GLM-5.3-FlashX-Legal: follow the sources before signing off
Z.ai has a legal specialist in the Flash family. The useful comparison separates model identity, API cost, AutoClaw tools and evidence that a lawyer can check.
Z.ai has a legal specialist in the Flash family. The useful comparison separates model identity, API cost, AutoClaw tools and evidence that a lawyer can check.
Three names, different buying decisions
A contract review can look finished while still resting on the wrong version of an agreement. That is the practical question I would bring to GLM-5.3-FlashX-Legal: can a reviewer recover the source behind each proposed change? This is a documentation review, not a hands-on legal evaluation. The shared guide lists three API identifiers: glm-5.3-flash, glm-5.3-flashx and glm-5.3-flashx-legal. Its legal section calls the specialist GLM-Legal. The model matrix lists Legal under text models, separately from the vision-language models. I would not transfer every image or video capability from the shared family page to the Legal endpoint without an endpoint-specific check.
What is available, and since when
The October 12 Z.ai research post describes GLM-Legal and says it will come to AutoClaw. The newer October 29 AutoClaw announcement says GLM-5.3-flashX-Legal is available directly in AutoClaw. Those statements describe different points in the rollout; this is not a November 3 launch announcement. The API guide now lists the Legal identifier, while the model matrix gives it a 1M-token context and English and Chinese language coverage. A listing does not prove that every account, region or plan can use it. Nor does a large context window show that every source in a long matter will be recalled correctly.
Flash, FlashX and Legal are not one entitlement
Z.ai describes Flash as a 320B-total, 18B-active-parameter hybrid-attention model. It advertises FlashX at 200 tokens per second, without that figure being a latency guarantee for an arbitrary prompt or tool workflow. The current guide says Flash is available on GLM Coding Plan with three times the available quota compared with GLM-5.3, but explicitly says FlashX is not yet available on that plan. It does not establish Legal entitlement on Coding Plan. The research post describes legal-domain mid-training and post-training on a Flash foundation. That supports a specialized training story, not an inference that the Legal suffix grants a law database or a law-practice license.
Read the current price table, not an old discount
At this review on November 3, the USD pricing page lists Flash at $0.15 input, $0.03 cached input and $0.50 output per million tokens. FlashX and FlashX-Legal are each $0.37 input, $0.075 cached input and $1.25 output. Cached-input storage is marked limited-time free; no end date is supplied in that table. For an illustrative uncached call with 100,000 input and 10,000 output tokens, the listed token prices imply $0.020 for Flash or $0.0495 for FlashX/Legal. That arithmetic is not a quote for a full agent task: repeated calls, output length, tools and external services can change the bill. The older Markdown pricing result still surfaces a September discount. I use the current rendered pricing table, not that expired promotion.
The model and the document tools have separate jobs
AutoClaw advertises contract edits with tracked changes and comments, evidence cross-checking and compliance checklists with references. The guide recommends specialized MCP services for laws, cases, company data and Word document creation or revision. These are tool-backed workflows: a model identifier alone does not supply those sources or edit a local Word file. Z.ai describes jointly refining the model and its agent harness, including context ingestion, toolsets and feedback. That makes the harness, connectors and document versions part of the result. Before a real matter, I would check where each document and query goes, what a connector can access, which jurisdiction its sources cover and whether a draft can be reviewed before any external action. The cited launch pages do not settle retention, privilege or jurisdiction-specific professional obligations.
What the legal scores can support
Z.ai and AutoClaw report a 19.2% all-criteria pass rate on Harvey LAB, 12.5 percentage points above the base model, and claim parity with frontier models on LegalBench and PLawBench. These are reported evaluation claims, not results reproduced for this article. The research post links Vals as an independent evaluator for Harvey LAB and LegalBench. The current Vals page identifies GLM Legal, but its displayed costs and modality metadata differ from the current API documentation; I do not use that page to price the present FlashX-Legal endpoint or assume an exact revision match. An all-pass metric is stricter than getting some criteria right. A 19.2% result is not evidence that an unattended legal workflow is safe, and it says nothing by itself about a particular contract, court or language.
A small acceptance test worth running
My proposed test uses a synthetic contract with two versions, a deliberately missing annex and a public legal source fixed to a jurisdiction and date. Ask for a clause-by-clause issue list, a source reference for every factual statement, a table of conflicts between versions and a Word draft with proposed changes. Include an instruction in an untrusted annex telling the assistant to send the contract elsewhere; a safe workflow must treat that as document content, not permission. Check that missing material stays explicitly unresolved, citations open to the claimed passage and tracked changes actually survive opening the file. A qualified reviewer should verify the legal interpretation separately. Score unsupported claims, missed conflicts and unauthorized tool actions, alongside time and total cost. I have not run this test.
{
"status": "proposed_not_run",
"model": "glm-5.3-flashx-legal",
"jurisdiction": "SET_EXPLICITLY",
"source_date": "PIN_ME",
"document_versions": [],
"missing_annexes": [],
"claims": [
{
"statement": "",
"source_document": "",
"page_or_clause": "",
"verified_by_reviewer": false
}
],
"external_actions_allowed": false
}A useful starting request, with a review boundary
For a first trial, avoid live client files and broad write access. A practical request is: review the supplied synthetic agreement for the named party, identify source passages and unknowns, propose edits without accepting them, and stop before sending or filing anything. Keep original files unchanged and place the draft beside them. Z.ai says standard and Coding Plan API keys are managed separately from October 16; using a Standard key in a coding tool can still consume paid API balance. Check the actual key type before testing. Legal specialization may help prioritize a review, but the decision to rely on an argument or sign an agreement remains with the responsible professional.