«Вастрик.Ай» digest for the previous 24 hours

The main topic was the release of GPT-6 Sol/Luna and hands-on tests. In the BitGN benchmark, Sol (especially high) significantly shifted the price/quality/speed balance and completed the AI coding set at 100%; Luna looks like a new inexpensive workhorse for simple tasks. There was no unanimous enthusiasm, however: some participants found Sol 6 verbose, prone to overengineering, and a regression from Sol 5.6. Some subjectively ranked Opus 5.5 above Sol and Astra, although it costs more and refuses more often because of safety policies. A practical suggestion from the discussion: use Luna for easy tasks and Sol as the main option; tune reasoning, and dynamically lowering it during routine CLI steps may save budget.

Participants discussed Codex limits and “banked reset”: resets have indeed started arriving, but it is unclear whether they change the date of the regular reset. There were complaints about model availability in corporate Claude accounts and unpredictable subscription limits.

Useful links: OpenAI on GPT-6 prompt caching — https://openai.com/index/better-prompt-caching-for-gpt-6/; ways to access Jev through OpenRouter and RouterAI. Participants also proposed a wrapper around Jev with projects, datasets, prompt versioning, evaluation, and caching; separately, they noted a case of Jev as a reranker for classifying 78,000 product codes.

Alongside that, they debated whether local hardware for large models pays off (the consensus was that a home build usually loses to subscriptions), and discussed local CI/CD in which two AIs check PRs, conformity to the description, and tests.