GLM-Legal vs Astra for Law: Two Ways to Build a Legal Model, and Only One of Them Sells You an API

The two most interesting legal models released in the last month solve the same problem by opposite means. GLM-Legal is a weights-level specialisation — Z.ai took GLM-5.3-Flash and put it through legal mid-training and reinforcement learning until it behaved like a practitioner. Astra for Law is a configuration — the vendor left GPT-6 Astra exactly as it was and bolted a 230-million-URL U.S. case-law index and a set of legal instructions onto the front of it. One is a smarter lawyer with the wrong library; the other is an ordinary lawyer with an extraordinary library. Which one you want depends almost entirely on whether your problem is reasoning or retrieval, and the two releases are graded on benchmarks that cannot be put side by side.

That last point is the reason this comparison is harder than the headline makes it look, so it is worth getting out of the way first.

Why no number in this piece compares them head to head

GLM-Legal's headline figure is 19.2% all-pass on Harvey LAB. Astra for Law's is 54.0% overall correctness on a private U.S. legal-research validation set. Anyone can divide 54 by 19.2 and conclude the configured version is nearly three times better. That arithmetic is meaningless, and here is exactly why:

• What is being graded — GLM-Legal on end-to-end law-firm workflow tasks across 24 practice areas vs Astra for Law on U.S. legal research questions. Different task families, not different difficulty settings of the same task.

• Grading strictness — Harvey LAB is all-pass: a task counts only if every step lands. The Astra figure is overall correctness, a partial-credit aggregate. All-pass scoring mechanically produces lower numbers on the same model, and a 19.2% all-pass rate is not comparable to any percentage that awards partial credit.

• Baselines — Z.ai measures GLM-Legal against GLM-5.3-Flash and reports a 12.5-point lift. The GPT maker measures Astra for Law against GPT-6 Astra with plain web search — 38.7% to 54.0% — and the retrieval is doing nearly all of that work. Both are delta claims against the vendor's own base, not against each other.

• Jurisdiction — legal effectiveness is jurisdiction-bound, and neither harness covers the other's ground. Harvey LAB and the Astra validation set are U.S.-focused; PLawBench, GLM-Legal's second benchmark, is Chinese. A model can be excellent on one and useless on the other and still score well on both releases' numbers.

• Independence — zero on both sides. No third party has re-run Harvey LAB with GLM-Legal, and no one outside the vendor controls the split the 54.0% was measured on.

The one public board that has touched this territory is Vals AI's Legal Research Bench, where Astra for Law has been scored at 9.4% correctness against 22.5% for an unnamed frontier model — a reminder that as strict grading rubrics replace vendor-chosen splits, headline numbers in this category tend to move down, not up. That is not a verdict on Astra for Law; it is a reason to hold every figure in this comparison loosely.

What each one is actually built to do

The architecture difference is not academic — it decides what each model can be good at.

• GLM-Legal — the model changed. Legal corpora in mid-training, then RL across what Z.ai describes as tens of thousands of realistic legal environments, with the agent harness co-evolving alongside the weights. It ships reasoning, not a library.

• Astra for Law — the model did not. Same GPT-6 Astra weights, plus a retrieval layer over U.S. case law, statutes and regulations, plus legal-specific instructions. It ships a library, not a better lawyer.

• GLM-Legal's evidence — a side-by-side against its own base. On the same prompts, GLM-5.3-Flash gives competent general advice; GLM-Legal resolves the numbers — reducing a convenience-termination notice from 60 to 30 days, removing the cure period for IP and confidentiality breaches, requiring return of work product whether or not it has been paid for, adding up to 90 days of transition assistance at the provider's rates.

• Astra for Law's evidence — a retrieval delta. 54.0% against 38.7% overall correctness, 24% more reference cases found, up to 54% more relevant passages retrieved from correct opinions, and roughly twice the answer length. The vendor's own numbers, on a split nobody else can run.

• The honest summary of the demos — the base model finds the facts, the specialised model builds the argument. In Z.ai's evidence-analysis example, GLM-5.3-Flash spots a contradiction between a 60%-complete client report and a 41% internal snapshot and says the documents "may support" the client's position. GLM-Legal links an earlier internal warning, the snapshot and an instruction to suppress the risk into a single concealment theory and names the claim elements still to test.

Read together, those two evidence bases describe a clean division. GLM-Legal is being sold on judgement; Astra for Law is being sold on not missing the case.

Where each one is likely to fail you

Both releases document their own weaknesses, which is more than most launches manage, and the failure modes are mirror images.

Astra for Law's most instructive result came from a third-party test rather than the launch itself. Across non-U.S. jurisdictions, three models tested by the legal-AI firm HAQQ each cited the correct provision while misstating what it says — Astra was legally correct on 66.7% of those items, against 100% for Claude Opus 5. The citation checker passes those answers, because the citation is real and correct. The client does not. This is the failure the Legal Search Index is structurally worst placed to fix: the index is U.S.-only, and the error is in the reading rather than in the retrieval.

GLM-Legal's weaknesses sit in the opposite direction. It is tuned toward Chinese and multi-jurisdictional statutory practice — the compliance demo resolves Singapore, Malaysian and Indonesian privacy obligations with section-level citations — and it carries no equivalent of a U.S. case-law index. On American research the architecture that finds 230 million URLs of precedent is doing something GLM-Legal simply does not do. Its published strengths are drafting, evidence linking and rubric-graded document work; its published evidence does not include a retrieval claim at all.

Neither vendor has published the governance paperwork a law firm's general counsel will ask for first. Astra for Law arrives with Zero Data Retention on the API and ChatGPT Enterprise usage excluded from human review by default, but that coverage is granted per organisation and the programme is invitation-based. GLM-Legal's beta terms are not published at all, which is not the same as being unacceptable — it is simply undocumented, and undocumented is what you cannot take to a risk committee.

Getting to try either one, and what it costs

Neither legal configuration is on a public rate card.

• Astra for Law — Trusted Access, invitation-based and aimed at large firms; no published price for the legal configuration; the API model id has been described as coming soon.

• GLM-Legal — a beta application aimed at law firms, legal-tech companies, in-house teams and independent practitioners; no published price; no public endpoint.

• The underlying models, though, are both live. GPT-6 Astra sits at $10 per million input tokens and $50 per million output on a 1.05M-token context, with a long-context reprice above 272,000 input tokens that legal document sets will cross routinely. GLM 5.3 Flash sits at $0.15 and $0.50 on a 1M-token context. That is a 66-fold difference in input price for two models that both take a contract set in one pass.

Both are on the OrcaRouter catalogue at those list prices with no markup added, which matters more than usual in this category: a vendor price change on either model is live here the same day rather than at the end of a contract cycle. One key covers both, so a legal engineering team that wants Astra for the retrieval-heavy work and GLM 5.3 Flash for high-volume first-pass review does not need two commercial relationships, two SDKs or two sets of credentials. The routing DSL splits traffic by task class, so the drafting path and the summary path can point at different models and still bill on one invoice.

And there is a deliberate restriction worth naming rather than hiding. The thing that actually distinguishes Astra for Law from GPT-6 Astra is the Legal Search Index, and that layer is not for sale — you cannot route to it, and neither can anyone else. What we route is the base model. A team that prototypes a legal workflow on GPT-6 Astra today is testing everything except the part that made the legal configuration better.

Which one to pick

If the work is U.S. case-law research, precedent-heavy analysis, or anything where the primary risk is missing an authority, Astra for Law is aimed squarely at that and GLM-Legal is not competing for it. If the work is drafting to a specification, statutory compliance across named regimes, or reasoning over a document set you have already supplied, GLM-Legal's evidence is more directly on point — and its demos show the specific behaviour you want, which is a model that produces numbers and clause language rather than a checklist of things to consider.

For most legal teams the practical answer this quarter is that neither is a decision you have to make yet. Both are application-gated, both are unpriced, and both have exactly one vendor-controlled benchmark run behind their headline numbers. The teams that will be ready when the gates open are the ones who spent the interval building the pipeline on base models that are actually callable — so that adopting the specialised version later is a model-id change and not a re-platforming.

The routing DSL splits traffic by task class, so the drafting path and the summary path can point at different models and still bill on one invoice.