AI News Flash · Daily Brief

OpenAI brings GPT-5.5 agentic tools to life sciences via updated GPT-Rosalind.

By and · Editorial team, AI News Flash

Platforms

OpenAI brings GPT-5.5 agentic tools to life sciences via updated GPT-Rosalind.

OpenAI has shipped a significant update to GPT-Rosalind, its vertically specialized life sciences model, incorporating GPT-5.5's agentic coding and tool-use capabilities alongside improved performance on drug-discovery and genomics tasks. The release adds new plugins targeting evidence retrieval and bioinformatics workflows, making the system more useful across the full research pipeline. OpenAI is also expanding a trusted-access research preview to eligible organizations globally, signaling a deliberate push to build specialized frontier models alongside its general-purpose GPT line and deepen its foothold in high-value scientific domains.

Why it matters: Life sciences organizations worldwide can now access frontier agentic AI tuned for drug discovery and genomics research.

Gemini 3.5 Pro misses Pichai's June deadline as enterprise preview stays limited.

Gemini 3.5 Pro remains unreleased as of June 9, more than three weeks after Sundar Pichai told the Google I/O audience to "give us until next month," a line that drew audible groans from the crowd. The model is currently confined to internal testing and a limited Vertex enterprise preview, with no public launch date confirmed. Google designed Gemini 3.5 Pro to support a 2-million-token context window and a new Deep Think reasoning mode intended to close the hard-reasoning gap left open by the already-shipped Gemini 3.5 Flash. The continued delay puts pressure on Google as competitors advance their own frontier releases.

Why it matters: Enterprise customers evaluating Vertex AI and developers planning around Google's roadmap face continued uncertainty about when Gemini 3.5 Pro will be available.

Meta says Muse Spark API is in partner testing and on track for June launch.

After the Wall Street Journal reported multiple delays and the absence of a firm launch date, Meta went on the record to defend its timeline: a spokesperson confirmed that the Muse Spark API is currently in testing with early partners and remains on track to ship later in June. The API has encountered repeated slips since the model made its public debut on April 8, when Meta Superintelligence Labs launched the consumer-facing Muse Spark product but deliberately held back developer access. The confirmation suggests Meta is trying to manage narrative risk as interest in the API grows among developers who have been waiting since the initial launch.

Why it matters: Developers building on Meta's AI platform can expect API access to Muse Spark within weeks, depending on whether the June timeline holds.

Capabilities

Harness-1 Hits 0.730 Recall on 8 Benchmarks by Offloading Bookkeeping from RL Policy

A team from UIUC, UC Berkeley, and Chroma has published Harness-1, a 20-billion open-weights retrieval subagent designed around a clean separation of concerns: a stateful harness manages working memory and evidence tracking, while an RL-trained policy focuses solely on deciding what to search and when to stop. Across eight retrieval benchmarks covering web search, finance, patents, and multi-hop question answering, Harness-1 achieves 0.730 average curated recall, surpassing the next strongest open search subagent by 11.4 points and outperforming GPT-5.4, Sonnet-4.6, and Kimi-K2.5 under identical evaluation protocols. Only Opus-4.6 scored higher. The most striking result is its generalization: on four held-out benchmarks not seen during RL training, performance gains were 2.2 times larger than on in-distribution tasks.

Why it matters: Open-weights developers and enterprises building retrieval pipelines now have a competitive 20B alternative to closed frontier models for complex search tasks.

Technology & Research

Z.ai's GLM-5.1 hits 400 tokens per second with no Nvidia hardware involved.

Z.ai has released a GLM-5.1-highspeed API variant that delivers 400 tokens per second, a throughput the company claims leads major LLM providers. The underlying GLM-5.1 is a 754-billion-parameter mixture-of-experts model with 40 billion active parameters per token, and it uses DeepSeek Sparse Attention to reduce compute costs on long-context inputs. Notably, the entire model was trained on Huawei Ascend 910B chips with no Nvidia hardware involved, making it a concrete data point that U.S. semiconductor export controls have not blocked Chinese labs from achieving frontier-class inference performance. The release is likely to intensify scrutiny of non-Nvidia AI infrastructure as a viable path to competitive deployment.

Why it matters: GLM-5.1 shows that China-based labs can reach top-tier inference speeds using domestic chips, complicating the assumed impact of U.S. export controls.

Regulation & Policy

Trump signs order creating voluntary 30-day government review for frontier AI models.

President Trump signed an executive order on June 2 directing the NSA and CISA to develop a classified benchmarking process to identify covered frontier models with advanced cyber capabilities. Under the order, AI developers may voluntarily provide the government with up to 30 days of pre-release access before broader distribution, but no mandatory licensing or pre-clearance requirement is imposed, resolving weeks of internal administration debate over regulatory scope. The order also establishes an AI cybersecurity clearinghouse to support vulnerability scanning and remediation across critical infrastructure. The framework stops well short of the mandatory review mechanisms that some officials had pushed for, representing a middle path between intervention and a hands-off approach.

Why it matters: Frontier AI developers face a new but voluntary government review window, setting a precedent that could shape future mandatory oversight proposals.

EU AI Act Omnibus clears final formal approval vote, shifting high-risk deadlines to 2027

The EU AI Act Omnibus, a provisional political agreement reached on May 7, is advancing toward final formal approval, with publication in the Official Journal expected in July and ahead of the August 2026 AI Act milestone. The package's most consequential provision defers Annex III high-risk system obligations, covering AI used in recruitment, credit scoring, and law enforcement, from August 2026 to December 2027, giving affected organizations more than a year of additional preparation time. The Omnibus also introduces two new Article 5 prohibitions targeting AI-generated non-consensual intimate imagery and child sexual abuse material, both effective December 2, 2026. It is the first set of amendments to the AI Act since its original adoption in June 2024.

Why it matters: Companies deploying AI in recruitment, credit, or law enforcement gain over a year of additional compliance runway, while new content prohibitions take effect by year-end.

AI Stocks

(ORCL) Oracle reports Q4 FY2026 earnings tonight with AI cloud in focus

Oracle will release its fiscal fourth quarter 2026 earnings after today's market close, with Wall Street consensus at $1.96 earnings per share on $19.10 billion in revenue. The central focus for investors is the trajectory of Oracle Cloud Infrastructure, which posted 243% year-over-year AI infrastructure revenue growth in Q3 and reached $553 billion in remaining performance obligations. Analysts are also watching whether FY2027 capital expenditure guidance approaches the $100 billion level that Scotiabank projects. Ahead of the print, multiple analysts raised their price targets, with Cantor Fitzgerald lifting its target to $284 from $229 and Scotiabank raising to $290 from $215, reflecting broad conviction in Oracle's AI infrastructure buildout.

Why it matters: Oracle's guidance on FY2027 capex and OCI growth will signal how aggressively hyperscalers outside the traditional big three are scaling AI infrastructure investment.