OpenAI announced Presence on July 22, 2026, a product designed to deploy AI agents for specific enterprise workflows in voice and chat. OpenAI says Presence combines model reasoning with policies, guardrails and an improvement loop powered by Codex.
OpenAI published a field report (July 28, 2026) on eight agent-assisted scientific computing projects, mainly in the life sciences. The report finds agents accelerate development but human validation and long-term stewardship remain necessary.
OpenAI says GPT‑5.6 Sol balances capability and cost, outperforming Claude Fable 5 on the Artificial Analysis Coding Agent Index at under half the cost. The company reports kernel, routing, caching, and agentic-harness changes that cut serving costs and raised token efficiency.
Textly released a macOS utility that captures text and tables from images, videos, PDFs and apps while keeping table rows and columns. The app exports to CSV, Markdown and TSV, offers instant translation and searchable clipboard history, and launched March 20, 2025.
avatarin partnered with Yamada Holdings to deploy the Kurashi-Marugoto AI Agent on Yamada Denki’s online store, using OpenAI’s GPT‑Realtime for multilingual voice and text support. In a two-week public campaign about 30,000 users engaged and 92% of survey responses were positive.
OpenAI says GPT-5.6 Luna will be 80% cheaper and Terra 20% cheaper, and introduces Fast mode for GPT-5.6 Sol with up to 2.5× speed at twice the price. Changes affect ChatGPT Work, Codex and the API, with updated token rates and rollout beginning July 30, 2026.
Circles built an AI Concierge on the OpenAI API to link support, personalization and operations in real time. The company reported a 22% ARPU uplift and 9% lower churn in Singapore, plus a 65% autonomous resolution rate from its CareX multi-agent system.
OpenAI released GPT-6.1 Terra, a model aimed at visual reasoning, computer-based tasks and clearer writing. OpenAI reported improved benchmark scores and lower error rates versus GPT-5.6 Terra, and made the model available in ChatGPT Work, Codex and via API.
OpenAI introduced BioBasicsBench, a 100‑task benchmark created with Ligo to test AI agents on raw wet‑lab data and error detection. At the ‘xhigh’ setting, GPT‑6 Astra scored 66.7% versus 43.8% for GPT‑5.6 Sol, according to OpenAI’s report.
ToyBlocks is an open simulation framework presented as the first FoundationWorld for controlled robot learning. The team reports automatic generation of demonstrations, datasets up to 24× real time on one GPU, and evaluations of Molmo 2, ManiFlow and frontier models.