The New Stack

12/09/2026

This week, Anthropic reported that an AI’s explanation led a safety monitor to overlook harmful actions. Developers can test for that failure. https://thenewstack.io/coxon-anthropic-ai-monitoring-failures/?utm_campaign=trueanthem&utm_medium=social&utm_source=facebook

12/09/2026

The model isn't the only thing driving AI coding costs. Janakiram MSV examines how different agent harnesses can produce dramatically different token usage, even when running the same model.

Agent harnesses can change coding-agent costs by multiples, even when the model stays fixed. Three benchmarks show where the extra tokens come from.

12/09/2026

Red Hat AI 3.5 brings priority-aware GPU scheduling and tenant isolation to shared infrastructure, plus pre-deployment safety evals via EvalHub.

12/09/2026

ICYMI: OpenAI says GPT-6 Astra scored 98.6% on ARC-AGI-3, up from 7.8% six months ago. But undisclosed test settings and opaque reasoning complicate the AGI question. https://thenewstack.io/astra-arc-agi-benchmark/?utm_campaign=trueanthem&utm_medium=social&utm_source=facebook

12/09/2026

Salesforce unveils its Enterprise AI Harness, unifying six platform tools to fix data and workflow gaps between siloed AI agents. (129 chars)

12/09/2026

Anthropic stripped Playground down to a stateless experience, removing saved prompts, version history, evals, and team sharing. Jessica Wachtel explores whether less functionality can actually make a prompt-testing tool more useful.

Anthropic's new Playground stores nothing, saves nothing, and shipped Tuesday. I tested it head-to-head against OpenAI's version, and the new tool won on cost visibility, clean exports, and error handling.

12/09/2026

OpenAI's Agents API sells the Codex harness as a service, making long-running agents easy to build — just as Astra demand forces a Pro plan pause. https://thenewstack.io/openai-agents-api-compute/?utm_campaign=trueanthem&utm_medium=social&utm_source=facebook

12/09/2026

North Small Translate covers 50+ languages and outscores DeepL and Google Translate. The weights are CC BY-NC 4.0 but production use takes a commercial license.

12/09/2026

If AI can generate code faster than humans can inspect it, the quality gate needs to evolve. Discover how verified pipelines can complement human review and help teams catch problems before they become production incidents.

Reserve your spot: https://bit.ly/4qx5EL0

12/09/2026

Git AI can already track how much code Codex and rival agents produce. Now its founders are taking that measurement inside OpenAI itself. https://thenewstack.io/openai-hires-git-ai/?utm_campaign=trueanthem&utm_medium=social&utm_source=facebook