https://modelsatwork.news/article/anthropic-claude-nist-ecdsa-curve-seeds-lrand48-timestamps Models at Worken 2026-10-16T22:20:00.000Z Anthropic says Claude traced NIST curve seeds to a 1990s random generator Security, Anthropic, Claude, Research https://modelsatwork.news/article/aws-sales-agents-nine-stage-evaluation-framework-agentcore-evals-wrong-answers Models at Worken 2026-10-16T20:58:00.000Z AWS sales agents: a nine-stage eval loop after $410.08 became $410.08 million Agent evaluation, Evals, Amazon Bedrock AgentCore, Multi-agent, Vendor-published https://modelsatwork.news/article/vercel-ai-gateway-production-index-october-2026-open-weight-71-percent-tokens Models at Worken 2026-10-16T20:05:00.000Z Vercel AI Gateway: open-weight models take 71.5% of tokens, 23% of spend Model selection, Open-weight models, Cost, Token cost, Vendor-published https://modelsatwork.news/article/hugging-face-community-post-self-improving-agents-what-must-survive-update-tests Models at Worken 2026-10-16T18:58:00.000Z Hugging Face post: test what an agent keeps when it self-improves Agent evaluation, Agent safety, Evals, Human in the loop, Community https://modelsatwork.news/article/konnect-228-tool-mcp-server-lets-claude-design-kicad-10-pcbs-agpl Models at Worken 2026-10-16T17:56:00.000Z Konnect: 228-tool MCP server lets Claude edit KiCad 10 circuit boards Show HN, Open source, Tool use, Context management, Developer tools https://modelsatwork.news/article/apple-mlx-four-mac-studios-run-816b-muse-spark-open-36-tokens-per-second-self-hosting Models at Worken 2026-10-16T17:00:00.000Z Apple: four Mac Studios run Muse Spark Open at 36 tokens a second Self-hosting, Open-weight models, Inference, Apple, Vendor-published https://modelsatwork.news/article/reflection-beam-501b-23b-active-apache-open-weight-coding-model-sagemaker Models at Worken 2026-10-16T15:58:00.000Z Reflection Beam: 501B open-weight coding model, Apache 2.0, on AWS Model release, Open-weight models, Coding agents, Self-hosting, Amazon SageMaker AI https://modelsatwork.news/article/mistral-fortran-77-to-cpp-40000-lines-parity-harness-human-gated-agents-vendor-published Models at Worken 2026-10-16T13:58:00.000Z Mistral migrated 40,000 lines of Fortran 77 with agents and a human gate Mistral, Coding agents, Human in the loop, Vendor-published, Case study https://modelsatwork.news/article/aws-sagemaker-jumpstart-deploy-test-delete-endpoint-one-prompt-9-seconds-vendor-published Models at Worken 2026-10-15T11:20:00.000Z AWS shows a deploy-test-delete script for SageMaker JumpStart models Amazon SageMaker AI, Vendor-published, Inference, Model selection, Evals https://modelsatwork.news/article/plimsoll-open-source-sandbox-agent-code-isolation-tier-floor-credential-broker Models at Worken 2026-10-16T12:10:00.000Z Plimsoll: open-source sandbox that states its isolation tier Agent security, Open source, Tooling, Agents, Access control https://modelsatwork.news/article/mistral-timestral-time-series-forecasting-209m-parameters-cpu-modified-mit Models at Worken 2026-10-16T09:58:00.000Z Mistral Timestral: 209M-parameter forecasting model runs on CPU Model release, Mistral, Open-weight models, Forecasting, Self-hosting https://modelsatwork.news/article/anthropic-950-claude-agents-21-hours-210m-tokens-enzyme-discovery Models at Worken 2026-10-16T00:55:00.000Z Anthropic ran 950 Claude agents for 21 hours to find a new enzyme Agents, Multi-agent, Anthropic, Vendor-published, Human in the loop https://modelsatwork.news/article/bank-of-america-4000-people-patching-vulnerabilities-found-by-ai Models at Worken 2026-10-15T22:57:00.000Z Bank of America puts 4,000 people on patching AI-found flaws Security, Banking, Project Glasswing, Vulnerability disclosure, Cost https://modelsatwork.news/article/accenture-caps-central-claude-funding-29-a-month-per-person Models at Worken 2026-10-15T21:58:00.000Z Accenture caps central Claude funding at $29 a month per person Cost governance, Token cost, Claude, Anthropic, Change management https://modelsatwork.news/article/aws-trust-in-depth-five-layers-agent-autonomy-agentcore-policy Models at Worken 2026-10-15T20:58:00.000Z AWS Trust in Depth: five layers for letting agents act alone Agent safety, Governance, AWS, Amazon Bedrock AgentCore, Vendor-published https://modelsatwork.news/article/meta-muse-spark-open-816b-42b-active-open-weight-coding-model-sagemaker-jumpstart Models at Worken 2026-10-15T18:58:00.000Z Meta Muse Spark Open: 816B open-weight coding model, 42B active Model release, Open-weight models, Meta, Coding agents, Self-hosting https://modelsatwork.news/article/anthropic-metr-employee-like-access-three-tier-independent-evaluators Models at Worken 2026-10-15T18:05:00.000Z Anthropic gives METR employee-like access, sets three evaluator tiers Governance, Anthropic, Agent safety, Safeguards https://modelsatwork.news/article/nvidia-nemotron-3-5-super-vl-120b-12b-active-1m-context-open-weights-bolt-24gb Models at Worken 2026-10-15T17:20:00.000Z NVIDIA Nemotron 3.5 Super VL: open 120B agent model, 12B active Model release, Open-weight models, NVIDIA, Agents, Self-hosting https://modelsatwork.news/article/theophyseval-gpt-6-astra-codex-49-of-50-physics-problems-under-3-dollars-each Models at Worken 2026-10-15T15:58:00.000Z TheoPhysEval: GPT-6-Astra in Codex solves 49 of 50 physics tasks Research, Benchmarks, Evals, GPT-6, Cost https://modelsatwork.news/article/ai2-gpu-scheduler-budgets-98-percent-owed-hours-debug-wait-30-seconds Models at Worken 2026-10-15T15:00:00.000Z Ai2's GPU budget scheduler: 98% of owed hours, 74% fewer repairs GPU capacity, GPU, Cost governance, Change management, Case studies https://modelsatwork.news/article/apertus-1-5-open-weights-8b-70b-262k-context-apache-2-full-training-data Models at Worken 2026-10-15T14:10:00.000Z Apertus 1.5: open 8B and 70B models, 262k context, Apache 2.0 Model release, Open-weight models, Open source, Evals, Governance https://modelsatwork.news/article/jevcrawl-open-source-agent-crawler-0-025-per-task-vs-0-194-llm-link-picking Models at Worken 2026-10-15T13:05:00.000Z Jevcrawl crawler for agents: $0.025 a task, 40% correct Open source, Agents, Agent cost, Show HN, Benchmarks https://modelsatwork.news/article/aws-sagemaker-jumpstart-model-test-script-one-prompt-qwen3-5-27b-9-seconds-delete-endpoint Models at Worken 2026-10-15T12:10:00.000Z AWS shows a script to test SageMaker JumpStart models AWS, Amazon SageMaker AI, Evals, Latency, Vendor-published https://modelsatwork.news/article/hugging-face-tokenizers-bert-50x-faster-ascii-skip-embedding-rag-pipelines Models at Worken 2026-10-15T10:05:00.000Z Hugging Face makes BERT tokenizing up to 50x faster Inference, RAG, Cost, Open source, Vector search https://modelsatwork.news/article/aws-agentcore-powerpoint-generator-sample-0-04-dollars-per-deck-vendor-published Models at Worken 2026-10-15T01:57:00.000Z AWS AgentCore deck generator sample: ~$0.04 and 70–90 seconds Agents, Amazon Bedrock AgentCore, Vendor-published, Cost, RAG https://modelsatwork.news/article/graphicle-open-source-rust-code-graph-memory-for-coding-agents-kubernetes-3-min-vs-26-min Models at Worken 2026-10-15T01:30:00.000Z Graphicle: open-source code graph and memory for coding agents, runs locally Coding agents, Open source, Agents, Cost, Community https://modelsatwork.news/briefing/2026-10-16 Models at Worken 2026-10-16T10:30:00.000Z The briefing, 2026-10-16: 404 Media reports Accenture funds only $29 a month per person for Claude Enterprise https://modelsatwork.news/briefing/2026-10-15 Models at Worken 2026-10-15T10:30:00.000Z The briefing, 2026-10-15: Microsoft and Hugging Face say the best agent passed all 20 runs on only 47.5% of 507 workflows