Results depend on how a model is run, not just on the model. Here's what Redrob adds around it, and how we measured the difference.
Three kinds of AI app
Web chatA city carYou upload the document and copy the answer out yourself. It can't open folders on your computer.For example claude.ai, chatgpt.com
Desktop work appA delivery van with a driverIt pulls material from the folders and apps you connect, and sticks with work for hours until the document or deck is done.For example Claude Cowork, ChatGPT Work
Developer toolA workshop truck with a mechanicIt reads and edits the working folder itself, and runs commands to check its own results.For example Claude Code, Codex
As of October 2026, from each vendor's documentation: Claude Help Center "Claude Cowork architecture overview", OpenAI "Introducing ChatGPT Work" (9 July 2026), Claude Code docs, OpenAI Codex docs.
The engine developers use. Now you just talk to it.
Type what you need, the way you always do. Behind the scenes, Redrob Code works like a developer tool: it reads the files it needs, builds the document and checks its own work before handing it over. No commands, no setup.
1You ask in plain words
2Redrob Code does the work
3You get work ready to review
Go deeper: harnesses and benchmarks
A harness is the software that wraps an AI model and runs it. It decides which tools the model can use, how it reads and writes files, and how it breaks work into steps and checks itself along the way. If the model is the engine, the harness is the body, the gearbox and the dashboard. Redrob Code is a harness, and the desktop app runs on it.
The research says the same thing. In one study of 100 coding tasks, keeping the model and changing only the harness moved scores by 8.5-13 points, while keeping the harness and changing only the model moved them by 2.5-5. Independent lab Epoch AI found the same model swinging by up to 11-15 points on harness alone, and called the harness the single biggest factor in performance.
One caveat: all of these numbers come from coding tasks. No public study has run the same comparison on legal or accounting documents yet. That's why Redrob measured document quality in its own internal evaluation. The results are further down this page.
1. Redrob internal evaluation: same model, same effort setting, same prompt, compared with a general model.
2. arXiv 2605.23950, "Stop Comparing LLM Agents Without Disclosing the Harness" (May 2026). On 100 coding tasks, the spread (variance) of scores when only the setup around a model changed was 7.8 times the spread when only the model changed. In score points: 8.5-13 from the setup, 2.5-5 from the model.
Sources and method
Where the numbers on this site come from.
Desktop app about +8%, coding agent app about +19%
Gains over web chat. They are estimates from public benchmarks that run the same AI model with and without tools, or in different setups, with estimated ranges of +3% to +15% and +8% to +33%. Sources: Anthropic's Claude Opus 5 system card (July 2026; GDP.pdf, Humanity's Last Exam), the Claude Opus 4.6 system card (February 2026; MMMU-Pro), Scale AI SWE-Bench Pro with Augment Code's same-model runs (February 2026), and arXiv 2605.23950 (May 2026). None of these measures legal or accounting document work, or web chat products themselves.
Task-built skills about +27%, two or three of your firm's own templates about +48%
Task-built skills (about +27%) and two or three of your firm's own templates (about +48%) are shown as the gain over web chat in Redrob's internal evaluation on several frontier AI models, same model and same prompt each time. Your results depend on your work and how you use Redrob.
About 20% better work, even on frontier models
Redrob's internal evaluation on several frontier AI models, same model and same prompt, with and without the skill.
100 GB, about 1 million pages
Page count assumes about 100 KB per page, typical for scans and office documents. Actual capacity depends on file types.
