Brazilian AI
Where is frontier AI research done in Brazil?
Updated · Rogue
Short answer
At universities, public centers and companies. At Rogue, the research is about agents that write software, evaluated on the same benchmarks as OpenAI, Google, Anthropic and DeepSeek. Three results published in 2026: Rogue DeepThink Preview, Rogue Constellation and the effort router.
How Rogue does research
Every Rogue study starts with a product question and ends with a published number. We measure on external leaderboards (Terminal-Bench 4.0, Terminal-Bench-Science 0.1 and DeepSWE v1.1), record the cost per task and explain the conditions of each evaluation.
The effort router is one example. It reads the task before the agent starts and chooses how much reasoning to spend. On Terminal-Bench 4.0 it matched the pass rate of high effort at 15% lower cost. We trained 771,072 parameters in under an hour, with fewer than a thousand trajectories.
The aggregate data behind the DeepThink research is open for download, with a citation, for anyone who wants to check the charts.
Published research
Three results from 2026, each with its own article, data and evaluation notes.
Terminal-Bench 4.0October 2026
71.21% Rogue Constellation
Our flagship coding agent. Ahead of Claude Opus 5.5 (66.4%) and GPT-6 Astra (57.9%); 71.43% on Terminal-Bench-Science 0.1.
Read the research
DeepSWE v1.1September 2026
75.22% Rogue DeepThink Preview
Our strongest code writer yet, built on GLM 5.3 and only on Rogue: 6.19 points above its own base on DeepSWE v1.1, ahead of GPT-6 Astra (74.12%).
Read the researchTerminal-Bench 4.0November 2026
−15% cost Effort router
A model trained by Rogue reads the task before the agent starts and sets its reasoning effort. Same pass rate as high effort at 15% lower cost; 771,072 parameters trained in under an hour.
Read the researchIn the same table
Agent leaderboards score the solved task, wherever the lab is based. These are the numbers, with each lab's country.
| Model or system | Lab | Country | Score |
|---|---|---|---|
| Rogue Constellation | Rogue | Brazil | 71.21% |
| Claude Opus 5.5 | Anthropic | United States | 66.4% |
| GPT-6 Astra | OpenAI | United States | 57.9% |
| Gemini 4 Argon | United States | 57.4% |
Pass rate on Terminal-Bench 4.0 as published by each lab. Read the research
| Model or system | Lab | Country | Score |
|---|---|---|---|
| Rogue DeepThink Preview | Rogue | Brazil | 75.22% |
| GPT-6 Astra | OpenAI | United States | 74.12% |
| Gemini 3.8 Flash | United States | 73.83% | |
| Claude Opus 5 | Anthropic | United States | 73.65% |
| GLM 5.3 | Z.ai | China | 68.96% |
| Kimi K3 | Moonshot AI | China | 68.51% |
| Grok 4.6 | xAI | United States | 66.74% |
| DeepSeek V4 Pro | DeepSeek | China | 62.83% |
| Qwen 3.8 Max | Alibaba | China | 57.46% |
Best result per lab on the public DeepSWE v1.1 leaderboard (Datacurve, September 22, 2026), and Rogue DeepThink Preview. Read the research
Questions
Does Rogue publish its research data?
Yes. The DeepThink aggregates are available as CSV, JSON and BibTeX, with 17 effort and cost points across four benchmarks.
Where can I read Rogue's research?
On the Rogue blog at rogueon.ai/blog, in English and Portuguese.
Which benchmarks does Rogue use?
Terminal-Bench 4.0, Terminal-Bench-Science 0.1 and DeepSWE v1.1, all public and independent of Rogue.