New AI models, more affordable coding, and a replacement for Git
The time for Riihisoft's weekly AI training!
Time for Riihisoft's weekly AI training! This week, we explored Claude Fable 5.2, GPT-6.2, GitHub Copilot Fusion, and the new version control of Anthropic.
Fable 5.2 surpasses Opus 5.5 in coding
Anthropic's new Fable 5.2 solved 69.9% of tasks in the FrontierSWE v2 test (Opus 5.5: 62.3%, Fable 5.1: 56.3%). The model is 30% faster than 5.1 and costs $8 / $40 per million tokens. The unrestricted Mythos 5.2 is only for approved security and medical operators. For everyday work, Anthropic still recommends Opus 5.5 and Fable for the most difficult tasks.
OpenAI released GPT-6.2: Astra dropped by 20%
The price of GPT-6.2 Astra's responses dropped from 50 dollars to 40 dollars per million tokens, reaching Fable 5.2. The lighter Sol costs only 10 dollars and lags 0.5 points behind Astra in the coding test, so it is enough for many tasks. Some testers have criticized the models, so try them on your own tasks.
Copilot Fusion rose to the top of the coding test
In the Terminal-Bench 4.1 test, Copilot Fusion Max solved 71.2% of the tasks (Opus 5.5 and Fable 5.2: 65.8%). Fusion is 88% cheaper than Fable 5.2 and 57% cheaper than Opus 5.5. You just select the level (Standard, High, or Max) and Copilot selects the models. Fusion works in VS Code, JetBrains, and Copilot CLI.
Fable 5.2 made the operating system in 8 days
In the experiment, over 1,000 agents controlled by Fable built a working operating system from scratch in 8 days. It boots on a real laptop and has its own kernel, drivers, windows, and installer. It is an experiment, not a finished product: the model divides the project into parts and guides the agents, and a human sets the goal and checks the result.
Anthropic replaced Git with its own system
Changes occurred 12 times more per day, and the repo copy contained 100 gigabytes of history, so Git was no longer enough. The new Scion retrieves only the necessary files, and the server handles history and merges. Server requirements are 200 times smaller and code retrieval is 100 times faster. Scion is made for huge monorepos, not as a general replacement. As agents increase, the bottleneck first arises for testing, reviewing, and merging, not for writing code.
From ideas to effective solutions
In our AI trainings, we do not focus solely on new models and technologies but also cover concrete AI-based implementations, proof of concept (POC) solutions, development practices, and work processes. This week, Janne Huuhtanen , Head of Partner Success at Plentics, presented the Billing Analyzer application, which he has effectively implemented using artificial intelligence. Billing Analyzer is a billing support tool. It compares CSV billing data over months. This way, trends in the number of devices in the endpoint management service's customer environments and devices can be observed.
Contact us, and let's turn your ideas into solutions!