Grok 4.7 vs GLM-5.2: The Licence Outlives the Model

The most valuable thing about GLM-5.2 is not its benchmark score. It is the four letters MIT. Z.ai released GLM-5.2 on June 16, 2026 under a standard MIT licence with no regional limits, no revenue threshold and no commercial-use condition beyond retaining the copyright notice — and when Z.ai shipped GLM-5.3 two months later, it did not do that again. The successor carries a custom "GLM-5.3 License" that adds a carve-out: a model-as-a-service operator with more than US$10 billion in revenue over any consecutive twelve months must pass a security review, with the scope of that review "reasonably determined by Z.AI." GLM-5.2 is the last text flagship Z.ai released without that string attached. Against Grok 4.7 — released by the vendor on September 21, 2026, closed-weight and a day old — that makes this a comparison between the strongest model on the board and the freest one.

One of these you can own outright

GLM-5.2 is a 753-billion-parameter mixture-of-experts model with roughly 40B active per token, published in BF16 and FP8 on Hugging Face under the zai-org organisation. It carries a 1M-token context window and a 128K maximum output. It is text in, text out — no native vision — with function calling, MCP integration, structured JSON output and context caching, and a thinking mode that is on by default with a reasoning-effort setting. Z.ai's first-party API lists it at $1.40 per million input tokens, $0.26 for cached input, and $4.40 per million output.

Grok 4.7 is closed. It holds a 500k-token context, takes text and images in, and lists at $2.00 in and $6.00 out per million below 200k prompt tokens, stepping to $4.00 and $12.00 above that line, with cached reads at $0.50 and $1.00. There is no weight download, no self-host path, and no way to keep calling it if xAI changes its mind — the same structural position as every closed flagship. Both models are capable; only one of them is an asset you can put on a disk.

What the index says, and what "deprecated" means here

Artificial Analysis scores GLM-5.2 at 34 on Intelligence Index v4.3.2 — the revision published September 19, 2026 — and Grok 4.7 at 46 on the same revision. That twelve-point gap is real, and the shape of it matters more than the size. GLM-5.2 also carries a "deprecated" label on its Artificial Analysis page, with a note that a newer model exists and that only the default 10k-token workload continues to be benchmarked. That label is about succession, not shutdown: Z.ai still serves glm-5.2 today and still lists it in its pricing table.

• Composite — Grok 4.7 (xhigh): 46 on Artificial Analysis Intelligence Index v4.3.2. GLM-5.2 (max): 34 on the same revision.

• Agentic terminals — Grok 4.7: Terminal-Bench 4.0 26. GLM-5.2: 1. This is the most extreme single figure in this batch of comparisons, and it needs the caveat that goes with it. Z.ai's own model card reports GLM-5.2 at 81.0 on Terminal-Bench 2.1, best harness 82.7. Both numbers are honest; they are on different benchmark versions, and the v4.3.2 index retired 2.1 in favour of 4.0. A vendor score on a retired benchmark tells you what a model could do against the old test, not what it does now.

• Workflow automation — Grok 4.7: AutomationBench-AA 66%. GLM-5.2: 28%. Both measured on v4.3.2, so this one needs no version caveat.

• Knowledge honesty — Grok 4.7: AA-Omniscience 32. GLM-5.2: 4. Again the same version on both sides.

• Reasoning and long context — Grok 4.7: HLE 43, CritPt 18, AA-LCR v1.1 77%, window 500k. GLM-5.2: HLE 41, CritPt 21, AA-LCR v1.1 78%, window 1M. GLM-5.2 wins two of these four, and the context window by a factor of two.

• Speed — Grok 4.7: not yet measured by Artificial Analysis. GLM-5.2: 73 tokens per second with a 5.98-second time to first token through Z.ai's API, described as slower than average for open-weights peers.

• Price — Grok 4.7: $2.00 / $6.00 per 1M, doubling at 200k prompt tokens. GLM-5.2: $1.40 / $4.40 per 1M with $0.26 cached reads, unchanged since launch.

The self-host economics are the real argument

Per-token, GLM-5.2 is roughly thirty per cent cheaper on input and twenty-seven per cent cheaper on output than Grok 4.7 at short prompts — a modest discount that would not decide anything on its own. The reason it matters is that the per-token price is optional. GLM-5.2's weights are downloadable under a licence that permits commercial use, modification, redistribution and sublicensing, which means the relevant comparison is not $1.40 against $2.00 but $1.40 against whatever your own GPUs cost per million tokens. A 753B model with 40B active is not a laptop workload; it is a serious inference deployment, and the FP8 checkpoint exists precisely for that. But it is a deployment you can run, cap and audit — and for a regulated workload or an air-gapped one, that is not a cost question at all.

The counterweight is succession risk in the other direction. GLM-5.2 is two releases behind: GLM-5.3 landed August 18 and GLM-5.3-Flash on August 26, and the newer flagships score higher — GLM-5.3 at 45 on v4.3.2, one point below Grok 4.7. So the honest framing is that GLM-5.2 is the free one, not the best one, and that Z.ai's own newer models are both stronger and more restrictively licensed. If permissive licensing is what you need, GLM-5.2 is the endpoint to freeze; if raw capability on an open weight set is what you need, the successor is the one to evaluate, with a lawyer's eye on the revenue carve-out.

Running both without running two stacks

OrcaRouter carries GLM-5.2, listed on its own model page at $1.40 in and $4.40 out with the 1M context window and 128K maximum output, because we pass provider list pricing through at 0% markup. With an open-weight model that pass-through has a second use: the hosted price and the self-hosted option are both live choices, and a router lets you move traffic between them per request rather than per contract. Z.ai has held GLM-5.2's list price flat since June while third-party hosts have moved theirs around, so a same-day pass-through is the difference between quoting today's rate and quoting one you copied last month. Automatic failover covers the case where a self-hosted deployment saturates and the hosted endpoint has to absorb the spill — the routing DSL composes the two into one call. Grok 4.7 is not in the OrcaRouter catalog as of this writing; calling it means going through xAI's own API and the third-party platforms that carry it.

The decision rule that falls out of all this is unusually clean. If you need to own the model — on-premise, audited, licensable without a revenue review — GLM-5.2 is the only one of these two that offers it, and the twelve-point index gap is the price of that freedom. If you need the model that finishes agentic work, the gap runs the other way and it is not subtle: 66 per cent against 28 per cent on workflow automation, 26 against 1 on terminal agents, 32 against 4 on knowledge honesty. Nothing in the licence question changes those numbers.

The call

Grok 4.7 wins the capability comparison and it wins it on the evaluations that predict whether an agent completes a task. GLM-5.2 wins on the two things no benchmark measures: it is cheaper, and it is yours. The detail that makes this worth writing down is that the second advantage is expiring — Z.ai's newer flagships abandoned the unrestricted MIT licence, so a model released in June 2026 under terms that ask nothing of you is now a historical artefact rather than a policy. If permissive open weights are load-bearing for your product, that is a reason to standardise on GLM-5.2 while it is still the cheap hosted option and still downloadable. If they are not, buy the twelve points.