OpenAI released an update to GPT‑Rosalind on June 3, 2026, combining GPT‑5.5 agentic coding with improved performance in medicinal chemistry and genomics. OpenAI reports GPT‑Rosalind leads on the new LifeSciBench; availability expanded Sept 11, 2026 and pricing starts Oct 5, 2026.
HOPE is a second-order expert-pruning method that accounts for pairwise interactions between experts in Mixture-of-Experts (MoE) layers. The authors report that HOPE improves benchmark ranks—especially at 40–50% pruning—and provide a three-step calibration, solve and prune workflow.
Endava has embedded OpenAI tools across its software delivery process, creating DavaFlow, an AI-native methodology used by developers and non-technical teams. CTO Matthew Cloke says agents now run in daily workflows to speed planning, reduce manual work, and enable internal apps.
Nextdoor reports that Codex lets product engineers own features end‑to‑end, speeding delivery and shifting focus to outcomes. Cory Dolphin says Codex helps debug complex issues and that upgrades in GPT‑5.4 and 5.5 improved persistence and problem solving.
OpenAI says Oracle customers can apply eligible Oracle Universal Credits to purchase access to OpenAI models and Codex through Oracle Marketplace. Updated Aug 7, 2026: purchases via the API, Codex and ChatGPT Work are available in the United States.
OpenAI announced the OpenAI Partner Network on June 14, 2026, a program to help enterprises deploy frontier models with global partners. The company said it will invest $150 million and aims to certify 300,000 consultants by the end of 2026.
OpenAI describes Deployment Simulation, a method that replays de‑identified conversations to estimate undesired behavior before release. In tests on GPT‑5‑series Thinking models, it analyzed about 1.3 million conversations and reported a median error of 1.5× in rate estimates.
Immunologist Derya Unutmaz used GPT‑5 Pro to reanalyze a 2022 experiment and identified a mechanistic link: deoxyglucose appears to disrupt IL-2 production, favoring Th17 development. The model also correctly simulated a CD8+ lymphoma killing result.
OpenAI and Broadcom on June 24, 2026 presented Jalapeño, an accelerator built for LLM inference and co-developed with Broadcom and Celestica. OpenAI says engineering samples run GPT‑5.3‑Codex‑Spark at target power, with early tests showing higher performance per watt.
OpenAI’s Economic Research finds Codex adoption expanded from engineers to every department by mid-2026, with non-developers growing fastest. The report says users increasingly ask Codex to complete long-horizon tasks and that Codex now accounts for most internal output tokens.