SAME CARD AS GLM-5.21M CONTEXT128K MAX OUTPUTTEXT ONLYPOST-TRAINING GAINS
GLM-5.3 API Pricing
GLM-5.3 is priced. Z.ai's International rate card lists $1.4/M input, $4.4/M output and $0.26/M cached input — the identical card GLM-5.2 and GLM-5.1 carry. That consistency is not a coincidence: Z.ai states GLM-5.3 uses the same base model as GLM-5.2, with every gain coming from post-training. A new generation at no extra cost per token. Until August 16, 2026 this page showed TBA — the API was listed as coming soon and access ran through the GLM Coding Plan subscription; the rate card now lists it.
Input - per 1M tokens
$1.40/M
Source Z.AI flat
Output - per 1M tokens
$4.40/M
Context 1M flat
Cached input - per 1M tokens
$0.26/M
Storage limited-time free -81%
Effective - agentic blend
$0.78/M
92/8 split - 82% cache
§ 01 / TERMINAL
GLM-5.3 API cost calculator.
Calculator pre-loaded with the live GLM-5.3 rates. Because the card matches GLM-5.2 exactly, every figure here doubles as a GLM-5.2 estimate — the choice between them is about capability, not cost.
$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
GLM-5.3 cost on real workloads.
CODING AGENT
Repo-wide feature build
$0.100/task
LONG DOC
Reading 150-page specs
$0.119/doc
RAG
Knowledge base answer
$0.00812/query
ASSISTANT
Research planning turn
$0.017/turn
Calibrated · measured on the vendor's tokenizer · 2026-06-10 Auto-counts as you type
Counts use a chars-per-token calibration measured on the vendor's own published tokenizer (zai-org/GLM-5, 2026-06-10). English prose is typically within a few percent; code and non-Latin scripts tokenize heavier. For billing-exact counts use the vendor's count-tokens API.
Characters 420
Words 68
Tokens (estimated) 80 tokens
Cost as input · uncached $0.00011 USD
Cost as output · uncached $0.00035 USD
Cost as cached input $0.00002 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| GLM-5.3 Current | $1.40 cache $0.26 | $4.40 | $0.78 agentic 92/8 | 1M | Newest GLM flagship |
| GLM-5.2 | $1.40 cache $0.26 | $4.40 | $0.78 same | 1M | The generation underneath 5.3 |
| GLM-5.1 | $1.40 cache $0.26 | $4.40 | $0.78 same | 200K | Older GLM-5, narrower window |
| GLM-5 Turbo | $1.20 cache $0.24 | $4.00 | $0.70 deprecated · −10% | 200K | Delisted August 2026 - kept for reference |
| GLM-5 | $1.00 cache $0.20 | $3.20 | $0.572 −27% | 200K | Cheapest of the GLM-5 line |
| GLM-4.7 FlashX | $0.07 cache $0.01 | $0.40 | $0.0511 −15× | 200K | High-throughput classification |
By comparison
GLM-5.3 has published scores on 14 benchmarks we track; its best placing is #2 of 9 on Terminal-Bench 3.0. 1 of these scores comes from an independent leaderboard; the rest are the lab's own reports.
| Benchmark | Score | Rank | Leader | Source |
|---|---|---|---|---|
| Terminal-Bench 3.0 Agents and computer use | 28.3% as reported | #2 of 9 | DeepSeek V4.1 Flash · 30% | Vendor-reported Z.ai (Zhipu) |
| Terminal-Bench 2 Agents and computer use · Terminal-Bench 2.1 | 88.2% as reported | #5 of 17 | DeepSeek V4.1 Flash · 90.6% | Vendor-reported Z.ai (Zhipu) |
| PostTrainBench AI research | 39.8% as reported | #2 of 6 | Gemini 4 Argon · 45.3% | Vendor-reported Z.ai (Zhipu) |
| AutomationBench Agents and computer use · Main set (as reported) | 48.2% v1.0.6 | #8 of 22 | MiniMax M3.1 Flash · 62.8% | Vendor-reported Z.ai (Zhipu) |
| Terminal-Bench 4.0 Agents and computer use | 41.8% max effort, Claude Code harness, 330 trials, ±3.2 pts (95% CI) | #13 of 34 | Kimi K3.1 Pro · 65.7% | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| FrontierSWE Coding · Version not stated | 78.1% as reported | #2 of 5 | Kimi K3 · 81.2% | Vendor-reported Z.ai (Zhipu) |
| Humanity's Last Exam Expert knowledge · With tools | 62.5% with tools | #8 of 20 | Claude Fable 5.2 · 74.2% | Vendor-reported Z.ai (Zhipu) |
| Agents' Last Exam Agents and computer use · CLI split | 28.5% CLI split | #3 of 6 | Kimi K3.1 Pro · 44.2% | Vendor-reported Z.ai (Zhipu) |
| CyberGym Cybersecurity | 84.5% as reported | #4 of 8 | MiMo-V2.6-Flash · 95.1% | Vendor-reported Z.ai (Zhipu) |
| NL2Repo-Bench Coding | 58% as reported | #4 of 8 | DeepSeek V4.1 Flash · 68.2% | Vendor-reported Z.ai (Zhipu) |
| DeepSWE Coding | 66.9% as reported | #14 of 27 | Gemini 4 Argon · 77.9% | Vendor-reported Z.ai (Zhipu) |
| Toolathlon Agents and computer use | 73% as reported | #5 of 9 | MiMo-V2.6-Pro · 76.9% | Vendor-reported Z.ai (Zhipu) |
| ExploitBench Cybersecurity | 54.4% as reported | #4 of 7 | GPT-6 Astra · 100% | Vendor-reported Z.ai (Zhipu) |
| ProgramBench Coding | 19% almost-solved rate | #5 of 7 | Kimi K3.1 Pro · 31.5% | Vendor-reported Z.ai (Zhipu) |
GLM-5.3 API pricing FAQ.
GLM-5.3 pricing questions, including what changed when the API opened.
Q · 01 How much does GLM-5.3 cost? +
$1.4/M input, $4.4/M output and $0.26/M on a cache hit, on Z.ai's International rate card. Under AI//COST's 92/8 agentic blend with 82% cache hits, the effective planning figure is $0.78/M.Q · 02 Is it more expensive than GLM-5.2? +
No — the two cards are identical, down to the cached rate. Z.ai states GLM-5.3 uses the same base model as GLM-5.2 and that the improvements come from post-training, so there is no price step between the generations. On cost alone there is no reason to stay on 5.2.
Q · 03 What changed since this page said TBA? +
The API opened. On August 16, 2026 GLM-5.3 had a model page marked New and shipped to GLM Coding Plan subscribers, while its own docs said "The GLM-5.3 API is coming soon" and neither the International nor the China rate card listed it — so this page showed TBA rather than guessing. As of August 23 the International card lists it.
Q · 04 How much better is it than GLM-5.2, really? +
Z.ai claims +50% over 5.2 on its own Z.ai Code Bench, and reports Terminal-Bench 3.0 rising from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Agents' Last Exam from 23.8 to 28.5. Every one of those is a vendor-run figure on a vendor page — we carry them because they are what exists, not because they are neutral. No third-party evaluation of GLM-5.3 was available at the time of writing.
Q · 05 What is the cached-input storage charge? +
Z.ai marks it "Limited-time Free" across the whole GLM line, GLM-5.3 included. It is a separate charge from the per-token rates — you pay for keeping a cache alive, not for reading it — and we do not model it, because it is free today and Z.ai publishes no rate for when it stops being free.
Q · 06 What are the model's limits? +
1M context and 128K maximum output, text in and text out. Thinking mode, streaming, function calling, context caching, structured output and MCP are all supported per Z.ai's model page.