What is GLM-5.3?
GLM-5.3 is Z.ai's current flagship. It keeps GLM-5.2's base model and adds more training on long, realistic engineering tasks, which Z.ai says lifts coding by 50% on its own benchmark and puts it top of the CyberGym vulnerability test. It reads and writes text only, holds 1M tokens of context and always reasons, at low, high or max effort.
You can call it on Z.ai's API, use it in Claude Code or OpenCode through the GLM Coding Plan, chat with it at chat.z.ai, or run the weights yourself. Its licence is free for commercial use, but unlike GLM-5.2's MIT licence it asks the very largest model hosts to pass a Z.ai security review first.
What changed from GLM-5.2?
Same base model, more training: it codes and finds vulnerabilities far better, but it always thinks and moves to Z.ai's own licence.
- Coding
- Terminal-Bench 3.0 rose from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9 in Z.ai's tests.
- Security
- ExploitBench more than doubled, from 24.4 to 54.4, and CyberGym rose from 77.2 to 84.5.
- Thinking
- Always on; you can no longer turn it off, and max effort is the default.
- Licence
- Z.ai's GLM-5.3 License replaces MIT and adds a security review for the largest model hosts.
- Price
- Unchanged at $1.4 / $4.4 per million tokens on the API.
Who is GLM-5.3 best for?
GLM-5.3 is best for long coding agent runs, security code review, teams that want to self-host open weights, and developers who want flat-rate coding in Claude Code or OpenCode.
Long coding agent runs
It holds 1M tokens of context and, in Z.ai's tests, finishes more agentic coding tasks than GLM-5.2 while writing fewer output tokens per task.
Security code review
Z.ai trained it on vulnerability discovery, it found 2,436 real vulnerabilities across 269 projects with partner security teams, and users say it takes on security work closed models refuse.
Teams that want to self-host
The weights are on Hugging Face, free for commercial use, with guides for vLLM, SGLang, Transformers and KTransformers.
Flat-rate coding in Claude Code
The GLM Coding Plan swaps per-token billing for a weekly credit allowance from $18 a month, and calls outside weekday peak hours count half.
What's included with GLM-5.3?
Three effort levels, OpenAI and Anthropic compatible endpoints, tool calling with streamed arguments, cheap cached input, Coding Plan MCP tools and downloadable weights.
- Effort levels
- Low, high and max reasoning settings; max is the default and the one Z.ai recommends for coding.
- OpenAI and Anthropic formats
- Z.ai serves it over OpenAI Chat Completions, OpenAI Responses and Anthropic Messages endpoints, so existing SDKs and Claude Code connect with a new base URL.
- Tool calling
- Function calling with streamed tool-call arguments, plus structured JSON output for wiring it into other systems.
- Context caching
- Cached input costs $0.26 per million tokens instead of $1.4, which matters in long agent loops.
- Coding Plan MCP tools
- Plan subscribers get Z.ai's web search, web reader, GitHub repository reader and image and video analysis tools, so the text-only model can still work from screenshots.
- Open weights
- Downloadable from Hugging Face, with support for SGLang, vLLM, Transformers, KTransformers and Unsloth, plus Ascend NPUs.
How much does GLM-5.3 cost?
GLM-5.3 costs $1.4 per million input tokens and $4.4 per million output tokens on Z.ai's API, or comes with the GLM Coding Plan from $18 a month, and its open weights are free to download.
- Cached input $0.26 per million tokens
- 1M-token context, up to 128K tokens out
- Text input only, with reasoning always on
- 10,000 credits a week, up to 2,000 per 5 hours
- About 48M to 97M GLM-5.3 tokens a week at a 95% cache hit rate
- Works in Claude Code, Cline, OpenCode and 20+ agent tools
- 6 times Lite usage: 60,000 credits a week
- Curated MCP tools and faster generation
- Priority access to new flagship models
- 14 times Lite usage: 140,000 credits a week
- Dedicated resources at peak times
- First access to new flagship models
Prices are in US dollars. The GLM Coding Plan is 20% cheaper billed quarterly (Lite $14.4, Pro $64, Max $134.4 a month) and 30% cheaper billed yearly ($12.6, $56 and $117.6 a month). Plan credits come from token use: GLM-5.3 uses about 3 times the credits of GLM-5.3-Flash for the same tokens, and calls outside weekday peak hours (14:00 to 18:00 Singapore time) use half. Team seats cost $88 (Standard) or $188 (Premium) per seat per month. Subscriptions are not refundable. Full pricing on z.ai. We checked these prices on .
How do readers rate GLM-5.3?
No one has rated GLM-5.3 yet. Be the first.
Sign in to rate. One rating per account.
GLM-5.3 pros and cons
It codes close to the closed frontier and takes on security work others refuse, but it reads text only, thinks at length, and Z.ai's service strains at peak times.
Pros
- Codes close to the closed frontier: in Z.ai's table it scores 88.2 on Terminal-Bench 2.1, level with Fable 5 (88.0) and GPT-5.6 Sol (88.8), and developers who have tried it on real work say the gap is small.
- Takes on security research and bug triage that closed models often refuse, several developers say, and it tops Z.ai's CyberGym table at 84.5%.
- Shows its full reasoning, which users say makes it easy to spot and stop a run that is going wrong.
- Costs $1.4 / $4.4 per million tokens, about a third of Claude Opus 5.5's input price and under a quarter of its output price, and the weights are free to self-host.
Cons
- Text only: it cannot read screenshots or images, which some web developers call a deal breaker; Z.ai points to GLM-5.3-Flash or its vision tool on the Coding Plan instead.
- Thinks at length: max effort is the default, and users report more tokens per task than closed models and higher bills than cheaper open models such as DeepSeek's Flash for the same work.
- Z.ai's own service strains at busy times: Coding Plan users report hitting the 5-hour limit fast in weekday peak hours, plus slow or failed requests on some days.
- Still behind the best closed models on the hardest work: Z.ai's table has it well below Fable 5 on ProgramBench and ExploitBench, and some users say it drifts off task or needs more hand-holding.
What are the best GLM-5.3 alternatives?
Two tools on Recaply do GLM-5.3's job. DeepSeek V4.1 Flash is open source, so you can run it on your own server.
DeepSeek V4.1 FlashOpen SourceChoose it when cost per task matters most; it charges $0.30 / $1.20 per million tokens at peak and half that off-peak, reads images, lets you turn thinking off and is MIT licensed.
Claude Opus 5.5PaidChoose it when you want a closed model with image input on Amazon Bedrock, Google Cloud or Microsoft Foundry, and can pay $4 / $20 per million tokens.
Questions about GLM-5.3
What is GLM-5.3?
GLM-5.3 is Z.ai's flagship open-weights model for long coding and agent work, built on GLM-5.2's base with more training and sold at $1.4 / $4.4 per million tokens.
Who is GLM-5.3 for?
GLM-5.3 is best for long coding agent runs, security code review, teams that want to self-host open weights, and developers who want flat-rate coding in Claude Code or OpenCode.
Is GLM-5.3 free?
Partly. GLM-5.3's weights are free to download from Hugging Face under Z.ai's own licence, which allows commercial use, and you can chat with it free at chat.z.ai after signing up. Using it through Z.ai's API or in coding tools costs money: $1.4 / $4.4 per million tokens, or a Coding Plan from $18 a month.
How much does GLM-5.3 cost?
GLM-5.3's paid plans are API at $1.4 / $4.4 per million input / output tokens, Coding Plan Lite at $18 per month, Coding Plan Pro at $80 per month and Coding Plan Max at $168 per month.
Is there an open source alternative to GLM-5.3?
Yes. DeepSeek V4.1 Flash is an open source tool that does the same job, so you can run your own.
What are the best GLM-5.3 alternatives?
The closest alternatives on Recaply are DeepSeek V4.1 Flash and Claude Opus 5.5.
Where can I try GLM-5.3?
On its official website, z.ai.
Should I use GLM-5.3 or GLM-5.3-Flash?
Use GLM-5.3 for the hardest coding and security work, and GLM-5.3-Flash for everyday tasks, images or a smaller bill. Flash costs about a ninth as much per token on the API, reads images and video, and uses about a third of the Coding Plan credits.
How do I use GLM-5.3 in Claude Code?
Point Claude Code at Z.ai's Anthropic-compatible endpoint, https://api.z.ai/api/anthropic, and use a Z.ai API key or a GLM Coding Plan key. Coding Plan keys can call only GLM-5.3 and GLM-5.3-Flash.
Does Z.ai use my GLM-5.3 prompts for training?
Not on the API: Z.ai's terms say it will not use API customers' content to improve its services unless they agree, and its data terms say API inputs are not stored. In the chat app, individual users give Z.ai the right to use their content. Data is processed in Singapore.
Can I get a refund on the GLM Coding Plan?
No. Z.ai says GLM Coding Plan subscriptions are non-refundable once bought, even if you have not used your quota. You can cancel on the subscription page at least 24 hours before the next billing date, and the plan stays active until it expires.
Our verdict
GLM-5.3 is the open-weights model to try first for long coding and security work.
Z.ai's tests put it level with the closed frontier on several coding benchmarks at $1.4 / $4.4 per million tokens, and developers who have used it on real work largely agree.
- Skip it for
- Work that needs image input, or strict limits on tokens per task.
- Check first
- On the Coding Plan, test your peak-hour usage against the 5-hour and weekly credit limits before paying for a year.
GPT-5.3-Codex
Rnj-1
Kimi K2.7 Code
KAT-Coder-Pro V2
GLM-4.7
MiMo-V2.5-Pro