AI Daily — July 10, 2026
2026-07-10
openai.com
GPT-5.6 released with improved efficiency and M365 Copilot integration
OpenAI released GPT-5.6, positioning it as a more compute-efficient frontier model with higher capability-per-token and better cost scaling. Simultaneously, GPT-5.6 becomes the default model powering Microsoft 365 Copilot across Word, Excel, PowerPoint, and collaborative features. The rollout marks another step in OpenAI's rapid minor-version cadence following GPT-5 earlier this year.
openai.com
ChatGPT Work: OpenAI launches long-horizon agentic product for enterprise tasks
OpenAI introduced ChatGPT Work, an agent mode that can operate across files and third-party apps for hours on a single goal, targeting complex multi-step professional workflows. It integrates with existing app ecosystems and is designed to sustain context and take actions across sessions rather than single-turn completions. This represents OpenAI's most direct push into autonomous productivity-agent territory.
openai.com
OpenAI launches GPT-5.5 Bio Bug Bounty program
OpenAI opened a dedicated bug bounty program focused on biological risks associated with GPT-5.5, inviting researchers to probe the model for unsafe biosecurity-relevant outputs or capability gaps in safety mitigations. The program signals OpenAI's effort to systematically red-team biological risk vectors before broader deployment. It is one of the first public bounty programs scoped specifically to a biosafety threat model.
www.technologyreview.com
Anthropic's Jacobian lens gives clearest view yet of LLM internal reasoning
Anthropic developed a mechanistic interpretability tool called the Jacobian lens that tracks how intermediate representations evolve as a model processes a prompt, revealing a latent 'reasoning space' where the model appears to iteratively resolve conceptual uncertainty before producing output. The findings include both expected behavior and results researchers describe as unnerving, though details remain limited in the public coverage. This is described as the most granular window into LLM cognition the company has achieved to date.