Last updated: November 4, 2026
GLM-5.3-FlashX-Legal is Z.ai’s legal-domain model, built on GLM-5.3-FlashX with extra legal training. It reads text, images, video and files, handles a 1M-token context, costs $0.37 input and $1.25 output per million tokens, and became available on OpenRouter on November 2, 2026.
Contents
- What GLM-5.3-FlashX-Legal is
- Legal vs FlashX vs Flash
- Where to use it and what it costs
- How to call it through the API
- Benchmarks Z.ai reported
- Legal workflows and prompt patterns
- Limitations
- Frequently asked questions
Z.ai describes the model on its GLM-5.3-Flash documentation page, which covers the whole Flash family. AutoClaw, Z.ai’s agent app, added the model on October 29, 2026, and OpenRouter listed it as z-ai/glm-5.3-flashx-legal on November 2. This guide covers what it does, how to call it and where its published evidence stops.
What GLM-5.3-FlashX-Legal is
| Item | Detail |
|---|---|
| Base model | GLM-5.3-FlashX, the faster tier of GLM-5.3-Flash (320B total, 18B active parameters, hybrid sparse and linear attention) |
| Legal training | Legal-domain mid-training and post-training, plus reinforcement learning in tens of thousands of environments modeled on legal workflows |
| Inputs and output | Video, image, text and files in; text out |
| Context | 1M tokens, up to 128K output tokens |
| Speed | Z.ai lists FlashX at 200 tokens per second |
| API features | Function calling, context caching, structured output and streaming |
| Target tasks | Document cross-checking, legal issue analysis, contract review and negotiation support, drafting, and tool integration |
Legal vs FlashX vs Flash
| Aspect | GLM-5.3-Flash | GLM-5.3-FlashX | GLM-5.3-FlashX-Legal |
|---|---|---|---|
| API model code | glm-5.3-flash | glm-5.3-flashx | glm-5.3-flashx-legal |
| Input / output price per 1M tokens | $0.15 / $0.50 | $0.37 / $1.25 | $0.37 / $1.25 |
| Cached input per 1M tokens | $0.03 | $0.075 | $0.075 |
| Legal-specific training | No | No | Yes |
| Open weights on Hugging Face | Yes | No listing found | No listing found |
| GLM Coding Plan | See Z.ai plan terms | Not included yet | Not included yet |
Prices come from Z.ai’s pricing page, where the Legal variant costs the same as standard FlashX. For comparison, the full GLM-5.3 model is listed at $1.40 input and $4.40 output. The base GLM-5.3-Flash weights are on Hugging Face, so you can self-host the general model, but not the legal tune.
Where to use it and what it costs
- Z.ai API: call the Chat Completion API with the model code
glm-5.3-flashx-legal, billed per token. - OpenRouter:
z-ai/glm-5.3-flashx-legalat $0.37 input, $1.25 output and $0.075 cache read per million tokens, with Z.ai as the only provider. - AutoClaw: Z.ai’s agent app, where the AutoClaw Legal Assistant adds Word and WPS plugins for tracked changes and comments.
- Cache storage: Z.ai lists cached input storage as free for a limited time.
OpenRouter shows a context length of 1,048,576 tokens and a maximum completion of 131,072 tokens. The model is not offered through OpenRouter’s EU in-region routing, which matters if your firm has data residency rules.
How to call it through the API
- Create an API key in the Z.ai developer console, or use an OpenRouter key.
- Send a POST request to
https://api.z.ai/api/paas/v4/chat/completionswith the headerAuthorization: Bearer YOUR_API_KEY. On OpenRouter, use its OpenAI-compatible endpoint instead. - Set
"model": "glm-5.3-flashx-legal"(orz-ai/glm-5.3-flashx-legalon OpenRouter) and pass yourmessages. Put scanned exhibits or screenshots inimage_urlcontent blocks. - Use Z.ai’s recommended settings:
temperature1,top_p0.95,reasoning_effortmax,thinking.typeenabled andthinking.clear_thinkingfalse. - For agent work, send
toolsand setstreamandtool_streamto true so tool calls stream as they form. - For checklists or clause tables, request JSON with a schema in
response_format, which OpenRouter lists as supported.
Thinking cannot be turned off on this model family, so expect reasoning tokens on every call and budget output costs with that in mind.
Benchmarks Z.ai reported
- Harvey LAB: 19.2% all-criteria pass rate, 12.5 percentage points above the base model. Z.ai says this was the best result among open-source models in the evaluation.
- LegalBench and PLawBench: Z.ai says the model matched frontier models, without publishing the scores on its AutoClaw post.
A 19.2% pass rate means most Harvey LAB tasks still fail at least one criterion. Treat these as vendor-reported numbers and check results on your own documents before relying on them.
Legal workflows and prompt patterns
Z.ai recommends running the model inside an agent setup with MCP services (Model Context Protocol tool servers) for statutes, case law, company information, and Word drafting and revision.
- Contract redlines. “Review this supply agreement from the buyer’s side. List each risky clause, propose replacement text and explain the change in one sentence.” In AutoClaw the edits land in Word as tracked changes and comments.
- Evidence cross-checks. Load records into the 1M-token window and ask it to map contradictions and link each finding to its source document and page.
- Compliance checklists. Ask it to turn requirements across markets into a prioritized checklist with a legal reference for each item.
- Issue spotting. Give the facts and jurisdiction, then ask for issues, governing rules and open questions as separate lists.
For other large models released this autumn, see our Tencent Hy4 guide, LongCat-2.5 guide, Kimi K3.1 Pro guide and MiniMax M3.1 Flash guide. To run any of them in a coding agent, start with our Claude Code guide.
Limitations
- It is not a lawyer. Verify every citation and statute against an official source before filing or advising.
- Only Z.ai serves the model on OpenRouter, so there is no provider failover and no EU in-region routing.
- No legal-specific open weights are listed, so you cannot self-host this variant.
- Thinking is always on, which adds latency and output tokens.
- The FlashX models are not yet part of the GLM Coding Plan.
Frequently asked questions
What is GLM-5.3-FlashX-Legal?
GLM-5.3-FlashX-Legal is a legal-domain version of Z.ai’s GLM-5.3-FlashX model. Z.ai added legal mid-training, legal post-training and reinforcement learning in tens of thousands of task environments modeled on legal workflows such as contract review, drafting and legal research.
How much does the GLM Legal API cost?
Z.ai lists GLM-5.3-FlashX-Legal at $0.37 per million input tokens, $0.075 per million cached input tokens and $1.25 per million output tokens, with cached input storage free for a limited time. OpenRouter lists the same rates.
Is GLM-5.3-FlashX-Legal open source?
As of November 4, 2026, no legal-specific weights are listed under Z.ai’s zai-org account on Hugging Face. The base GLM-5.3-Flash weights are published there, but the Legal variant is available through the Z.ai API, OpenRouter and AutoClaw.
How did GLM Legal score on Harvey LAB?
Z.ai reports a 19.2% all-criteria pass rate on Harvey LAB, 12.5 percentage points above the base model and the best result among open-source models in that evaluation. Z.ai also says it matched frontier models on LegalBench and PLawBench but did not publish those scores on its model page.
Can I turn off thinking in GLM-5.3-FlashX-Legal?
No. Z.ai’s documentation lists enabled as the only supported value for thinking.type on the GLM-5.3-Flash family, and recommends reasoning_effort set to max with clear_thinking set to false.
Is GLM-5.3-FlashX-Legal in the GLM Coding Plan?
Not yet. Z.ai’s documentation says the FlashX models are not currently included in the GLM Coding Plan, so you pay per token through the API.
Sources
- GLM-5.3-Flash model documentation (Z.ai)
- Pricing (Z.ai Developer Docs)
- GLM-5.3-flashX-Legal Is Now Available in AutoClaw (AutoClaw, October 29, 2026)
- Z.ai: GLM-5.3-FlashX-Legal, pricing and providers (OpenRouter)
- zai-org/GLM-5.3-Flash model card (Hugging Face)
Image: Z.ai company logo, public domain (trademarked), via Wikimedia Commons. Z.ai is a trademark of its owner. FreeAnswers is not affiliated with or endorsed by Z.ai.
.svg){kind=link}