Z.ai · Open weights
GLM-5.2
Long-horizon flagship
- Context
- 1M tokens
- Parameters
- 753B / 40B active
- License
- MIT
00 / Reading note
Scores are only comparable when the test is.
Benchmark versions, agent harnesses, reasoning effort, token budgets, and tool access can materially change a result. A dash means that no directly comparable figure was reported in the cited first-party release. Higher is better unless noted otherwise.
01 / Model cards
Three efficiency tiers, three different bets
DeepSeek · Open weights
V4 Flash
Fast reasoning model
- Context
- 1M tokens
- Parameters
- 284B / 13B active
- License
- MIT
OpenAI · Hosted
GPT-5.6 Luna
High-volume efficiency tier
- Context
- 1.05M tokens
- Max output
- 128K tokens
- Weights
- Proprietary
02 / Publisher-reported evaluations
Reported capability results
Percent unless noted
| Evaluation | GLM-5.2 | DeepSeek V4 Flash | GPT-5.6 Luna |
|---|---|---|---|
| GPQA Diamond | 91.2publisher setup | 88.1Max · Pass@1 | 92.3publisher setup |
| SWE-bench Pro | 62.1OpenHands | 52.6Max · resolved | 62.7publisher setup |
| Humanity's Last Exam | 40.5text-only | 34.8Max · Pass@1 | —not reported in release table |
| Terminal-Bench | 81.0v2.1 · Terminus-2 | 56.9v2.0 · Max | 84.7v2.1 |
| BrowseComp | —not reported | 73.2Max · Pass@1 | 83.3publisher setup |
| MCP-Atlas public set | 76.8thinking | 69.0Max · Pass@1 | —not reported |
| DeepSWE | 46.2mini-swe-agent | —not reported for Flash | 67.2v1.1 |
Terminal-Bench uses different versions for DeepSeek and the other two models. SWE-bench Pro harnesses also differ. These rows are preserved as reported records, not presented as controlled head-to-head experiments.
03 / Independent snapshot
Artificial Analysis, comparable reasoning tiers
Intelligence Index v4.1
GLM-5.2 max51
GPT-5.6 Luna high46
DeepSeek V4 Flash max40
Observed output speed
179 t/s
GLM-5.2 max, provider median snapshot
161 t/s
GPT-5.6 Luna high
116 t/s
DeepSeek V4 Flash max
Published API price
GLM-5.2$1.40 / $4.40
DeepSeek V4 Flash$0.14 / $0.28
GPT-5.6 Luna$1.00 / $6.00
USD per 1M input / output tokens. GLM figure is the independent provider median; DeepSeek and OpenAI figures are publisher prices.
04 / What the records suggest
There is no single winner—only a clearer workload fit.
GLM-5.2 leads this independent intelligence snapshot and posts strong agentic results while remaining open-weight.
DeepSeek V4 Flash is the price outlier: materially cheaper, smaller per-token activation, and competitive at Max effort.
GPT-5.6 Luna combines strong publisher-reported coding and browsing scores with a hosted, high-volume product profile.
05 / Sources
Primary records and independent methodology
- Z.ai · GLM-5.2 official model card
- DeepSeek · V4 Flash official model card and technical report
- OpenAI · GPT-5.6 launch and evaluation tables
- OpenAI · GPT-5.6 Luna model reference
- Artificial Analysis · Intelligence, speed, and pricing methodology
Compiled 05 August 2026 · Figures may change as providers update models, harnesses, and pricing.