AI Daily · 2026-09-11

OpenAI’s announcements today center on putting frontier models into more organizational workflows: a new Data agent in ChatGPT Work turns approved data sources into answers, dashboards, and follow-up actions; a multi-year GSA agreement removes license fees and halves usage costs for eligible governments while discounting Daybreak Blue and opening Daybreak Red for approved red-team use; ChatGPT for Financial Services embeds paid financial data and GPT‑6 Astra for investment banking and equity research. Elsewhere, Google’s ToolGrad reverses the usual data-generation order by deriving user queries from tool-use chains to lower the cost of long-horizon tool-use data, and NVIDIA highlights Skild AI’s S1 model for learning long tasks from a single video. On the developer infrastructure side, Replit and Databricks hit GA with native Lakebase support, Datasette shipped security releases after frontier-model audits, and trynix.dev runs any Nix package in the browser; LangChain’s OpenWiki use case shows repository documentation agents becoming a prerequisite for coding agents.

North America · First-hand

OpenAI

⭐⭐⭐ [Product Update] Now everyone can put data to work

OpenAI introduced a new Data agent in ChatGPT Work that lets users turn company data into answers, interactive dashboards, and action simply by asking, without writing queries or learning a new analytics tool. The agent connects to approved data sources including Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake, and more, and can bring files and documents from Google Drive and SharePoint into the analysis. It interprets data using an organization's business terms, metric definitions, custom calculations, and data relationships, drawing that context from semantic layers and trusted sources such as Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and BI dashboards. Users can direct and refine the analysis within a single conversation and share the resulting dashboards. The announcement includes comments from platform partners including AWS, ClickHouse, Databricks, Snowflake, MongoDB, Redis, and G2.

⭐⭐⭐ [Other] Expanding AI access and cyber defense for federal, state, local, and tribal governments

OpenAI announced a multi-year agreement with the U.S. General Services Administration that gives eligible federal, state, local and tribal governments $0 license fees (normally $15 per user per month), no minimum commitment, and 50% off usage costs. More than one million government employees already have ChatGPT access under existing agreements, and the expanded partnership extends eligibility across a U.S. public-sector workforce of roughly 23 million people; the announcement names GPT‑6 Astra among the tools available to them. On cyber defense, every verified government entity will be approved for Daybreak Blue at 50% off standard commercial pricing, and can request Daybreak Red for advanced vulnerability research, exploit validation and red teaming at standard commercial pricing. The announcement builds on Daybreak for Frontline Defenders, announced the prior week with a $1 billion global commitment. It also cites earlier cases such as the CDC producing most initial public-health literature reviews in under 30 minutes and Georgia's Department of Revenue cutting tax-form digitization from up to two weeks to 15 minutes.

⭐⭐ [Product Update] Introducing ChatGPT for Financial Services

OpenAI introduced ChatGPT for Financial Services, a tailored ChatGPT Work experience that combines built-in financial data with GPT-6 Astra's reasoning to help teams produce research, financial models, and customized client materials. The product was shaped through design partnerships with Morgan Stanley and Evercore, starting with investment banking and equity research. It includes premium data from providers such as Daloopa, PitchBook, LSEG News, and Crunchbase, indexed and hosted by OpenAI for higher accuracy and granular citations so bankers can trace figures and claims back to their sources. GPT-6 Astra and newer models are available natively out of the box, with centralized access and data-connection management backed by ChatGPT's enterprise security and governance controls. OpenAI said its work with partners will inform post-training, product improvements, and expansion into other financial services categories.

Google

⭐⭐ [Research] ToolGrad: Efficient tool-use dataset generation with textual "gradients"

Google Research presented ToolGrad at ACL 2026, a tool-use dataset generation framework that reverses the usual paradigm by generating ground-truth tool-use chains first and then annotating the corresponding user queries. The authors argue that an explicit tool-use solution carries more unambiguous information than a prompt, so annotation from tool usage to user query needs only a single LLM step, yielding more complex long-horizon tool-use data at lower cost than prior approaches that generate hypothetical instructions and then search for solutions with a DFS agent. ToolGrad adapts the textual-gradient idea from TextGrad, using plain-text feedback from an LLM critic to iteratively build API workflows, and consists of four modules: API Proposer, API Executors, API Selector, and LLM Updater. The authors report that LLMs trained on ToolGrad data outperform those trained on baseline methods and even match state-of-the-art proprietary LLMs on out-of-distribution datasets with unseen tools.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐ [Model Release] Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video

An NVIDIA blog post introduces Skild AI's new S1 robot foundation model, which is designed to learn previously unseen, long-horizon tasks from a single video demonstration. The post notes that manufacturing floors, warehouses and production lines rarely stay fixed—tasks change, layouts shift and new products arrive—and most robots can't keep up without significant reprogramming. According to the post, S1 launched last week and taps NVIDIA's Physical AI, as stated in the headline. The article is brief and does not disclose model size, training data or performance figures.

Product Update

⭐⭐ [Product Update] Replit | Databricks Integration is Now Generally Available with Native Lakebase Support

Replit announced that its integration with Databricks is now generally available, adding native Databricks Lakebase support on top of the foundation from its public preview announced last June. Under the integration, Replit accelerates application creation while Databricks hosts and provides governed access and security for live enterprise data, with Lakebase offering a managed database for storing and updating application data. Native Lakebase support lets teams build full-stack, context-rich apps that combine live Databricks warehouse data with data created and updated through the app. When an app built with Replit is ready to deploy, Replit Agent automatically provisions its Lakebase database, removing manual setup. A new automated preview deploy capability also creates a separate preview environment that keeps test data isolated from live business data in Databricks.

⭐⭐ [Product Update] Datasette 1.0a39 and 0.65.4 security releases

Datasette released two security patch versions, 1.0a39 for the current alpha series and 0.65.4 for the stable 0.65.x family; the fixes should be applied by anyone running a Datasette instance on the public web, particularly if it mixes public and private tables. After issues were reported by Sevban Dönmez, Alex Garcia and Simon Willison ran an extensive audit of Datasette using Claude Fable 5.1, GPT-5.6 and GPT-6 Astra, then spent nearly a week collaborating on and reviewing the fixes; the models helped find some very subtle bugs, and frontier-model security audits will be part of all future development work. The two split the work in a shared private repository: one wrote automated tests highlighting each issue while the other implemented the fix, ensuring two humans reviewed every issue in addition to coding agents running different models.

⭐⭐ [Product Update] Any Nix package, live in your browser

Simon Willison highlights Farid Zakaria's trynix.dev in a link post, which Zakaria calls his magnum opus of Nix work. The site provides a qemu-wasm powered x86_64 Linux virtual machine that runs entirely in the browser via WebAssembly and can boot any Nix package from the past 13 years. The VMs are URL addressable, so visiting https://trynix.dev/?pkg=python3%403.6.2 and clicking Load yields an interactive shell running Python 3.6.2 from 2017. Zakaria has also built trynix-preview on top of it: a GitHub Action that comments a link on a pull request so reviewers can boot that PR's build in the browser, with no servers involved.

Other

⭐⭐ [Other] How Credit Genie keeps codebase docs fresh with OpenWiki

Credit Genie, a mobile-first financial wellness platform, uses LangChain's open-source repo-documentation agent OpenWiki across its AI and ML engineering teams. Previously the team kept project docs in Notion pages, README files, and AGENTS.md files, but in a fast-moving environment that content quickly went stale and became hard to find, leaving both engineers and coding agents without a reliable source of truth. Mattia Ciollaro, Credit Genie's Head of AI and ML, adapted OpenWiki from a side project into a self-serve portal hosted on GitHub Pages that aggregates OpenWiki docs from all onboarded repositories into one searchable web interface. Documentation lives in each repository's openwiki/ folder; OpenWiki runs nightly, checks for commit changes, and opens a pull request with doc updates when it finds meaningful differences, while onboarding a new repo requires just one change in the portal config. The system serves two audiences: team members who search and read docs in the portal, and coding agents that are instructed to check the openwiki/ folder before acting.