, start with a free automation audit.
Where Z.ai is the wrong choice
Three situations, and none of them are about the models. The first is anything business-critical without a failover. There is no published uptime SLA, so when the endpoint is down mid-workflow you have no contractual remedy and no credit to claim. If an automation genuinely matters, either build failover to a second provider on day one or use a provider that will commit to an availability number in writing.
The second is anything where you have to show a client, an insurer or an auditor what happens to their data. As of August 2026 we could not open a privacy policy at any path we tried, which means you currently cannot hand anyone a link. That is the same objection we raise about xAI and it is a procurement blocker regardless of how good the underlying practice may be. Ask Z.ai for the document; if it arrives and says what it used to, this objection disappears.
The third is any workload that leans on overnight batching. No batch discount and no off-peak discount appear in the pricing documentation, so a job that would run at half price on OpenAI, Anthropic or Google runs at full price here. Given how low the full price already is, that rarely changes the answer, but it does narrow the gap more than the headline rates suggest.
Where it is emphatically the right choice: cost-sensitive volume work on data whose sensitivity you can assess yourself, coding agents through the Anthropic-compatible endpoint, scanned-document pipelines using GLM-OCR at $0.03 per 1M, and any deployment where the MIT licence lets you move the whole thing on-premise if the requirements tighten. That escape hatch is worth more than most of the paperwork it substitutes for.
Strengths for agent workloads
- The cheapest paid tier in the whole comparison at $0.07 in and $0.40 out, with two free Flash tiers below it for prototyping
- MIT-licensed open weights on the flagship, which is as permissive as it gets and better than Meta's licence terms
- A 1M context window with up to 128K output on GLM-5.2, documented on the model page
- Function calling, MCP support, context caching, thinking mode and structured outputs all documented on the flagship
- First-class Claude Code compatibility with an Anthropic-format endpoint and an official setup helper
- A dedicated OCR model at $0.03 per 1M in and out, which makes scanned-document pipelines almost free to run
- Cheap vision tiers, with GLM-4.6V-FlashX at $0.04 in and a free vision Flash tier
Weaknesses
- We could not reach Z.ai's privacy policy at any path we tried in August 2026, so retention and training defaults are currently unverifiable from where we sit
- No published uptime SLA, which is a genuine problem for anything business-critical
- No batch or off-peak discount appears in the pricing documentation, so caching is the only lever
- Context windows are not published on the pricing page at all, so capacity planning means hunting per-model pages
- Capability claims come from vendor benchmarks, including direct comparisons against named competitor models, that we cannot independently verify
- Subscription coding plan prices are not published where we could find them, making them impossible to compare against per-token billing
- The China endpoint is a separate service with separate terms, and it is easy to point at the wrong one by mistake
How to access Z.ai (GLM)
- The international API at api.z.ai, operated from Singapore. This is the endpoint to use outside China.
- An Anthropic-compatible endpoint path, which means Claude Code and Claude-shaped agent harnesses work with a base URL and key change. Z.ai publishes an official setup helper for this.
- The separate China platform at open.bigmodel.cn, which is a distinct service with distinct pricing and terms. Check which one your SDK defaults to.
- OpenRouter, where GLM-5.2 is carried by third-party hosts at prices that differ from first-party, sometimes lower on input and higher on output.
- Self-hosting the MIT-licensed weights from Hugging Face. The full flagship needs multi-node H100-class hardware, but quantised community builds bring smaller deployments within reach.
Data privacy and residency
- We have to correct our own earlier position here. Previous versions of this page reported that Z.ai's published privacy policy states API prompts are processed in real time and not stored. When we re-checked in August 2026, every privacy policy path we tried on z.ai and docs.z.ai returned a 404, and the pricing documentation says nothing about data handling. We are therefore treating that no-storage claim as unverified rather than repeating it as fact. If it matters to your decision, ask Z.ai for the current policy document and read it yourself.
- That is not an accusation. A vendor reorganising its site is the most likely explanation, and the claim may well be accurate. But the standard we hold every provider to on this site is that we cite what we can read, and this month we could not read it.
- The international service is operated from Singapore, not mainland China, so an ordinary transfer to Singapore is what you are documenting rather than a transfer to the PRC. Confirm your integration points at api.z.ai and not the separate China platform.
- The company is headquartered in China. For some businesses, and for some of their clients and insurers, that is itself a decision point regardless of where the servers sit and what the policy says. That is a commercial and reputational judgement rather than a technical one, and it is yours to make.
- There is no published uptime SLA that we could find. For a business-critical automation that is a real gap: a good privacy policy does not help you when the endpoint is down and you have no contractual remedy. Design for failover to a second provider if you build anything important on it.
- The MIT licence on the weights gives you the ultimate privacy escape hatch. If the policy position is not enough for a particular workload, run the model on your own infrastructure and no policy is involved at all.
What we build on Z.ai (GLM)
- Cost-sensitive high-volume classification and extraction where the FlashX tier makes the economics trivial
- Coding agents through the Claude-compatible endpoint, as a much cheaper backend for an existing harness
- Prototyping on the free Flash tier before committing to a paid provider
- Self-hosted deployments under MIT where you want a capable model with no vendor relationship at all
- The volume tier of a two-model stack, paired with a US flagship for judgement steps
Model choice is a line item, not a strategy. If you want the workflow designed, built and handed over with a plain-language report of what it ran and what it saved, start with a free AI automation audit or see the full done-for-you automation services.
Z.ai (GLM) FAQ
How much does the Z.ai GLM API cost?
As of August 2026, GLM-5.2 and GLM-5.1 list at $1.40 per 1M input, $0.26 cached and $4.40 output. GLM-5 is $1.00 and $3.20, GLM-4.7 is $0.60 and $2.20, GLM-4.5-Air is $0.20 and $1.10, and GLM-4.7-FlashX is $0.07 and $0.40 with cached input at $0.01. GLM-4.7-Flash and GLM-4.5-Flash are free subject to rate limits. There is also a GLM-OCR model at a flat $0.03 in and out.
What would a real GLM automation cost per month?
A 1,500-run inbox agent at about 3,000 input and 400 output tokens per run costs roughly $0.56 a month on GLM-4.7-FlashX, or $0.34 with the stable prefix cached. A 4,000-document extraction job at 6,000 input and 300 output tokens each is about $2.16. On GLM-4.7 those become about $4.02 and $17.04, and on GLM-5.2 about $8.94 and $38.88. There is no batch discount published, so those are the numbers.
Does Z.ai store or train on API data?
We could not verify this in August 2026 and we are not going to guess. Every privacy policy path we tried on z.ai and docs.z.ai returned a 404, and the pricing documentation says nothing about data handling. Earlier versions of this page reported a published no-storage claim; we are treating that as unverified until we can open the document. Ask Z.ai for the current policy in writing before sending anything sensitive.
Where is Z.ai data processed?
The international API at api.z.ai is operated from Singapore. The China platform at open.bigmodel.cn is a separate service with separate terms and pricing. Check which endpoint your SDK or integration actually points at, since defaulting to the wrong one is an easy mistake.
Can I self-host GLM models?
Yes, and this is a genuine strength. The flagship weights are published on Hugging Face under the MIT licence, which places no commercial restrictions on you at all. The full model needs multi-node H100-class hardware, but quantised community builds bring smaller deployments within reach of a single well-specified server.
Does GLM work with Claude Code?
Yes. Z.ai publishes an Anthropic-compatible endpoint and an official setup helper, so Claude Code and Claude-shaped agent harnesses work with a base URL and key change. That makes evaluating GLM as a cheaper backend a fifteen-minute experiment rather than an integration project.
What is the catch with Z.ai?
Three things, and they are all paperwork rather than product. There is no published uptime SLA, so if the endpoint goes down mid-workflow you have no contractual remedy: build failover to a second provider if the automation matters. We could not reach a privacy policy at all in August 2026. And capability claims come from vendor-run benchmarks we cannot independently verify. The models are cheap and capable; the documentation around them is the weak part.
Does Z.ai have a batch discount?
None appears in the pricing documentation we could read as of August 2026, and no off-peak discount either. Cached input at roughly a fifth of the miss price is the only lever, which puts the whole optimisation burden on prompt structure: keep the system prompt, tool schema and examples at the front where they cache, and the customer name, timestamp and retrieved documents at the end where they do not need to.
What is the GLM-5.2 context window?
The pricing page publishes no context windows for any model, but the GLM-5.2 model page documents a 1M-token context window with up to 128K maximum output. We did not confirm figures for the other tiers, so check the specific model page for whichever one you intend to use rather than assuming the family shares a limit.
Is GLM-4.6 still current?
No. GLM-4.6 has been superseded, with GLM-5.2 the flagship as of August 2026 and GLM-4.7 and 4.7-FlashX the value tiers. There is no Air variant of the GLM-5 line. As with any fast-moving provider, keep the model name in configuration rather than hardcoded so a lineup change is a config edit.
