GLM-5 vs MiniMax M3: Open Models Got Serious
Two open-weight releases in June 2026 — Zhipu's GLM-5.2 and MiniMax's M3 — closed most of the gap with the closed frontier while costing a fraction of the price. One is the capability leader; the other is the value pick with 1M context and native multimodality. If you've been waiting for open models to be a serious production option, the wait is over. Here's how to choose.
For two years, the honest advice about open-weight models was: great for experimentation, fine for narrow tasks, not what you bet a production system on. June 2026 ended that era. In the span of two weeks, MiniMax shipped M3 (June 1) and Zhipu shipped GLM-5.2 (June 13) — and between them, the open-weight tier now beats last year's closed flagships on several benchmarks that matter, at prices that make the closed vendors' invoices look like a rounding error with a margin problem.
The head-to-head numbers
GLM-5.2 is the open-weight capability leader, full stop. It scores 51.1 on the Artificial Analysis index — well clear of the 42-44 cluster where MiniMax M3, DeepSeek V4 Pro, and the rest of the chasing pack sit — and its Terminal-Bench 2.1 score of 81% doesn't just lead the open tier, it lands within four points of Claude Opus 4.8's 85% and above GPT-5.5's 84%. Read that again: an MIT-licensed model you can run on your own hardware is now within arm's reach of the closed workhorse tier on agentic terminal work.
MiniMax M3's answer isn't to out-benchmark GLM on coding — it's to change what the comparison is about. M3 is the first open-weight model combining frontier-adjacent coding, a 1M-token context window, and native multimodality: it reads text, images, and video, and it can operate a desktop. On BrowseComp, the autonomous web-research benchmark, M3 scores 83.5% — above Claude Opus 4.7's 79.3%. And its MSA attention architecture delivers roughly 9.7x faster prefill and 15.6x faster decode than standard full attention, which is why it's priced the way it is.
| Metric | GLM-5.2 | MiniMax M3 |
|---|---|---|
| AA Index (open-weight tier) | 51.1 — open leader | ~42-44 cluster |
| SWE-bench Pro | 62.1% | 59.0% |
| Terminal-Bench 2.1 | 81.0% | 66.0% |
| MCP Atlas (tool use) | 77.0% | 74.2% |
| BrowseComp (web research) | not published | 83.5% (Opus 4.7: 79.3%) |
| Context window | 1M tokens | 1M tokens |
| Modalities | Text only | Text, image, video, desktop control |
| License | MIT | Open-weight (custom) |
| API price per 1M tokens (in / out) | $1.40 / $4.40 | $0.30 / $1.20 (launch promo) |
| Released | June 13, 2026 | June 1, 2026 |
The cost gap deserves concrete numbers because percentages hide it. Generating 100 million output tokens — a month of a moderately busy agent fleet — costs about $440 on GLM-5.2 and about $120 on MiniMax M3. The same volume on Claude Opus 4.8 is $2,500; on Claude Fable 5, $5,000. Even the open-weight capability leader is roughly 5x cheaper than the closed workhorse tier, and M3 is 3.7x cheaper again. This is the number that's pulling high-volume workloads toward open weights.
How this generation got here
Neither model came from nowhere. The spring generation — GLM-5.1 (April) and MiniMax M2.7 (March) — traded blows in the high 50s on SWE-bench Pro (58.4% vs 56.2%), with M2.7 pulling off its result using only 10 billion activated parameters, about 94% of GLM's coding performance at a fifth of the price. The June releases each doubled down on their existing bet: Zhipu pushed capability (GLM-5.2 gained four points of SWE-bench Pro and eighteen points of Terminal-Bench over 5.1), while MiniMax pushed scope and efficiency — 1M context, multimodality, and the MSA speedups. Both trend lines are steep, and neither company shows signs of slowing to a comfortable annual cadence.
When open weights are the right call — and when they aren't
- Pick GLM-5.2 when the workload is agentic coding or terminal automation and you want the most capable model you can self-host — the MIT license is as permissive as licenses get, and the Terminal-Bench score is within striking distance of closed workhorses.
- Pick MiniMax M3 when volume economics dominate, or when the workload needs eyes: document-image extraction, video understanding, browser and desktop automation. Nothing else open touches its BrowseComp score, and the price makes high-volume experimentation nearly free.
- Stay closed (for now) when the task rides the frontier — complex multi-step reasoning where GLM's 51 index score versus Fable 5's 60 shows up as real quality gaps — or when a vendor's compliance posture, uptime SLA, and safety evaluations are what your auditors want to see.
- The hybrid pattern that actually ships: open weights for the high-volume commodity steps, a closed frontier model on the escalation path, one routing layer in front of both. This is where most cost-conscious production systems land.
A candid note on operations: self-hosting a 1M-context open model is real infrastructure work — GPU capacity planning, inference-server tuning, monitoring, patching. Hosted API endpoints for both models remove that burden at prices that still embarrass the closed tier. Self-hosting earns its complexity when data can't leave your network or when utilization is high enough to beat the API price; otherwise start hosted.
GLM-5 vs MiniMax M3 — common questions
Are open-weight models actually ready for production in 2026?
For a substantial and growing class of workloads, yes — and June 2026 is the month the claim stopped needing caveats. GLM-5.2's 81% on Terminal-Bench 2.1 sits within four points of Claude Opus 4.8 and above GPT-5.5, and MiniMax M3 beats Claude Opus 4.7 outright on autonomous web browsing. Those aren't toy benchmarks; they're the evaluations closest to real agentic work. The remaining honest gaps: the open tier still trails the closed frontier by nine or more index points, which shows up on genuinely hard reasoning; vendor safety evaluations and uptime SLAs matter to auditors; and self-hosting is real operational work. The pattern we recommend to clients is hybrid — open weights for high-volume commodity steps where the 5-40x cost advantage compounds, closed frontier models on the escalation path for the hard steps, and a routing layer that makes the split invisible to the application.
Which is better for coding: GLM-5.2 or MiniMax M3?
GLM-5.2, and it isn't close on the agentic side. It leads SWE-bench Pro 62.1% to 59.0% — a modest gap — but the Terminal-Bench 2.1 spread is fifteen points, 81% to 66%, and terminal-driven work is where coding agents spend most of their time: running tests, chasing build errors, operating tooling. GLM also edges MCP Atlas tool-use, 77.0% to 74.2%. MiniMax M3's counterargument is economic and architectural: at $1.20 per million output tokens against GLM's $4.40, you can afford to run M3 three times — with a verification pass — for less than one GLM run, and its 15.6x decode speedup means iteration loops feel faster. For a primary coding agent, take GLM-5.2. For high-volume, lower-stakes code tasks — test generation, boilerplate, batch refactors with review — M3's economics are hard to argue with.
What does MiniMax M3 do that GLM-5.2 cannot?
See. GLM-5.2 is text-only; M3 natively handles text, images, and video, and it can operate a desktop environment. That difference defines entire categories of work: extracting data from scanned invoices and shipping documents, understanding screenshots in a support workflow, QA-testing a web app by actually looking at it, monitoring video feeds, driving legacy desktop software that has no API. M3 is also the stronger autonomous researcher — its 83.5% BrowseComp score beats not just every open model but Claude Opus 4.7 — and its OSWorld-Verified 70% makes it the most capable open computer-use model available. Add the 1M context window shared with GLM and the 3.7x output-price advantage, and M3 is less a cheaper GLM alternative than a different tool: GLM-5.2 is the best open coding engine; M3 is the best open perception-and-action engine.
Should we self-host an open model or use a hosted API?
Start hosted, and let two specific conditions pull you to self-hosting rather than defaulting to it. The conditions: data that genuinely cannot leave your network (regulatory or contractual, not just preference), or sustained GPU utilization high enough that owned or reserved hardware beats the per-token API price — which typically requires steady multi-million-token daily volume, not spiky experimentation. Self-hosting a modern 1M-context model is serious infrastructure: multiple high-memory GPUs, inference-server tuning, KV-cache management, monitoring, and someone on call when it degrades. Hosted endpoints for GLM-5.2 and MiniMax M3 deliver the same open-weight economics — still 5-40x cheaper than closed flagships — with none of that burden, and they preserve the strategic benefit that matters most: because the weights are open, you can move from hosted to self-hosted later without changing models, retraining prompts, or renegotiating with a vendor who knows you're locked in.
Do Chinese open-weight models pose a data or compliance risk?
Separate the two questions, because they have different answers. Data risk is an infrastructure question, not a model question: open weights are static files, and a model running on your own GPUs — or on a US or EU hosting provider you choose — sends nothing anywhere. That's the core advantage of open weights over closed APIs, where your data necessarily transits the vendor's servers. Using Zhipu's or MiniMax's own hosted APIs is a different posture and deserves the same vendor review you'd give any offshore data processor. Compliance is more situational: some regulated industries and government-adjacent contracts restrict models by origin regardless of hosting, licenses differ (GLM-5.2 is straight MIT; M3's open-weight license has custom terms worth a legal read), and export-control rules in this space have shifted more than once in 2026 — as Anthropic's own 19-day Fable suspension showed, this cuts in every direction. For most commercial buyers, self-hosted or western-hosted open weights clear both bars comfortably; check your specific regulatory surface before betting a flagship workload on it.