GLM-Legal: $0.25 a Contract Review, Not $0.16

Z.ai announced GLM-Legal on 12 October 2026. Its API name is glm-5.3-flashx-legal, and it's GLM-5.3-Flash with extra legal training on top. On Harvey's legal agent benchmark (Harvey LAB) it passes 19.2% of tasks on every criterion, which Z.ai says is 12.5 points better than the base model. Vals AI ran that test independently and put it 7th of 79, at $0.15 per test.

The price is the first thing people get wrong. Vals lists GLM-Legal at $0.075 per million input tokens and $0.25 per million output. Z.ai's own pricing table says $0.37 in, $0.075 for cached input and $1.25 out. Below I cost one contract review session at Z.ai's real rates, set it against Claude Sonnet 5.5 and Haiku 5.5, and walk through the MCP setup Z.ai recommends, including what it leaves out for Indian law.

Table of contents

• What GLM-Legal is
• The price on Vals is not the API price
• One contract review session, costed
• Connecting the Word editing MCP
• What India gets
• Catches before you build on it
• FAQ
• Sources

What GLM-Legal is

Z.ai trained it in two stages. Mid-training added a larger set of legal text. Post-training then ran reinforcement learning in "tens of thousands" of task environments that copy real legal work: reading the file, spotting the issues, calling tools and drafting the document. The model ships inside the same docs page as GLM-5.3-Flash and FlashX, and the docs give it the same 1M context and 128K output.

The Vals update text and the model page don't quite agree on ranks. The update says 14th of 150 on LegalBench and 24th of 78 on Legal Research Bench. The page header says 15th of 151 and 27th of 89, probably because more models were added later. The scores themselves match.

The price on Vals is not the API price

Vals writes: "GLM Legal is priced at $0.075 per million input tokens and $0.25 per million output tokens." Z.ai's pricing page lists GLM-5.3-FlashX-Legal under text models at $0.37 input, $0.075 cached input and $1.25 output.

So Vals has the cached input rate as the input price, and an output price one fifth of Z.ai's list. If you budget from Vals you'll undercount fresh input by 4.9x and output by 5x. For the session below, that's $0.16 on paper against $0.25 on the bill. Vals's "$0.15 per test" for Harvey LAB is their own measured spend, and I couldn't tell from the page which rates they billed it at.

One contract review session, costed

The workload: a lawyer or in-house team member works through one services contract with an agent.

• A 60K-token prefix: the contract itself, the system prompt and the loaded skills. It stays the same every turn.
• 30 turns. Each turn adds 4K tokens of new text (instructions, a clause, a tool result) and returns 1.5K tokens.
• Turn 1 pays full price for the prefix. Turns 2 to 30 read it from cache.
• I've ignored conversation history growing turn by turn, so treat these as a lower bound for every model alike.

That's 1,920,000 input tokens and 45,000 output tokens per session. Z.ai's caching is automatic and has no write fee, so the first turn bills the prefix as normal input. Anthropic bills it as a 5-minute cache write.

The 88 sessions are 4 a day over 22 working days. Rates used: Sonnet 5.5 at $2 in, $2.50 for a 5-minute cache write, $0.10 cache read and $10 out. Haiku 5.5 at $0.10 in, $0.125 write, $0.01 read and $0.50 out, which applies because every prompt here stays under its 100,000-token line.

Three things stand out.

GLM-Legal is a quarter of the Sonnet bill. Most of the gap is output: Sonnet's $10 against $1.25 makes $0.45 versus $0.06 on the same 45K tokens.

Cache reads are half of GLM-Legal's bill. $0.1305 of the $0.2534 is the contract being read again 29 times. Z.ai's $0.075 cached rate is a lot closer to Sonnet's $0.10 than its fresh input rate is to Sonnet's $2, so the saving shrinks as the contract gets longer and the turns get shorter. Without caching the same session would cost $0.77.

Haiku 5.5 is cheaper still, at about a quarter of GLM-Legal. It has no legal training, and I haven't seen Haiku 5.5 results on Harvey LAB, so the cost table can't tell you whether it's good enough. Plain GLM-5.3-Flash ($0.15 in, $0.03 cached, $0.50 out) comes to $0.10. Z.ai's own blog shows Flash giving vaguer contract advice than GLM-Legal on the same task, which is the case for paying 2.5x more.

To try your own token counts, put them into the API cost calculator. It now lists GLM-5.3-FlashX-Legal under Z.ai.

Connecting the Word editing MCP

Z.ai suggests you run GLM-Legal inside an agent connected to its MCP servers. The model plans and reasons, and the servers supply laws, cases and Word editing. The one Z.ai marks "recommended to install" for every country is Edit Word DXE. It edits .docx files with tracked changes.

DXE uses Streamable HTTP. You need an agent that supports that transport plus file upload and download, and a standard Z.ai API key. Z.ai's page says a Coding Plan key won't work. The config from the docs:

{
  "mcpServers": {
    "legal-dxe": {
      "url": "https://api.z.ai/api/mcp-broker/proxy/legal-dxe/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_ZAI_API_KEY"
      }
    }
  }
}

It exposes nine tools, all listed as free: dxe_save, dxe_transform, dxe_guide, dxe_upload, dxe_read, dxe_inspect, dxe_validate, dxe_ensure_header and dxe_ensure_footer. The test Z.ai gives is simple. Hand the agent a contract and ask it to change the payment term from 30 days to 15, keep tracked changes and return the file. If you paste the key into a JSON file, keep that file out of git. The JSON formatter will catch a missing comma before your agent fails to start.

Z.ai also publishes skills as zip files, such as contract-review, contract-desensitize (replaces names and figures and keeps a mapping), case-brief, tabular-review and escalation-flagger. You load them into the agent next to the MCP.

What India gets

Z.ai's install list covers 36 sections, from a global set to individual countries. India's section has the same general skill set as the others. Under MCP it says "None" and tells you to search these sites instead:

• India Code (laws and regulations)
• Supreme Court of India (case law)
• NCLAT (case law and rulings)
• Competition Commission of India
• Ministry of Corporate Affairs
• SEBI

The US, EU, Japan and Hong Kong each get a dedicated MCP that queries the official source, such as the Federal Register and eCFR for the US. In India the agent has to fetch these sites through its own web tools, and Z.ai's built-in web search costs $0.01 per use. Ten lookups add $0.10 to the $0.25 session, which is a 40% jump. Budget for it if your reviews cite statutes or judgments.

Catches before you build on it

API access may need a form. The pricing table and docs list the model code, but the launch blog says "Join the GLM-Legal Beta" and links an application at z.ai/consultation/legal. Test a call with your own key before you promise anyone a workflow.

DXE billing isn't final. Z.ai's own page says the DOCX revision tests passed, and "USD billing is still pending verification." The tools show as free today. That may change.

"Open weights" isn't verifiable yet. Vals tags GLM-Legal as open-weight. On 3 November I found no GLM-Legal repo under zai-org on Hugging Face, only GLM-5.3-Flash and its relatives. Don't plan a local install from the Vals tag. If a repo does appear, it will need the same memory as GLM-5.3-Flash, which our GLM-5.3-Flash local guide puts at 93 GB even at 1-bit.

The benchmark bar is low in absolute terms. 19.2% all-criteria pass means about four in five Harvey LAB tasks miss at least one requirement. Z.ai's own examples end with the model telling the team to check before concluding. A qualified lawyer still signs off.

Recommended settings. The GLM-5.3-Flash docs say temperature: 1, top_p: 0.95, reasoning_effort: max, and thinking.type only accepts enabled. Thinking tokens bill as output, so on long reasoning turns the $1.25 output rate will push the session above $0.25.

FAQ

Is GLM-Legal on the GLM Coding Plan? The docs say GLM-5.3-FlashX is not on the plan yet and only GLM-5.3-Flash is. GLM-Legal is a FlashX variant, so expect pay-as-you-go API billing.

Does it read scanned contracts? The model takes image and file input, so a scanned PDF can go in directly. For a clean text layer first, a dedicated OCR model helps; we covered Unlimited OCR 2 on 29 October.

Does caching need any setup on Z.ai? No. It's implicit. Send the same prefix again and the response shows the cached count in usage.prompt_tokens_details.cached_tokens. Keep the contract at the start of the prompt and put changing text after it, or the cache won't match.

Sources

• Z.ai, "GLM-Legal: Advancing Agentic Legal Intelligence" (z.ai/blog/glm-legal, dated 12 October 2026, read 3 November 2026 in a browser): training method, Harvey LAB result, AutoClaw examples, beta application link.
• Z.ai docs, Pricing (docs.z.ai/guides/overview/pricing): GLM-5.3-FlashX-Legal, GLM-5.3-Flash and GLM-5.3 rates, web search at $0.01 per use.
• Z.ai docs, GLM-5.3-Flash/FlashX (docs.z.ai/guides/vlm/glm-5.3-flash): model code, context, max output, input types, recommended settings, Coding Plan status, the GLM-5.3-FlashX-Legal section.
• Z.ai docs, Recommended Installation List by Country/Region (docs.z.ai/guides/legal-mcp): DXE config, tool list, billing status, skills, the India section.
• Z.ai docs, Context Caching (docs.z.ai/guides/capabilities/cache): implicit caching and the cached_tokens field.
• Vals AI, GLM Legal model page (vals.ai/models/zai_glm-legal, read 3 November 2026): release date, scores, ranks, listed prices, the open-weight tag.
• Anthropic, Claude pricing (platform.claude.com/docs/en/about-claude/pricing, read 3 November 2026): Sonnet 5.5 and Haiku 5.5 rates and the Haiku 100,000-token line.
• Hugging Face API, models by author zai-org (read 3 November 2026): no GLM-Legal repository listed.
• Session costs, monthly totals and ratios are calculated by ToolMintX from the rates above.