Transcript · 2026-10-20

Welcome to Overnight AI. I'm Alex. And I'm Jamie. We're your AI-generated hosts, and here's your AI news briefing for October 20th, 2026. Today, faster models, older software, and a different way to organize AI agents. Let's start with that older software. OpenAI published new evaluations for programming languages that have less code and documentation available for models to learn from. That includes COBOL, Fortran, and Visual Basic. Right. One evaluation, MultiLangBench, asks models to build a database engine compatible with SQLite across 20 programming languages. A separate evaluation asks them to migrate COBOL programs while preserving what those programs actually do. So, building something new and preserving something old are being measured separately. What did OpenAI report? On MultiLangBench, it reports an average pass rate of 87.4% for GPT-6 Astra, compared with 26.8% for GPT-5.6 Sol. Those are the company's own results. And on the COBOL migration test, GPT-6.1 Sol reaches 86.6% at low reasoning effort, and 97.9% at higher effort. Again, those are OpenAI's measurements, not an independent production migration. That's the useful boundary here. OpenAI is releasing the MultiLangBench setup and tests so developers can inspect them, but a strong score doesn't establish that a migration preserves a particular system's business rules, dependencies, and edge cases. Okay. So expert review and system-specific testing remain part of the job. Now, Inception is tackling a different constraint: how long applications wait for a model to respond. It released Mercury 3, a diffusion-based language model. Inception reports 1,068 tokens per second on Nvidia GPUs, with a 1 million token context window. It's available through its API and partners including OpenRouter, Baseten, and Vercel. What's different about diffusion in this setting? It generates tokens in parallel rather than relying only on sequential generation. That could matter for interactive tools and agents that repeatedly wait for model responses, but the speed figure is a vendor measurement. Inception also says its benchmark quality meets or exceeds GPT-6 Luna and Claude Haiku 5.5. That doesn't establish the same ranking for every application. Reliable tools, retrieval, and correct answers still determine whether speed helps. Which brings us to Meta's engineering case study. Its team described an agent that knew too much about too many subsystems and started mixing those concepts in production. Their response was to give the central router less to do. Meta describes an orchestrator routing more than 20 workflows to 14 specialist agents. The central layer sequences work and performs reliability checks. Domain decisions go to the specialists. So, this isn't simply a recommendation to add more agents? No. It's a case study about clearer responsibilities and loading the knowledge needed for each task. Meta reuses specialists through explicit workflow templates. It also says future automated changes to agent knowledge would still require human review. The takeaway is narrower context and explicit checks, rather than assuming one all-purpose agent will manage every detail. Let's turn to a more familiar interface: spreadsheets. Google announced conversational BigQuery analysis inside Connected Sheets. Users who have permission to query the connected data can ask Gemini questions about the full dataset, follow up, and request a plan for a new tab with connected formulas, pivot tables, and charts. The full dataset matters here. The analysis isn't limited to the rows currently visible in a spreadsheet. Google says rollout began October 19th and may take up to 15 days. It also says this feature is on by default with no separate admin control for the feature. Existing query permissions still apply, and generated analysis still needs checking before people rely on the conclusions. Finally, Accenture Federal Services and Google Public Sector announced Gemini Launchpads for federal agencies. Embedded engineers would help build agents connected to government systems and datasets. They're also developing prebuilt agents for functions including human resources, finance, and supply chains. It's an implementation program rather than independently measured evidence of better agency outcomes. The announcement emphasizes integration, training, and performance tracking. Those are practical requirements for moving from a demonstration into an operational service. Actual agency-specific outcomes remain to be established. You'll find the source links and written briefing in the show notes. Thanks for listening to Overnight AI. We'll be back with the next briefing.