News

OpenAI taught AI to rewrite legacy programs: up to 97.9% accuracy in its own test.

In particular, the model successfully migrated 86.6% of COBOL programs to other languages at a low reasoning level.

COBOL runs more than 40% of online banking systems and 95% of ATM transactions.

The previous model managed only in 5.4% of cases.

Nvidia released an open 120-billion-parameter model that, after fine-tuning, outperformed Claude Opus 5 on one task.

According to Artificial Analysis, it is 5 times faster than its closest competitor in its class.

Without fine-tuning, the base model did not beat the closed flagships. For general everyday tasks, closed models like Opus 5.5 are more convenient because they work without any setup.

Open models have their own advantages: they can be fine-tuned on your own data and run on your own servers, and they are fast for large volumes of repetitive work.

Meta entered the "frontier": the company released Muse Wonder and Muse Wonder Full, which, according to it, compete with the best AI models in the world, including those from OpenAI and Anthropic.

Key features: Active Memory (remembers you between conversations) and Active Focus (doesn't drop a task, but tries other approaches and asks before important actions).

Meta itself published the test results in its announcement; there is no independent verification yet. We're waiting for real-world tests.

Free Gemini no longer lets you choose a model: now Google decides for you. Accounts have been switched to "Auto" mode.

The manual model picker has disappeared, and only one option remains in the list: "Auto" with "Gemini 3 models". Most requests will be routed to Flash-Lite, the fastest and lightest one.

Everyone has also gotten "effort" levels (Low / Medium / High) instead of the old "Extended."