GLM 5.2 API Pricing 2026: GLM-5.2, GLM-5.3 and Every Z.ai Model per Million Tokens, Cache, Free Models, Coding Plan and Cost per Conversation

Published 29 October 2026. Every Z.ai price below was read from the Z.ai pricing page and every renminbi price from the bigmodel.cn API pricing page on 29 October 2026, and every price is as of that date.

GLM-5.2 costs $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens on Z.ai's API, as of 29 October 2026, with cache storage free for a limited time and one flat rate across its 1-million-token context window. It is no longer Zhipu's flagship. GLM-5.3, which Z.ai released on 18 August 2026, is built on the same base model with all of its gains from post-training, and it costs exactly the same on every line. The cheaper member of the family is GLM-5.3-Flash at $0.15 input, $0.03 cached and $0.50 output, and three older Flash models are still free.

So "GLM 5.2 pricing" now has a slightly odd answer: the model you searched for and the model that replaced it are priced identically, and the real choice between them is about reasoning, not rate. GLM-5.3 always thinks and cannot be told not to; GLM-5.2 can still switch thinking off, which on a chat workload is the larger cost lever. In the worked example further down, a 40-turn conversation costs $0.728 on either model without caching and $0.210 with the cache warm, which is a third less than Kimi K3.1 Pro or Claude Sonnet 5.5 before any reasoning tokens are counted, and $0.078 and $0.024 on GLM-5.3-Flash.

GLM API price per million tokens, every current model on Z.ai

US dollars per million tokens from the Z.ai pricing page, read 29 October 2026; context windows from Z.ai's model overview. GLM-5.3, GLM-5.2 and the GLM-5.3-Flash family have a maximum output of 128K tokens.

Vision models sit on the same page: GLM-4.6V costs $0.30 input, $0.05 cached and $0.90 output, GLM-4.6V-FlashX $0.04, $0.004 and $0.40, GLM-4.5V $0.60, $0.11 and $1.80, GLM-OCR $0.03 in and out, and GLM-4.6V-Flash is free. GLM-5.3-Flash is itself natively multimodal, taking image, video and file input at the same per-token rates, so for most new vision work it is the model to price first.

Three things in this table matter more than the rest. First, the top three rows are one price: GLM-5.3, GLM-5.2 and GLM-5.1 all cost $1.40 and $4.40, so there is no discount for staying on an older flagship, and the only reasons to pin glm-5.2 are behaviour you have already tested and the thinking switch covered below. Second, the cache hit rate is $0.26, which is 18.6% of input, not the 10% that Claude, Kimi and Qwen's explicit cache charge. Third, the Flash tier is an order of magnitude cheaper than the flagship and, according to Z.ai's release notes, runs on a different 320-billion-parameter hybrid-attention architecture with 18 billion parameters active, so it is a separate model rather than a throttled one. Z.ai describes it as "delivering stronger intelligence than GLM-5.2 at an exceptionally low cost", which is a vendor claim worth testing on your own prompts before you move a workload.

GLM prices in CNY on bigmodel.cn

Zhipu runs two platforms with separate price lists. The international one at z.ai bills in US dollars under a Singapore company; the China platform at bigmodel.cn bills in renminbi, and its list is not a conversion of the dollar one.

Renminbi per million tokens from docs.bigmodel.cn, read 29 October 2026. Cache storage is listed as a limited-time free item in renminbi per million tokens per hour, and the page says it does not yet show the price that will apply when the free period ends.

Beyond currency, the lists differ in structure. On bigmodel.cn, GLM-5.1, GLM-5 and GLM-4.7 are priced in tiers by input length, while Z.ai charges one flat rate for each. The China list carries models the international one does not, GLM-5-Turbo and GLM-5V-Turbo among them. And the cache discount is smaller in renminbi: a GLM-5.3 cache hit costs a quarter of input on bigmodel.cn (¥2 against ¥8) and under a fifth on Z.ai ($0.26 against $1.40). bigmodel.cn also offers its Batch API at half the standard price for supported models; Z.ai's international pricing page lists no batch rate. Budget in the currency and on the platform you will be billed on.

How GLM caching is billed, and the three numbers Z.ai gives for it

Caching on the GLM API is implicit. Z.ai's context caching guide says the system "automatically identifies and reuses" repeated context with no configuration, and reports the hits in usage.prompt_tokens_details.cached_tokens. There is no cache-write charge: tokens that miss bill at the normal input rate, tokens that hit bill at the cached rate, and storage of the cached content is free for now on both platforms.

Z.ai's own documents disagree on how large the discount is. The pricing table puts a GLM-5.2 or GLM-5.3 hit at $0.26 against $1.40, which is 18.6% of input. The Z.ai help FAQ says hits are "calculated at approximately 1/5 of the original price", which matches the table. The caching guide's billing section says cache hit tokens are billed "usually 50% of standard price". Budget on the table, which is the number your invoice uses, and treat the 50% line as out of date. The same FAQ adds that caching "is currently undergoing open beta testing" and that the rules for "how cache hits are triggered, how long cached content is retained" have "not yet been announced".

That last point is the practical difference from Claude or Kimi. You cannot mark a breakpoint, choose a TTL or guarantee a hit. Zhipu's Chinese caching guide gives the most concrete advice either platform publishes: the repeated prefix should be long, with 500 tokens or more recommended, and the cache takes effect asynchronously, so a second request sent immediately after the first may not hit. Keep the system prompt and tool definitions first and byte-identical, keep timestamps out of them, and read cached_tokens on real traffic before you put a hit rate into a budget.

Reasoning: GLM-5.3 always thinks, GLM-5.2 can stop

Reasoning is where GLM-5.2 and GLM-5.3 part company on cost. Z.ai's deep thinking guide says GLM-5.3 and GLM-5.3-Flash "no longer support disabling thinking", and a request with thinking.type: "disabled" returns an error. Their reasoning_effort accepts low, high and max, and max is the default. GLM-5.2 instead uses dynamic thinking by default, where the model decides whether to think, and it still accepts thinking.type: "disabled"; its reasoning_effort defaults to max, and the values none or minimal stop it thinking.

Z.ai's migration guide spells out the consequence for anyone moving up: an application that sends thinking.type: "disabled" must change it to enabled and set reasoning_effort to low before switching the model ID to glm-5.3, "otherwise, the request will fail". Z.ai's guide also warns that "the thinking process will consume extra tokens". In the example below we bill reasoning at the output rate, which is how it is billed on every other reasoning API we have priced.

A second setting, clear_thinking, decides whether reasoning is paid for once or many times. Z.ai's thinking mode guide describes "preserved thinking", where reasoning from earlier assistant turns stays in the context. It is "disabled by default on the standard API endpoint" and enabled by default on the Coding Plan endpoint, and you turn it on for the API with "clear_thinking": false, sending the full reasoning_content back unmodified. Left at the API default, old reasoning drops out of the next request, so each turn's thinking is paid for once, as output. Turn preservation on and it is paid again as input on every later turn.

Run the conversation below with 1,000 reasoning tokens on every reply to see the size of it. On GLM-5.3 at the API default, the 40-turn bill rises from $0.728 to $0.904 without caching and from $0.210 to $0.386 with it. With preserved thinking switched on, it rises to $2.00 and $0.633. The 1,000-token figure is an illustration, not a measurement, and at the default max effort real traces can be much longer. On GLM-5.2 you can remove the reasoning line entirely for turns that do not need it, which GLM-5.3 does not allow.

Free models and how you pay

Z.ai has no free tier for its paid models, but it lists three models at $0 on every line: GLM-4.7-Flash, GLM-4.5-Flash and the vision model GLM-4.6V-Flash. They are the right place to build a prototype, and none of them is a GLM-5 model. bigmodel.cn lists GLM-4.7-Flash, GLM-4.6V-Flash and several older Flash models as free. Neither pricing page lists a sign-up credit for the paid models.

Paid API calls draw on an account balance that you top up from the billing page. Two details from Z.ai's FAQ catch people out: billing runs a day behind, with the billing history showing the previous day's consumption and today's spend not yet visible, and credit-card top-ups do not support 3DS verification. With a one-day lag, a runaway agent can spend for a full day before the balance shows it, so cap max_tokens (GLM-5.3 allows up to 128K of output per call) rather than relying on the dashboard.

Rate limits are per account and per model. Z.ai points API users to their rate limits page in the console and publishes no table, and Zhipu's China rate-limit guide describes the mechanism: concurrency caps that differ by model and by account level, a 1302 error when your account hits its cap and a 1305 error when the model itself is overloaded, and a form to request more concurrency that Zhipu says it reviews within 10 working days.

The GLM Coding Plan, and why an API product cannot use it

Most of the noise around GLM pricing is about the GLM Coding Plan, a subscription that Z.ai's overview says starts "at just 18 USD per month", with Pro and Max tiers above it. The Pro and Max prices sit on Z.ai's subscription page, which renders them in the browser, and we did not read them from a page today, so we do not quote them; several incumbent guides quote launch-week promotions that have since changed.

For anyone building a product, the plan is not an alternative to the API. Z.ai's usage policy says it "may only be used within officially supported tools and products", meaning coding agents such as Claude Code, Cline, OpenCode and Z.ai's own ZCode, and that sharing an account is prohibited. Since 16 October 2026, according to Z.ai's API key guide, Standard API keys and Coding Plan keys are created separately and "the key type determines how calls are billed": a Standard key used inside a coding tool bills your balance even if you hold a plan. The plan also cannot pin GLM-5.2: requests for GLM-5.2 or GLM-5.1 "will be automatically routed to GLM-5.3", and GLM-4.7 requests go to GLM-5.3-Flash. If you need GLM-5.2 itself, for a reproducible evaluation or a compliance record that names the model, the pay-as-you-go API is the only way to get it from Z.ai.

For a developer who does use it inside a coding agent, the plan moved to credits on 30 July 2026. Each tier has a 5-hour and a weekly allowance, and a call consumes (input tokens times 6.9, plus cached tokens times 1.7, plus output tokens times 24) divided by 10,000 credits on GLM-5.3, or 2.3, 0.56 and 8 on GLM-5.3-Flash. Calls outside peak hours, which are Monday to Friday 14:00 to 18:00 Singapore time, cost half, as do weekends.

Source: Z.ai's Coding Plan overview, read 29 October 2026. The low end of each range assumes all usage at peak, the high end all off-peak.

Those multipliers explain why a coding agent fits the plan so well. Output costs 24 credits per 10,000 tokens and a cached input token costs 1.7, so a coding session that re-reads the same files with a 95% hit rate is dominated by cheap cached input, which is the case Z.ai's table models. Z.ai's claim that the plan saves "up to 92%" against pay-as-you-go GLM-5.3 assumes that profile with "fully utilizing the off-peak discounts", so a team that codes mostly in Singapore afternoons on weekdays will see less. Purchases are non-refundable and renew automatically unless cancelled at least three days before the billing date.

What a 40-turn conversation costs on GLM, Kimi, Qwen, DeepSeek, Groq, Gemini and Claude

Take a chat assistant with a 2,000-token system prompt, where each user message is 150 tokens and each reply is 350 tokens; these are the same assumptions as our Kimi K3, Qwen, Groq and DeepSeek pricing guides, so the figures line up. Turn 1 sends 2,150 input tokens and turn 40 sends 21,650, because the API is stateless and every request carries the exchanges before it. The whole 40-turn conversation consumes 476,000 input tokens to produce 14,000 output tokens, and a 100-turn conversation consumes 2,690,000 input tokens and 35,000 output tokens. Reasoning tokens are left out of the table, which flatters GLM-5.3 and GLM-5.3-Flash because they cannot turn reasoning off.

GLM rows are our arithmetic on rates read 29 October 2026. The other rows are the figures published in our Kimi K3, Qwen, DeepSeek, Mistral, Groq, Gemini and Claude guides on the dates shown, where the method for each is set out. In renminbi on bigmodel.cn, the same 40-turn conversation costs ¥4.20 uncached and ¥1.47 cached on GLM-5.3 or GLM-5.2, and ¥0.42 and ¥0.16 on GLM-5.3-Flash.

"Cache warm" on GLM means that from turn 2, the previous request's prompt is read at the cached rate and the 500 new tokens are billed at the normal input rate, with no write charge, because that is how Z.ai's implicit cache bills. It is a best case: Z.ai does not guarantee hits or publish a retention period, so a real conversation will hit somewhat less. The DeepInfra GLM-5.2 row includes its 5-minute cache write at 1.25 times input; DeepInfra's GLM-5.3 page did not list cache-write terms on 29 October, so we left its warm figures out.

Read across, GLM-5.3 lists at $1.40 and $4.40 against $2.00 and $10.00 for both Kimi K3.1 Pro and Claude Sonnet 5.5, so before reasoning it costs a third less than either for the same conversation, $0.728 against $1.09. With the cache warm the gap narrows to $0.210 against $0.240 for Sonnet 5.5, because a Claude cache hit costs 10% of input and a GLM hit costs 18.6%. Against the low-cost models, GLM-5.3-Flash lands where the field has settled for this conversation, $0.078 uncached next to $0.078 for Qwen3.8-Flash and $0.080 for DeepSeek, Groq and Mistral Small 4, but it pays more for cached tokens than DeepSeek does and it always reasons.

At 10,000 forty-turn conversations a month, GLM-5.3 or GLM-5.2 costs $7,280 without caching and $2,100 with it, and GLM-5.3-Flash costs $784 and $239. Add 1,000 reasoning tokens per reply on GLM-5.3 at the API default and the cached figure rises to $3,860.

Going from 40 to 100 turns multiplies turns by 2.5 and the uncached GLM-5.3 bill by 5.4, because every turn carries every turn before it. Caching cuts the price of the repeated tokens; it does not stop them being sent, and a cache whose lifetime Z.ai has not even published does nothing for a user who comes back tomorrow. The other lever is to stop resending the whole transcript. We build Maximem Synap, a memory layer that returns a ranked set of facts about the user, trimmed to a token budget you set, in place of the full history. When that beats caching, and when it does not, is worked through in how to reduce LLM token costs in long conversations.

What third-party hosts charge for open-weight GLM

GLM-5.2, GLM-5.3 and GLM-5.3-Flash all have open weights on Hugging Face, so other inference providers serve them at their own prices. On 29 October OpenRouter's endpoint data listed 49 providers for GLM-5.3 and 34 for GLM-5.2.

Prices per million tokens in US dollars, read from each provider's page on 29 October 2026; OpenRouter figures from its public endpoints API on the same day.

Together and Fireworks match Z.ai's list price, so the reason to use them is operational: an existing contract, consolidated billing, or LoRA fine-tuning on Fireworks. DeepInfra is the one that changes the bill. Its promotional GLM-5.3 rate is 60% below Z.ai on input and 43% below on output, which brings the uncached 40-turn conversation from $0.728 to $0.303, but it serves 4-bit weights, it is labelled promotional, and its GLM-5.2 page adds a cache-write premium that Z.ai does not charge. Price the standard rate ($0.90 and $4.00 on GLM-5.3) into anything that has to survive the promotion.

OpenRouter's cheapest numbers need the same caution as with every open-weight model. The lowest GLM-5.3 input, about $0.07, comes from endpoints that charge $1.40 to $4.38 for output, and many of the cheapest listings serve fp4 or nvfp4 quantized weights according to the endpoint data, which is a different trade from Z.ai's fp8 endpoint. A cached prefix also only pays off when consecutive requests land on the same provider, so if you route through OpenRouter, pin a conversation to one provider.

The GLM-5.3 license is not MIT

Guides that describe GLM as MIT-licensed are right about GLM-5.2 and GLM-5.3-Flash: both carry the MIT license on Hugging Face. GLM-5.3 does not. Its license is a custom "GLM-5.3 License" that grants the usual permissions to use, modify, distribute and sell, with one added condition: a licensee operating a "Model as a Service" business whose group revenue exceeds 10 billion US dollars in any 12 months "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose". The definition excludes end-user products with the model embedded in specific features, and plain relaying of requests to a model someone else hosts.

For nearly every company reading a pricing page this changes nothing. It matters to a large cloud deciding whether to host GLM-5.3, and it is one reason the GLM-5.2 weights, which are MIT and share GLM-5.3's base model, will stay in circulation.

Data use, the contracting entity and the law

Z.ai's Terms of Use, last updated 14 April 2026, are an agreement with JINGSHENG HENGXING TECHNOLOGY PTE.LTD, governed by Singapore law with arbitration at the Singapore International Arbitration Centre. The terms treat individual users and API users differently. For individual users, Z.ai reserves "the right to process any User Content to improve our existing Services". For enterprises and developers using the API, "we will not use your User Content for developing or improving Services unless you explicitly agree to such use". The terms also prohibit using GLM outputs to train models that compete with Z.ai's.

If no data may leave your infrastructure, the open weights are the answer, and DeepInfra marks its GLM-5.3 endpoint "zero retention" for teams that want a hosted route without a separate agreement.

Which GLM model to pick on cost

For a new product on Z.ai's API, start from what the workload needs from reasoning. If every turn benefits from thinking, use GLM-5.3: it is the current flagship, it costs the same as GLM-5.2, and Z.ai reports large coding gains over it, which you should verify on your own tasks. If many turns are short answers, routing or extraction, GLM-5.2 is the cheaper flagship in practice, because it is the only model in the family at that price that lets you switch thinking off. If cost leads the decision, price GLM-5.3-Flash first; at $0.15 and $0.50 it is about a tenth of the flagship on every line, and it takes images and video.

Off Z.ai, Together and Fireworks are list-price second suppliers, and DeepInfra's promotional rate is the cheapest well-known host for GLM-5.3 today, with quantized weights and a promotion that will end. If you are a developer looking for cheap GLM inside a coding agent, the Coding Plan is the better deal; if you are building a product, it is not available to you.

How to estimate your monthly GLM bill, and cut it

A GLM request costs (cache-hit tokens times the cached rate) plus (uncached input tokens times the input rate) plus (output and reasoning tokens times the output rate). For a conversation, the input on each turn is the system prompt plus every earlier turn, and also every earlier turn's reasoning if you set clear_thinking to false, so estimate per conversation as in the table above and multiply by conversations per month. Add web search at $0.01 per call, and remember that the balance shown in the console is a day behind.

The levers that move a GLM bill, roughly in order of size:

1. Use GLM-5.3-Flash where it is good enough. It is about a tenth of the flagship on every token.
2. Control reasoning. Use reasoning_effort: "low" on GLM-5.3, or switch thinking off on GLM-5.2, for turns that do not need it.
3. Leave clear_thinking at its API default unless an agent loop measurably needs preserved reasoning.
4. Keep the prefix long and stable, with 500 tokens or more of unchanging system prompt and tools first, so the implicit cache can hit.
5. Cap max_tokens. A single 128K-token reply on GLM-5.3 costs $0.56 in output alone.
6. Stop resending the whole conversation once chats run long, by summarising or retrieving only what the next turn needs.

Frequently asked questions

How much does GLM 5.2 cost?

As of 29 October 2026, GLM-5.2 costs $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens on Z.ai's API, with cache storage free for a limited time. The price is flat across its 1-million-token context window. On bigmodel.cn it costs ¥8 input, ¥2 cache hit and ¥28 output per million tokens.

Is GLM-5.3 more expensive than GLM-5.2?

GLM-5.3 costs the same as GLM-5.2 on Z.ai: $1.40 input, $0.26 cached and $4.40 output per million tokens. GLM-5.3 always reasons and cannot disable thinking, while GLM-5.2 can, so on workloads that do not need reasoning GLM-5.2 can cost less in practice.

How much does GLM-5.3-Flash cost?

GLM-5.3-Flash costs $0.15 per million input tokens, $0.03 cached and $0.50 output on Z.ai, and GLM-5.3-FlashX, the faster variant, costs $0.37, $0.075 and $1.25. Both have a 1M-token context and accept image, video and file input.

Is there a free GLM API?

Z.ai lists GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash as free on input, cached input and output. The GLM-5 family is paid, and neither Z.ai's nor bigmodel.cn's pricing page lists a sign-up credit for it.

How does GLM context caching work and what does it cost?

GLM caching is automatic, with no write charge and free cache storage for now; a cache hit on GLM-5.2 or GLM-5.3 costs $0.26 per million tokens, 18.6% of the input price. Z.ai says the cache is in open beta and has not published its retention rules, and Zhipu recommends a repeated prefix of 500 tokens or more.

Can I use the GLM Coding Plan for my own API product?

Not for a product: the GLM Coding Plan can only be used inside officially supported coding tools, and since 16 October 2026 Coding Plan keys and Standard API keys are separate, with the key type deciding the billing. A product calling GLM needs a Standard key billed at the API rates. The plan also routes GLM-5.2 requests to GLM-5.3.

How much does the GLM Coding Plan cost?

Z.ai's documentation says the GLM Coding Plan starts at 18 US dollars a month for Lite, with Pro and Max above it. Lite gives 2,000 credits per 5 hours and 10,000 a week, Pro 12,000 and 60,000, and Max 28,000 and 140,000, with off-peak and weekend calls costing half.

Where is GLM cheapest to run?

Among well-known hosts on 29 October 2026, DeepInfra's promotional GLM-5.3 rate of $0.563 input and $2.50 output per million tokens was the lowest, on 4-bit weights; Together and Fireworks match Z.ai's $1.40 and $4.40. OpenRouter lists 49 GLM-5.3 providers, some cheaper on input but more expensive on output.

Is GLM-5.3 open source?

GLM-5.3's weights are open on Hugging Face under a custom GLM-5.3 License that is permissive except that Model-as-a-Service operators with over 10 billion US dollars of annual revenue must pass a Z.AI security review before commercial use. GLM-5.2 and GLM-5.3-Flash are MIT-licensed.

Does Z.ai train on my API data?

Z.ai's Terms of Use say that for enterprises and developers using the API it will not use your content to develop or improve its services unless you explicitly agree, while content from individual users may be used. The contract is with a Singapore company under Singapore law.

Is GLM cheaper than Kimi, Qwen or Claude?

Before reasoning tokens, a 40-turn conversation costs $0.728 on GLM-5.3 or GLM-5.2 without caching, against $1.09 on Kimi K3.1 Pro and Claude Sonnet 5.5 and $1.04 on Qwen3.8-Max. With caching, GLM costs $0.210 and Sonnet 5.5 $0.240. GLM-5.3-Flash costs $0.078, level with Qwen3.8-Flash.

Sources

Z.ai, Zhipu and host sources retrieved 29 October 2026; comparison figures from our guides as dated in the table.

• Z.ai: pricing, model overview, release notes, GLM-5.3, GLM-5.2, GLM-5.3-Flash, migrate to GLM-5.3, thinking mode, deep thinking, context caching, help FAQ, Coding Plan overview, Coding Plan usage policy, API key management, plan update notice, Terms of Use and subscription terms
• Zhipu (bigmodel.cn): API pricing in CNY, context caching, rate limits and Coding Plan overview
• Weights: GLM-5.3, GLM-5.2 and GLM-5.3-Flash on Hugging Face
• Hosts: Together AI pricing and GLM-5.3, Fireworks GLM-5.3, DeepInfra GLM-5.3 and GLM-5.2, OpenRouter GLM 5.2 and its public endpoints API, and Groq models
• Maximem: Kimi K3 API pricing, Qwen API pricing, Groq API pricing, DeepSeek API pricing, Mistral API pricing, Gemini API pricing and Claude API pricing

Worked examples are our own arithmetic on those rates.