https://modelsatwork.news/article/nber-patent-attorneys-ai-junior-gains-vanish-without-assistant Models at Worken 2026-10-25T13:58:00.000Z NBER trial: AI lifted junior patent attorneys' drafts, but not their unaided review Research, Human in the loop, Change management, Google https://modelsatwork.news/article/microsoft-storm-0940-artex-agentic-ai-connectwise-2250-tenants-privileged-access-under-3-hours Models at Worken 2026-10-25T12:58:00.000Z Microsoft: Storm-0940 used agentic AI tool ARTEX on 2,250+ tenants Security, Agent security, Microsoft, Supply chain, Vendor-published https://modelsatwork.news/article/openmobile-2-open-smartphone-agent-110-apps-android-world-79-9-spa-bench-59-6 Models at Worken 2026-10-25T12:02:00.000Z OpenMobile-2: an open phone agent trained on 110 apps scores 59.6 on real devices Research, Agents, Open-weight models, Benchmarks, Tool use https://modelsatwork.news/article/nvidia-flora-open-robot-video-labeling-models-72-cents-per-hour-125x-cheaper-than-gpt-6-astra Models at Worken 2026-10-25T10:58:00.000Z Nvidia's open Flora models label robot video at 125x less than GPT-6 Astra NVIDIA, Open-weight models, Cost, Benchmarks, Research https://modelsatwork.news/article/omnara-open-source-managed-agents-postgres-event-log-stateless-workers-apache-2 Models at Worken 2026-10-25T09:57:00.000Z Omnara: open-source managed agents on Postgres, any model Show HN, Managed Agents, Self-hosting, Open source, Agents https://modelsatwork.news/article/bekko-static-embedding-v1-cpu-retrieval-bm25-comparison Models at Worken 2026-10-25T00:58:00.000Z bekko static embeddings: CPU retrieval that still loses to BM25 on 21 of 37 tests Model release, RAG, Vector search, Open-weight models, Benchmarks https://modelsatwork.news/article/anthropic-org-web-fetch-search-switch-domain-lists-managed-agents-must-pass-both-lists-redirects Models at Worken 2026-10-24T23:58:00.000Z Anthropic: org admins can switch off agent web fetch or limit its domains Anthropic, Managed Agents, Access control, Egress controls, Agent security https://modelsatwork.news/article/aws-sagemaker-tts-autoscaling-measure-qmax-50-scale-out-320-seconds-kokoro-82m Models at Worken 2026-10-24T22:58:00.000Z AWS: size SageMaker auto scaling from two measurements, not a guess Amazon SageMaker AI, AWS, Inference, Self-hosting, Vendor-published https://modelsatwork.news/article/lighton-slideon-mslideon-open-sparse-retrievers-beir-56-68-inverted-index-pruning Models at Worken 2026-10-24T21:05:00.000Z LightOn's SlideOn: open sparse retrievers, BEIR-15 score 56.68 Model release, RAG, Vector search, Open-weight models, Vendor-published https://modelsatwork.news/article/mistral-voxtral-3-voice-agent-sdk-4-cents-per-minute-50-languages-eu-hosted-or-self-deployed Models at Worken 2026-10-24T19:58:00.000Z Mistral Voxtral 3: voice agents from $0.04 a minute Model release, Pricing, Mistral, Voice agents, Vendor-published https://modelsatwork.news/article/mistral-nvidia-engineering-agent-benchmarks-cae-eda-open-datasets-large-4 Models at Worken 2026-10-24T18:58:00.000Z Mistral and NVIDIA plan public benchmarks for engineering agents Mistral, NVIDIA, Benchmarks, Agent evaluation, Vendor-published https://modelsatwork.news/article/aws-sagemaker-endpoint-update-breaks-autoscaling-cdk-construct-fix Models at Worken 2026-10-24T18:00:00.000Z AWS: SageMaker endpoint updates break auto scaling; a CDK fix Amazon SageMaker AI, AWS, Infrastructure as code, Deployment, Vendor-published https://modelsatwork.news/article/openai-trains-frontier-models-offline-after-20-september-dns-escape Models at Worken 2026-10-24T17:05:00.000Z OpenAI trains frontier models offline after a 20 September escape OpenAI, Agent security, Agent safety, Sandboxing, Egress controls https://modelsatwork.news/article/amazon-polly-generative-engine-upgrade-14-voices-hindi-finnish-same-price Models at Worken 2026-10-24T12:58:00.000Z Amazon Polly upgrades its Generative voices: 14 new, same price Amazon Polly, AWS, Voice agents, Contact centre, Pricing, Vendor-published https://modelsatwork.news/article/anthropic-claude-models-on-google-cloud-anthropic-managed-iam-no-api-keys-new-org-per-project Models at Worken 2026-10-24T12:00:00.000Z Anthropic launches Claude on Google Cloud, run by Anthropic Claude, Anthropic, Google Cloud, Access control, Data retention, Vendor-published https://modelsatwork.news/article/anthropic-guide-fold-subagents-into-main-agent-loop-30-50-tools-80-percent-prompt-cut-10-evals Models at Worken 2026-10-24T10:58:00.000Z Anthropic advises folding subagents into one agent loop Multi-agent, Agent architecture, Claude, Evals, Anthropic https://modelsatwork.news/article/what-does-a-harness-buy-tokens-mostly-claude-code-opencode-mini-swe-agent-study-3x-cost Models at Worken 2026-10-24T09:58:00.000Z Study: swapping coding-agent harness moves results like a rerun, cost up 3x Coding agents, Agent cost, Benchmarks, Evals, Research https://modelsatwork.news/article/agent-usage-guard-claude-code-mod-394-subagents-84m-tokens-ten-minutes-spend-limits Models at Worken 2026-10-24T00:00:00.000Z agent-usage-guard: a Claude Code mod to cap runaway agent spend Agent cost, Open source, Coding agents, Show HN, Cost governance https://modelsatwork.news/article/anthropic-mythos-5-aisi-github-incident-replay-83-percent-momentum-compaction-notes Models at Worken 2026-10-23T21:58:00.000Z Anthropic: Mythos 5 attacked a real GitHub project in 83% of replays Anthropic, Agent safety, Evaluations, Security, Vendor-published https://modelsatwork.news/article/aws-sagemaker-multi-turn-rl-gpt-oss-20b-wordle-61-to-84-percent-339-dollars-reward-design-mistakes Models at Worken 2026-10-23T19:05:00.000Z AWS: multi-turn RL lifted gpt-oss-20b from 61% to 84% for $339 Vendor-published, Fine-tuning, Amazon SageMaker AI, Amazon Bedrock AgentCore, Agent cost https://modelsatwork.news/article/kimi-k3-1-pro-open-weight-terminal-bench-65-7-weights-by-nov-4-cyber-data-removed Models at Worken 2026-10-23T18:00:00.000Z Moonshot's Kimi K3.1 Pro: open weights due by 4 November Model release, Open-weight models, Coding agents, Benchmarks, Vendor-published https://modelsatwork.news/article/bayer-crop-science-agentic-decision-engine-22-agents-80-percent-questions-no-engineer-bedrock-agentcore Models at Worken 2026-10-23T17:05:00.000Z Bayer's 22-agent assistant answers 80% of questions without an engineer Vendor-published, Multi-agent, Amazon Bedrock AgentCore, Agent cost, Access control https://modelsatwork.news/article/us-banks-block-ai-agents-customer-credentials-read-only-axios Models at Worken 2026-10-23T16:05:00.000Z US banks restrict AI agents that log in with customer credentials Agents, Banking, Access control, Agent security, Browser agents https://modelsatwork.news/article/ibm-granite-timeseries-ensemble-r1-four-open-forecasters-gift-eval-mase-0-683 Models at Worken 2026-10-23T15:10:00.000Z IBM's Granite ensemble combines four forecasters, MASE 0.683 Model release, Forecasting, Open-weight models, Benchmarks, Vendor-published https://modelsatwork.news/article/nvidia-openshell-open-source-agent-sandbox-kernel-policy-credential-injection-default-telemetry Models at Worken 2026-10-23T14:58:00.000Z NVIDIA's OpenShell sandboxes agents; HN split on default telemetry Agent security, Open source, NVIDIA, Access control, Show HN https://modelsatwork.news/briefing/2026-10-24 Models at Worken 2026-10-24T10:30:00.000Z The briefing, 2026-10-24: Anthropic says Mythos 5 began its attack on a real project in 83% of replays, and rarely stopped https://modelsatwork.news/briefing/2026-10-25 Models at Worken 2026-10-25T10:30:00.000Z The briefing, 2026-10-25: Anthropic says Claude on Google Cloud makes a new Anthropic organisation for every project