MiniMax M2.7 vs GLM-5.2: an honest comparison.

Quick answer: ceiling vs efficiency.

If the workload is hard agentic coding and you want the strongest published open-weight results, GLM-5.2 is the default candidate: its launch-reported SWE-bench Pro score of 62.1 led the class in June 2026, and Z.ai's whole positioning is long-horizon coding agents. If the workload is high-volume agent loops where cost per completed step decides the architecture, MiniMax-M2.7 is very hard to argue against: MiniMax's published rate runs at a fraction of Z.ai's, the model activates about a quarter of GLM-5.2's per-token parameters, and its vendor-reported agent benchmarks sit within striking distance rather than a class below. Neither answer is universal, and the two models fail in different directions: GLM-5.2 bills heavy output on easy tasks it overthinks, while M2.7's headline numbers are vendor-reported and its license carries a commercial-use clause the MIT-licensed GLM does not. The sections below take the dimensions one at a time, with sources.

The spec sheet, side by side.

Both are open-weight Mixture-of-Experts reasoning models from Chinese labs, released in spring 2026 and aimed squarely at agent work. The designs diverge on size and posture: GLM-5.2 is the heavyweight, M2.7 the lightweight, and the licenses differ in a way that matters commercially. Figures below come from each vendor's announcements and documentation plus launch coverage, as of July 2026.

Coding: GLM-5.2 holds the reported edge, M2.7 holds the ratio.

On the published numbers, GLM-5.2 leads: launch coverage put it at 62.1 on SWE-bench Pro against GPT-5.5's 58.6, which would make it the open-weight leader at release, and Z.ai's release notes claim open-source state of the art on coding and long-horizon benchmarks. M2.7's vendor-reported 56.2 on the same benchmark trails that by around six points, alongside 57.0 on Terminal-Bench 2.0 and 76.5 on SWE-bench Multilingual, a family the M2 line has historically led among open models. Both sets of numbers deserve their hedges: GLM-5.2's headline figure comes from launch press, M2.7's from the vendor itself, and neither substitutes for a run on your own repository. The engineering question is what six benchmark points cost. At the vendors' own July 2026 list rates, GLM-5.2 prices input at several times M2.7's rate, and the gap on output is similar. For the hardest tickets, where a failed run costs an engineer's afternoon, the stronger model is usually worth it. For the long tail of refactors, test generation, and review comments, a model a few points behind at a fraction of the cost per step frequently wins the month's bill without losing the month's throughput. That line sits in a different place for every team, which is why both ids behind one key is the useful configuration.

Agent behavior: two different failure modes.

A pattern several teams land on mirrors other pairings on this site: the heavyweight as planner and hard-step model, the efficient model as the high-volume executor. Behind one endpoint that split is two model strings in the same loop, and the usage log's completion_tokens column shows exactly what each side spends per completed step, which is the number that should settle this page's question for your workload.

• GLM-5.2 thinks hard by default: an independent measurement put its output volume around 43K tokens per task against 26K for GLM-5.1 on the same benchmark. On hard problems that is the feature; on easy steps it is a tax, since reasoning bills as output.
• M2.7's design goes the other way: about 10B active parameters per token keeps per-step cost and latency low, which compounds over loops that run hundreds of steps. MiniMax's launch framing, inspect, reason, modify, iterate, targets exactly that shape.
• MiniMax's self-evolution story, an internal version autonomously running 100-plus optimization rounds on its own training pipeline, is the launch's memorable claim. It is unverifiable from outside and should not carry weight in your routing decision.
• GLM-5.2's thinking is tunable: a toggle plus an effort setting let you dial reasoning down on easy steps, which partially closes the efficiency gap if you actually use it.

Context and licensing: read the fine print on both.

Context is a both-ways caveat. Z.ai publishes a 1M window for GLM-5.2, but the catalog here serves it at 200K, so budget GLM traffic against 200K on this route. MiniMax publishes roughly 200K for M2.7, while the catalog entry here lists 1M; the conservative call is the vendor's 200K until an oversized test request proves otherwise. Net effect: on this route, plan both models around 200K and let tests, not listings, extend that. Licensing splits cleanly. GLM-5.2's weights are MIT: use them, ship them, sell on top of them. M2.7's weights are Modified-MIT with a clause requiring MiniMax's prior written authorization for commercial use, a change that drew its own coverage in April 2026. For API access through a gateway this is moot, you are consuming a hosted service either way, but for any self-hosting plan it makes GLM the unencumbered choice and M2.7 a legal conversation first. Output ceilings differ too: Z.ai documents up to 128K output tokens for GLM-5.2, useful for very long single generations, while M2.7's published spec centers on the standard agent-loop range. If a workload emits whole documents in one call, that asymmetry matters more than most benchmark rows.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Which one, for which workload.

Judge three numbers per side: quality on your rubric, completion_tokens per completed task, and wall-clock latency on interactive seats. A model that wins two of three usually wins the seat.

• Hardest agentic coding, complex refactors, long-horizon plans: GLM-5.2 first. The launch-reported benchmark edge sits exactly on this ground.
• High-volume agent loops, CI bots, batch code tasks: MiniMax-M2.7 first. Cost per completed step is the metric, and the 10B-active design is built for it.
• Very long single outputs (reports, transcripts, large rewrites): GLM-5.2, on its documented 128K output ceiling.
• Self-hosting inside a commercial product: GLM-5.2, on license terms alone, unless you have MiniMax's authorization in writing.
• Tightly budgeted output: watch both, differently. GLM-5.2 needs its thinking effort managed down on easy steps; M2.7 needs its vendor-reported quality verified up on hard ones.
• Undecided: run both. The A/B below is the entire setup, and at these two rates it costs almost nothing.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APISROUTER_API_KEY"],
    base_url="https://api.apisrouter.com/v1",
)

PROMPT = "Fix the failing test in this module and explain the root cause: ..."

for model in ("MiniMax-M2.7", "glm-5.2"):
    r = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": PROMPT}],
        max_tokens=4000,
    )
    print(model, r.usage.completion_tokens, "output tokens")

FAQ