We’re introducing Rogue DeepThink Preview,
a method for solving harder problems with large language models. Built on GLM 5.3, DeepThink uses additional computation at test-time to tackle difficult software-engineering tasks.
Results
| Benchmark | DeepThink Preview | GLM-5.3 | Kimi K3 | Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 51.5% | 41.8% | 21.5% | 57.9% | 58.2% |
| FrontierSWE v2 | 36.8% | 30.2% | 25.9% | 56.3% | 65.5% |
| Terminal-Bench-Science 0.1 | 11.4% | 8.1% | 7.1% | 40.0% | 64.6% |
| DeepSWE v1.1 | 75.2% | 69.0% | 68.5% | 67.4% | 74.1% |
How it works
DeepThink uses GLM 5.3 as a base model and spends additional test-time compute on parallel thinking: exploring different approaches to the same problem at once.
On difficult tasks, deciding how to approach the problem is itself part of the work. An initial assumption can shape an entire line of reasoning, even when the subsequent steps are sound. Exploring other approaches creates opportunities to uncover an overlooked constraint or find a solution that the initial framing would have missed.
How it scales
We compared DeepThink with GLM 5.3 at its highest reasoning setting, max, across four benchmarks. DeepThink’s low, medium, high, and max settings use increasing amounts of test-time computation beyond that baseline.
Terminal Bench 4.0
How much is enough
Coding agents often expand a task as they solve it. A bug fix becomes a refactor; a small feature acquires a new framework. Even when the requested behavior works, the user inherits additional code to understand, review, and maintain.
Knowing when to stop is part of good engineering. Additional reasoning should help an agent judge what a specification requires, what can be left alone, and when the work is complete.
We examined submitted patches on 77 matched DeepSWE v1.1 tasks, measuring both added lines and complete patch bytes. This lets us compare how solution size changes across Fable-5’s reasoning settings and across DeepThink’s low, medium, and high settings, with GLM 5.3 at max as its baseline.
Mean added lines
DeepSWE v1.1 · 77 matched tasksIncreasing reasoning effort changes how much code Fable-5 produces: its mean added lines more than double from low to max. DeepThink’s output stays nearly constant from low through high, in both added lines and patch bytes.
DeepThink starts from a larger GLM 5.3 baseline and adds modestly to it. The distinction is not smaller patches, but that further increases in effort do not induce even larger solutions.
Knowing when the work is done
A specification leaves many decisions unstated. A coding agent has to determine which parts of the codebase need to change, which constraints must be preserved, and which apparent improvements fall outside the request. Solving the problem includes knowing where to stop.
DeepThink improves benchmark performance as we give it more test-time compute. In our separate patch-size study, increasing its effort leaves the size of its changes nearly constant. Together, these findings point toward something we want from coding agents: more thought invested in a change, without a growing implementation for the user to manage.
At Rogue, we want people to build software by describing what it should do. That requires agents we can trust with the decisions we haven’t specified - not just the instructions we have. DeepThink Preview is available to everyone today at app.rogueon.ai/build, and we hope you enjoy it!
Our coding harness for desktop will be available in the coming weeks...
Public data and citation
Download the author-reported aggregates behind the two interactive figures. This package preserves the plotted values and does not include the cross-model comparison table or current leaderboards.
Version 1.0.0 · packaged October 4, 2026
Benchmark effort and cost
17 reported points across four benchmarks, including the GLM 5.3 baseline. Cost in USD per task; score in percent.
Download effort and cost CSVSubmitted patch size
9 mean measurements on 77 matched DeepSWE v1.1 tasks. Added lines and complete patch bytes.
Download patch-size CSVDownload both tables as JSONDownload BibTeX citation
View all 17 effort and cost points
| Benchmark | Model | Effort | Cost (USD/task) | Score (%) |
|---|---|---|---|---|
| Terminal Bench 4.0 | GLM 5.3 (baseline) | max | 8.27 | 42.1212 |
| Terminal Bench 4.0 | Rogue DeepThink Preview | low | 16.6344 | 44.6212 |
| Terminal Bench 4.0 | Rogue DeepThink Preview | medium | 24.9996 | 46.8939 |
| Terminal Bench 4.0 | Rogue DeepThink Preview | high | 33.3968 | 49.0909 |
| Terminal Bench 4.0 | Rogue DeepThink Preview | max | 41.826 | 51.5152 |
| FrontierSWE v2 | GLM 5.3 (baseline) | max | 97.22 | 30.1814 |
| FrontierSWE v2 | Rogue DeepThink Preview | low | 194.5344 | 33.5405 |
| FrontierSWE v2 | Rogue DeepThink Preview | medium | 291.8496 | 34.9002 |
| FrontierSWE v2 | Rogue DeepThink Preview | high | 389.1968 | 35.7649 |
| FrontierSWE v2 | Rogue DeepThink Preview | max | 486.576 | 36.7775 |
| Terminal Bench Science 0.1 | GLM 5.3 (baseline) | max | 32.39 | 8.0952 |
| Terminal Bench Science 0.1 | Rogue DeepThink Preview | low | 64.8744 | 10.95 |
| Terminal Bench Science 0.1 | Rogue DeepThink Preview | medium | 97.3596 | 11.4286 |
| DeepSWE v1.1 | GLM 5.3 (baseline) | max | 3.99 | 69.0265 |
| DeepSWE v1.1 | Rogue DeepThink Preview | low | 8.0744 | 72.0501 |
| DeepSWE v1.1 | Rogue DeepThink Preview | medium | 12.1596 | 72.2345 |
| DeepSWE v1.1 | Rogue DeepThink Preview | high | 16.2768 | 75.2212 |
View all 9 patch-size means
| Model | Effort | Mean added lines | Mean complete patch bytes |
|---|---|---|---|
| Fable-5 | low | 548.51 | 22,907.18 |
| Fable-5 | medium | 782.51 | 32,251.45 |
| Fable-5 | high | 981.26 | 40,305.37 |
| Fable-5 | xhigh | 1,189.16 | 49,763.23 |
| Fable-5 | max | 1,404.17 | 58,510.24 |
| GLM 5.3 (baseline) | max | 1,567.72 | 61,816.77 |
| Rogue DeepThink Preview | low | 1,639.33 | 64,855.81 |
| Rogue DeepThink Preview | medium | 1,642.19 | 65,040.32 |
| Rogue DeepThink Preview | high | 1,632.85 | 64,850.79 |
These are author-reported chart aggregates, not a reproduction package. Task-level data, prompts, configurations, logs and the evaluation harness are not included. Missing points are omitted, not treated as zero or extrapolated.
Costs include inference and additional processing; they are not product prices or measured latency. FrontierSWE v2 scores are mean fractional reward, not a binary task pass rate. The 77-task patch-size study is separate from the 113-task DeepSWE score evaluation and has no matched accuracy measurements.
Cite this data
Rogue (2026). Rogue DeepThink Preview: public chart aggregates. Version 1.0.0, packaged 2026-10-04. https://rogueon.ai/blog/rogue-deep-think-preview/tecnica#public-data
Benchmark definitions
These primary sources describe the benchmarks and their versions.