On November 6, 2026, China Mobile Cloud and Zhipu Technology signed a cooperation agreement covering model services, AI coding products and industrial applications. Chinese media picked it up two days later: Zhipu's GLM series is slated for integration into China Mobile Cloud's MoMA model service platform, with a joint R&D center on token efficiency to follow. It did not name versions, publish timelines or show a single benchmark number.
That gap matters more than the press release, because GLM is no longer one product. As of mid-November 2026 there are three models a team can realistically deploy: GLM-5.3, GLM-5.3 Flash and GLM-4.7 Flash. Three price points, three licenses, three jobs. Here is the scorecard for whichever one a carrier channel, or your own pipeline, actually wants.
What the China Mobile Deal Actually Confirms
The agreement is a partnership announcement, not a launch. The reported facts are narrow, and it is worth separating what was signed from what coverage implied.
- Signed November 6, 2026, reported by Chinese media on November 8.
- Scope: model services, AI coding products and industrial applications.
- GLM is to be integrated into China Mobile Cloud's MoMA platform, with no confirmed version list or live status.
- A joint R&D center on token efficiency is planned, aimed at optimizing models for domestic computing hardware.
- No efficiency targets, no timetable, no verifiable performance data published.
The strategic logic still comes through. A carrier channel is how enterprise AI gets bought in China, so MoMA integration is a distribution win. The token-efficiency center is the more revealing item: the fight has moved from raw scores to cost per unit of useful output on domestic silicon.
The Scorecard: Capability, Cost, Control
Round 1 - Raw capability. Independent numbers are the only ones worth ranking. Artificial Analysis Intelligence Index v5.0.1, read on November 12, 2026, gives a clean separation.
- GLM-5.3 (max): 31.7 on the index. Same 744B base as GLM-5.2, 40B active parameters, 1M-token context, text input only, up to 128K output tokens.
- GLM-5.3 Flash: 27.8. A separate 320B multimodal model, 18B active parameters, taking video, image, text and file input at a 1M-token context.
- GLM-4.7 Flash: not competitive here. A 30B mixture-of-experts model, roughly 3B active parameters and about 200K context, built for autocomplete, refactors and short scripts.
Zhipu's own GLM-5.3 chart is far more flattering: Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, CyberGym from 77.2 to 84.5 percent, on the same base model. Even at face value it sits more than twenty points below Claude Opus 5.5 at 59.1 and GPT-6.1 Sol at 54.6. Verdict: GLM-5.3 wins capability, and it is still not frontier.
Round 2 - Cost per job, not per token. The rate card looks like a rout for the cheap models, and on arithmetic it is. Published Z.ai prices, US dollars per million tokens:
- GLM-5.3: $1.40 input, $0.26 cached input, $4.40 output.
- GLM-5.3 Flash: $0.15 input, $0.03 cached, $0.50 output. FlashX sits at $0.37 and $1.25.
- GLM-4.7 Flash: free on the Z.ai API. Paid GLM-4.7-FlashX runs $0.07 input and $0.40 output.
Then the trap. Artificial Analysis spent $5.01 per index task on GLM-5.3 against $1.27 on GPT-6.1 Sol, and $4.13 of that GLM bill was cached input read again and again. GLM-5.3 charges $0.26 per million cached tokens against $0.10 on GPT-6.1 Sol, and re-read roughly six times as much context. Cheap per token is not cheap per job. Flash came in at $0.58 per index task. Verdict: Flash wins cost on both the rate card and the measured bill.
Round 3 - Control and licensing. Here the three stop being interchangeable. GLM-5.3 Flash ships MIT weights, so fine-tuning and self-hosting are unrestricted. GLM-5.3 broke the family pattern: it launched on August 14 without weights, published them on August 28 under the GLM-5.3 License, and that license adds a security review before commercial use by any group earning over $10 billion a year that runs a model-as-a-service business. GLM-4.7 Flash is the outlier in the other direction, small enough for single-GPU deployment with vLLM and SGLang support. Verdict: split decision, MIT for products and the custom license for scale.
Verdicts and a Deployment Shortlist
Three winners, no single champion. GLM-5.3 is the one to rent or self-host when the task is long-horizon agentic coding or vulnerability discovery and a wrong answer is expensive. GLM-5.3 Flash is the default for everything with a budget. GLM-4.7 Flash is the offline and free-tier option, and nothing more.
For a carrier that implies obvious sequencing: Flash behind consumer and internal-assistant traffic where volume dominates, GLM-5.3 for coding and security work that justifies a premium, the 30B model as the on-premise escape hatch.
The subscription math reinforces it. The GLM Coding Plan runs $18 for Lite, $80 for Pro and $168 for Max per month, with 30 percent off annual billing. Credits are metered rather than prompts: 2,000, 12,000 and 28,000 per five hours, or 10,000, 60,000 and 140,000 per week. Usage outside the peak window of 14:00 to 18:00 Singapore time on weekdays costs half the credits.
What would flip these verdicts? Three things. GLM-6.0, which founder Tang Jie named on the August 31 results call as a full self-training build with no release date, could reset the capability column. A published efficiency number from the joint center would rewrite the cost column. And a confirmed MoMA version list would show whether the carrier channel is selling the flagship or the Flash tier.
The honest bottom line is that GLM's case was never the leaderboard. It is weights you can hold, a price you can forecast and a deployment you can move to your own silicon when a vendor relationship sours. GLM-5.3 gives the strongest version of that argument and the worst license terms; Flash gives the cheapest and the cleanest. Anyone signing a carrier contract this quarter should ask which of the two they are actually getting.
Comments