GLM API Pricing: $0.15 In, $0.50 Out per 1M Tokens

What you pay, in one table

The glm api pricing here has no tiers, seats or monthly plans. You buy credit, and every request deducts from it according to the tokens it used.

A short note on naming: the domain matches what people search for, but the service is an independent, OpenAI-compatible endpoint for a single open-weight text model, uncensored. It is not the official API of the vendor behind GLM, so these rates are not that vendor's rates. If you are comparing, look up their current pricing on their own site and use the method further down.

Turning rates into costs

Per-million prices are awkward to feel. A better unit is cost per request. For any request, multiply input tokens by $0.15, output tokens by $0.50, and divide each by 1,000,000. Four workloads show how it plays out.

A retrieval app with long context

A question-answering app stuffs 50,000 tokens of retrieved text into the prompt and gets back 800 tokens. Input is 50,000 x $0.15 / 1M = $0.0075. Output is 800 x $0.50 / 1M = $0.0004. One answer costs $0.0079, so 1,000 answers cost $7.90. The input side dominates, so shrinking retrieved context pays more than shortening answers.

A chat product

Each message sends about 3,000 tokens of history and persona, and the reply is 400 tokens. That is $0.00045 + $0.0002 = $0.00065 per message. A user who sends 40 messages a day costs $0.026 per day, which is $0.78 over 30 days. You can use that figure to set your own pricing with a sensible margin.

A batch job with structured output

You need to classify 200,000 short documents. Each prompt is 600 tokens, and JSON mode keeps each reply to about 20 tokens. Input is 200,000 x 600 = 120M tokens, which costs 120 x $0.15 = $18.00. Output is 200,000 x 20 = 4M tokens, which costs 4 x $0.50 = $2.00. The whole job is $20.00, exactly one entry-level top-up. Remember the rate limit when you plan wall-clock time: a full key allows 300 requests per minute, so 200,000 requests takes about 667 minutes at the cap.

A tool-calling agent

An agent makes 8 calls per task, each with 5,000 tokens in and 300 out. Per task that is 40,000 in ($0.006) and 2,400 out ($0.0012), so $0.0072. A hundred tasks cost $0.72. Agents are the workload where history growth bites hardest, since every step re-sends earlier tool results, so log real token counts after a few dozen tasks and replace these estimates with your own averages.

Top-ups and bonuses

You choose any whole dollar amount from $20 to $500. Bigger top-ups add a percentage on top of the credit you receive:

The last column divides the rate by the credit-to-payment ratio. At $100 you receive $110, so you pay $100 / $110 = 0.909 of list price: $0.15 becomes $0.136 and $0.50 becomes $0.455. The 15% bonus starts at $200, so there is no extra discount for going from $200 to $500. Pick the amount that matches how fast you spend, not the biggest one available.

Payment is by crypto (USDT on TRC20, USDC on Base). Because credit never expires, topping up $200 once for the bonus is safe if you expect to use it, and a smaller top-up is just as safe if you prefer to start low. Setup steps for either are on the quickstart page.

One practical rule for choosing: take your expected monthly spend from the scenarios above and top up about one to two months of it. For the support-style chat product that costs $0.78 per active user per month, 100 active users come to $78, so a $100 top-up (which adds $110 of credit) covers the first month with room to spare. Top up again when the balance header shows a week of runway left.

The free trial

New accounts can claim $0.50 of trial credit with one click on the API key page after sign-up. Sign-up is with Google or with email and password. The credit needs no card, is valid for 7 days, and is limited to one per person.

That amount is about 3.33M input tokens or 1M output tokens. In request terms, a 2,000-in, 500-out chat call costs $0.00055, so the trial covers roughly 900 of them. It is enough to test streaming, tool calling and JSON mode on your own prompts. The trial key runs under lower limits than a paid key, 2 concurrent requests and 60 per minute, which is fine for manual testing.

Start on the API key page. If you want to see all request fields before you spend anything, the quickstart lists them.

Watching your balance

Every response includes an X-Balance-USD header with your remaining credit, plus X-Request-Id for support questions. Streamed responses end with a usage block, so you can compute the exact cost of each call. A small logger is all you need:

Copyfrom openai import OpenAI

client = OpenAI(base_url="https://api.glmapi.cc/v1", api_key="YOUR_KEY")

raw = client.chat.completions.with_raw_response.create(
    model="uncensored",
    messages=[{"role": "user", "content": "Summarise the plot of a heist film."}],
)
r = raw.parse()
cost = r.usage.prompt_tokens * 0.15 / 1e6 + r.usage.completion_tokens * 0.50 / 1e6
print(round(cost, 6), raw.headers.get("X-Balance-USD"))

Set an alert on the balance header in your own monitoring, and top up with a margin ahead of any large batch.

Comparing with another provider's rates

Rates alone do not decide a comparison, because token counts differ between model families and so does verbosity. Take the same 100 prompts, run them through both services, and record real input and output tokens. Then apply each provider's two rates. Whatever the list prices say, only that measurement tells you your bill.

Beyond price, check four things on the other side: whether failed or refused requests are billed, whether credit expires, what the default rate limits are, and whether the model declines lawful requests that you need answered. Refusals are the hidden line item, since they cost retries and prompt rewrites. Our model is tuned to answer without content refusals for lawful adult use. It does keep one hard limit: no sexual content involving minors.

It is also fair to say when this is not the right tool. It is text only, with no images, audio, embeddings or fine-tuning, and one model rather than a menu. If you need those, check their current pricing elsewhere. The privacy-relevant fact is that prompts are not used for training.

Finally, budget for the limits as well as the price. A full key allows 8 concurrent requests and 300 per minute, a trial key 2 and 60. With 8 requests in flight and a reply that takes 10 seconds, you finish about 48 replies per minute, well under the per-minute cap, so concurrency is usually the tighter limit for long generations while short replies hit the request cap first. Streamed requests hold their slot until the stream ends, which matters if your client reads slowly. If you need more headroom, ask on the support page instead of guessing.

More guides

• GLM API Base URL: Connect Any OpenAI Client
• GLM API Python: Chat, Streaming and Tools Examples
• Gemma API: Myths vs Facts