Faster models, legacy-code tests and narrower agents move AI toward practical work

OpenAI tests coding beyond the most common languages

OpenAI published two evaluations aimed at software written in languages with less readily available code and documentation, including COBOL, Fortran and Visual Basic. MultiLangBench asks models to build a SQLite-compatible database engine across 20 languages; a separate test measures whether migrating COBOL programs preserves their observable behavior.

The company reports an average MultiLangBench pass rate of 87.4% for GPT-6 Astra, compared with 26.8% for GPT-5.6 Sol. On its COBOL migration evaluation, GPT-6.1 Sol reaches 86.6% at low reasoning effort and 97.9% at higher effort. These are OpenAI’s own results, not an independent assessment of production migrations.

OpenAI is releasing MultiLangBench’s setup and tests. That makes inspection possible, while leaving the practical question intact: can a proposed migration preserve the business rules, dependencies and edge cases of a particular deployed system? A benchmark score does not remove the need for expert review and system-specific testing.

Inception releases Mercury 3 with a focus on speed

Inception launched Mercury 3, a diffusion-based language model available through its API and partners including OpenRouter, Baseten and Vercel. The company reports 1,068 tokens per second on NVIDIA GPUs and a one-million-token context window, positioning it for interactive tools and agents that repeatedly wait for model responses.

Inception says its benchmark quality meets or exceeds GPT-6 Luna and Claude Haiku 5.5. Those comparisons come from the vendor. They do not establish the same ranking for every workload, serving configuration or latency measurement.

The architectural distinction is useful: diffusion generates tokens in parallel rather than relying only on sequential generation. For developers, the test is whether that speed remains helpful when the application also needs reliable tool use, retrieval and correct answers.

Meta gives its agent router less work

Meta Engineering described replacing an overloaded central agent with a deliberately narrow orchestrator for onboarding data columns across its recommendation infrastructure. Its reported implementation routes more than 20 workflows to 14 specialized agents; the central layer sequences work and performs reliability checks rather than making domain decisions.

The team says its earlier all-purpose agent mixed concepts from different subsystems in production. Its alternative loads specialist knowledge when needed and reuses workers through explicit workflow templates.

This is a case study, not proof that adding more agents always improves results. Its useful lesson is about boundaries: keep the routing and checks clear, then give each worker the context needed for its task. Meta says future automated improvements to agent knowledge would still require human review.

Gemini connects spreadsheet questions to BigQuery

Google announced conversational BigQuery analysis inside Connected Sheets. Users with permission to query the connected data can ask Gemini questions about the full dataset, follow up, and request a plan for a new tab containing connected formulas, pivot tables and charts.

The rollout began October 19 and may take up to 15 days. Google says the feature is on by default, with no separate admin control for this feature. Existing query permissions still apply.

The distinction from ordinary spreadsheet assistance is the data boundary: the query can cover the connected BigQuery dataset, rather than only the rows visible in a sheet. Generated analysis still needs checking before someone relies on its conclusions.

Accenture and Google target federal agent deployment

Accenture Federal Services and Google Public Sector announced Gemini launchpads, pairing embedded engineers with agencies to build agents connected to government systems and datasets. They also described prebuilt agents under development for functions including human resources, finance and supply chains.

The announcement sets out an implementation program, not independently measured results. Its emphasis on integration, workforce training and performance tracking reflects the work between a demonstration and an operational service; agency-specific outcomes remain to be established.