Volume 343, Issue 3 · 7 articles
OpenAI makes its Decisions API, built on GPT-6 Luna, generally available for fast, low-cost bounded decisions.
OpenAI has made the Decisions API generally available.
It is built on GPT-6 Luna and handles questions with bounded answers, such as classification and routing, up to 10x faster than the Responses API.
Since the beta, OpenAI has added GPT-Live integration, prompt caching, US and EU inference residency, native JSON inputs and up to 400 questions per request.
The post also describes how internal teams use the API for transaction categorization and media coverage classification.
Cloudflare uses OpenAI Daybreak with live traffic and security signals to find vulnerabilities and propose fixes.
OpenAI profiles how Cloudflare combines OpenAI models accessed through Daybreak with its network and threat data in Vulnerability Discovery and Remediation, an early-access service in Cloudflare Managed Defense.
The models examine code and assess potential vulnerabilities, while Cloudflare adds live traffic data and validation to prioritize findings and propose patches.
Cloudflare engineers also use Codex, and OpenAI says one product was rebuilt in three days for about $1,100 in tokens.
OpenAI reports stronger model abilities for building and migrating code in languages like COBOL, Fortran and Visual Basic.
OpenAI says its latest model advances improve building and migrating software in low-resource languages such as COBOL, Fortran and Visual Basic.
These languages have little training data but support critical systems, and OpenAI notes COBOL underpins most ATM transactions.
The work responds to developer and enterprise feedback that coding models struggle with these languages.
OpenAI built two evaluations, including MultiLangBench, which tests building a SQLite-compatible database engine in 20 languages, plus a COBOL migration evaluation.
OpenAI explains how it scaled Azure Cosmos DB beneath Habitat, the storage platform behind its products.
OpenAI describes the second part of its work on Habitat, the storage platform that serves more than 1 billion users across 40 regions and over 500 petabytes of data.
This installment focuses on scaling the underlying Azure Cosmos DB layer.
OpenAI says close work with the Azure team led to a multi-tenancy approach built on co-managed containers.
The post walks through the tradeoffs the team made to become reliable first, then fast, then operable at massive scale.
OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind agents to analyze data from millions of brain cells.
OpenAI describes work with PerturbAI and Scripps Research in which thousands of AI agents powered by GPT-Rosalind analyzed experimental data from millions of mouse brain cells in 32 minutes.
The agents linked the effects of gene disruptions to findings in scientific literature.
OpenAI says this uncovered unexpected biological connections relevant to diseases such as Alzheimer's and Parkinson's.
The company presents the work as a way to help researchers explore more of what large experiments reveal.
OpenAI publishes a blueprint with proposals for governments and industry to put advanced AI in cyber defenders' hands.
OpenAI has published a blueprint for empowering cyber defenders, building on its earlier cybersecurity action plan and frontier safety blueprint.
The company argues that defenders have a limited window to use advanced AI before offensive capabilities spread.
The proposals cover strengthening safety and security for frontier models and agents, including reliable shutdown controls, and modernizing cybersecurity with better tools and training.
OpenAI says its own misalignment and security incidents show the need for stronger industry safeguards.
OpenAI's head of forward deployed engineering explains how the FDE team works with customers on their hardest problems.
Colin Jarvis, head of Forward Deployed Engineering at OpenAI, explains what the FDE team does and how it works.
He says he founded the team in early 2025 to close the deployment gap between what models appear capable of and what people can reliably use.
FDEs work alongside customers on their hardest problems, make repeatable solutions with Product, and bring model shortcomings back to Research.
Jarvis says the goal is for each engagement to raise the floor so future customers need less help.