GLM-5.3 vs DeepSeek-V4-Pro: the open-weight value war

Zhipu and DeepSeek are fighting the same war from opposite ends. Both ship open weights. Both carry a million tokens of context. Both land within 2.5 points of each other on our capability index. And one of them charges 3.5 times more than the other for the privilege.

GLM-5.3 landed at $1.20 in and $4.00 out: a blended $1.90 per million tokens. DeepSeek-V4-Pro (the 0813 build) runs $0.435 in and $0.87 out: a blended $0.54. Same license shape, same context window, same weight class: 62.5 versus 60.0 on capability.

The scoreboard could not separate them either. GLM posts 51.9 reasoning, 42.8 coding, 39.6 agentic. DeepSeek answers with 50.1, 40.7, 37.9. Every category, a margin too small to build a decision on.

• Reasoning and coding: GLM by a whisker (51.9/42.8 vs 50.1/40.7). Not a decision.
• Throughput: GLM, 2.3x. The only category where the flagship pitch holds.
• Cost at scale: DeepSeek, 3.5x cheaper at blended rates. The volume argument ends here.
• The wildcard: GLM's own Flash tier: 55.3 capability at $0.24 blended, a 232.8 value score. It undercuts both flagships and its own big brother.

specs

	glm-5.3	deepseek-v4-pro	glm-5.3-flash
org	zhipu ai	deepseek	zhipu ai
capability (hle-based)	62.5	60.0	55.3
price in / out	$1.20 / $4.00	$0.44 / $0.87	$0.15 / $0.50
blended	$1.90	$0.54	$0.24
value score	32.9	110.3	232.8
context	1m	1m	1m
throughput	147 tok/s	65 tok/s	81 tok/s
license	open weights	open weights	open weights

The capability gap is 2.5 points. The price gap is 3.5x. One of these numbers is a rounding error.

where deepseek wins

The value math is not close. A month of agentic work at 300 million input tokens, 60 million out, and 270 million cache reads lands DeepSeek at roughly a quarter of GLM's invoice, and the capability you give up is 4% on our index. Per-task economics tell the same story: any workload where tokens are the cost center routes to DeepSeek until GLM's throughput premium buys back more than it costs.

The 0813 build also ships the pattern DeepSeek has run all year: ship the checkpoint, undercut the market, let the benchmarks catch up. Our compute-stack piece covered the infra thesis; the pricing here is the same play with the same discipline.

where glm wins

Throughput first. 147 tokens per second against 65 is the difference between an agent that feels like a terminal and one that feels like a form. For interactive coding, chat surfaces, and anything a human watches token by token, GLM's speed is a product feature the value score cannot price.

The second argument is the headroom: 62.5 capability is the stronger starting point for anything that compounds on model quality, and Zhipu's own cybersecurity report reads like a lab that knows where its gaps are and publishes them anyway. That is rarer than it sounds.

the flash wildcard

Here is the part both pricing pages hope you skip: GLM-5.3-Flash scores 55.3 on our index at a blended $0.24. That is a 232.8 value score: seven times DeepSeek's, and it sits inside GLM's own lineup, undercutting its flagship by 8x on price for a 7-point capability trade.

If the workload tolerates the drop, the Flash tier is the correct answer for most volume. The flagship-vs-flagship framing only survives if your tasks genuinely need 60+ capability, and most do not.

the fine print

Our capability numbers lean on HLE-sourced scoring, which weights hard-reasoning benchmarks; the gpqa and swe columns are empty for all three, so coding and agentic depth come from the aggregator medians rather than our index. The benchmark picture is one referee among several, and the aggregators disagreed about bigger models than these.

Both models ship open weights, which means the real floor is whatever inference stack you run them on. List prices are the vendors' API rates; self-hosting changes every number in the table.

And the volume math assumes cache-friendly agent workloads: the moment your traffic is human-shaped, throughput starts pricing back in.

Default DeepSeek for volume, GLM for latency, Flash for everything in between. The capability spread here is too small to pay 8x for.