Brazilian AI
Rogue is a Brazilian artificial intelligence company. We research agents that write software, and we publish every result on the public benchmarks that labs in the United States and China use, with score, cost and conditions.
Short answer
Rogue is a Brazilian AI company that does research and builds product. Its research appears in the same tables as OpenAI, Google, Anthropic, Z.ai, Moonshot AI and DeepSeek: Rogue Constellation scored 71.21% on Terminal-Bench 4.0, and Rogue DeepThink Preview scored 75.22% on DeepSWE v1.1. The same research powers app.rogueon.ai, where you build websites, web apps, slides and games.
Every feature that reaches the product starts as an experiment with a published score, cost and conditions.
We measure on independent benchmarks such as Terminal-Bench and DeepSWE, where the frontier labs are measured too.
A better answer at twice the price is not always better. Each study reports cost per task next to the score.
The team works from Brazil, the product speaks Portuguese and English, and the research ships in both.
Three results from 2026, each with its own article, data and evaluation notes.
Terminal-Bench 4.0October 2026
Five agents on every task, returning the solution with the strongest evidence that it works. Ahead of Claude Opus 5.5 (66.4%) and GPT-6 Astra (57.9%); 71.43% on Terminal-Bench-Science 0.1.
Read the researchDeepSWE v1.1September 2026
Deeper reasoning at test time on top of GLM 5.3: 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
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 researchAgent 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. On the official leaderboard (tbench.ai, checked November 6, 2026), the leader is Claude Code with Opus 5.5 at 64.8%. Rogue ran all 66 tasks, with a standard error of about 5.5 points. 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's internal evaluation. DeepThink Preview builds on GLM 5.3. Off the leaderboard, Meta reports 75.4% for Muse Spark 1.3 and Anthropic 74.2% for Claude Opus 5.5. Read the research
Who does what, based on what each company publishes about itself.
| Company | Type | What it does | Source |
|---|---|---|---|
| Rogue | Agent research and product | Agents that build websites, web apps, slides and games, plus Rogue Code for Mac. Research published with score, cost and conditions on Terminal-Bench and DeepSWE. | Rogue research |
| Maritaca AI | Portuguese language models | The Sabiá family (Sabiá 4, Sabiazinho 4, Sabiá 4 Thinking) and S5, specialized in Portuguese and Brazilian contexts, with a focus on legal work. Its technical report positions the models on cost-performance. | maritaca.aiSabiá-4 report |
| WideLabs | Portuguese language models | Maker of Amazônia IA, a language model built in Portuguese with Brazilian cultural data, delivered as full-stack AI projects, mostly for the public sector. | Mobile Time, Feb 2026 |
| Adapta | AI hub | Presents itself as “all the best AIs in one place”: ONE gives access to other companies' models, such as ChatGPT, Claude and Gemini, in one subscription with courses; Skip builds software without code. Its site lists no model of its own and no research. | adapta.org |
| SoberanIA | Public project | A model trained in Portuguese, launched in December 2025 by the Ministry of Science, the Government of Piauí, Telebras and partners, as part of the Brazilian AI Plan. | DPL News, Dec 2025 |
Sources checked on November 6, 2026.
Research
OpenAI, Anthropic and Google lead among model labs, with Z.ai, Moonshot, DeepSeek and Alibaba from China. Rogue, from Brazil, has the top published score on two agent tests and sits above the public leaderboard on the third.
Research
Who competes on international leaderboards, who published new research this year and who is most cited and downloaded worldwide. Five labs, with the rule and the sources in the open.
Product
Describe the game and Rogue delivers a playable version in the browser. The games page has three games built by Rogue Constellation to play right now.
