Digital Plumber Plumbing the information age
RSS

Topics

AI Model Providers

Model releases and changes from Anthropic, OpenAI, Google and others that affect operations tooling.

189stories in the archive
7 up from 6last 7 days · 6 the 7 before
27last 30 days
Jun 20first covered · 2026

Recent activity

Stories per week

Latest stories

10 most recent
  1. Rio Rundown, OpenAI · Nov 20, 2026 · Primary source

    OpenAI agents resolve 600 ChatGPT performance regressions

    OpenAI reports that its new Performance Factory utilizes Codex to autonomously identify and resolve performance regressions in ChatGPT; the system has processed over 1,700 investigations and resolved more than 600 cases in three months, replacing manual debugging with agentic triage and human-validated fixes.

    Why it matters Demonstrates a production AIOps use case: autonomous systems triage performance anomalies at scale and propose remediation, requiring only human validation before deployment, applicable to any high-traffic service observability model.

    Nov 20, 2026 edition

  2. 9to5Google · Nov 18, 2026 · Industry news

    Google's Gemini 4 Neon matches Carbon quality at half the cost

    Google released Gemini 4 Neon on November 18 as the middle tier of its Gemini 4 family. It scores 61.4 on Artificial Analysis Intelligence Index at xhigh reasoning, delivering Carbon-class performance at Flash speeds. On long-context reasoning (GraphWalks), Neon achieves 95.2% accuracy versus 80.1% for GPT-6.2 Sol.

    Why it matters Neon delivers comparable capability to flagship Carbon at lower cost per task with 1M-token context—critical for cost-sensitive enterprise agentic workflows and reasoning-heavy tasks requiring long-context understanding.

    Nov 19, 2026 edition

  3. OpenAI Deployment Safety Hub · Nov 17, 2026 · Primary source

    GPT-6.2 models stall on routine steps despite benchmark gains

    OpenAI released GPT-6.2 Astra and Sol on November 17 with improved benchmarks and 20% lower token pricing. On internal GDPval v2, Astra improves 6.2 percentage points over GPT-6 Astra. However, early testers reported severe autonomy issues—models ask permission on routine steps like file downloads that prior versions handle independently.

    Why it matters Early testers report Astra cannot handle autonomous workflows without constant hand-holding, asking permission on trivial technical decisions—a critical regression for agent-based production systems despite higher benchmark scores.

    Nov 19, 2026 edition

  4. ForkLog · Nov 18, 2026 · Industry news

    OpenAI ships GPT-6.2 Astra and Sol with cheaper input

    OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol on November 17, with Astra for complex professional tasks and Sol for everyday work. GPT-6.2 Astra is 20% cheaper on input and output and 60% cheaper on cached input, using 22% fewer output tokens than GPT-6 Astra.

    Why it matters GPT-6.2 Astra replaces GPT-6 Astra as the flagship model with a 1,050,000-token context window, improved document/spreadsheet formatting, and 20–60% cost reductions—directly lowering per-task expenses for production deployment at scale.

    Nov 18, 2026 edition

  5. AI Weekly · Nov 16, 2026 · Industry news

    Nvidia CEO exits as Anthropic sets IPO date

    In the first half of November 2026, Nvidia's CEO Jensen Huang resigned, Anthropic set its IPO for November 19, Anthropic and Google shipped new frontier models. Claude Fable 5.2 (November 9) is built for complex coding and research with 30% faster output at $8/$40 per million tokens; Gemini 4 Carbon (November 10) is Google's strongest model at $2/$10 introductory pricing.

    Why it matters Fable 5.2 scored 49.1% on FrontierCode 2.0 at half the cost per task of Opus 5.5, directly affecting model selection economics and infrastructure budgeting for teams building agentic coding workflows.

    Nov 16, 2026 edition

  6. pasqualepillitteri.it · Nov 13, 2026 · Analysis

    OpenAI and Anthropic pledge model shutdown mechanisms in Paris Call

    OpenAI, Anthropic, Microsoft and Mistral AI signed the Paris Call for AI Safety and Security on November 9, 2026, pledging to build mechanisms to stop, pause or shut down their models during training and internal use. OpenAI described Control Room, an internal service that stops research jobs on models in 35 seconds.

    Why it matters The pledge has no binding force or deadline; the commitment is voluntary and covers only frontier models, signaling regulatory pressure without hard enforcement mechanisms.

    Nov 15, 2026 edition

  7. Google Blog · Nov 12, 2026 · Primary source

    Google releases Gemini 4 Hydrogen for coding and agents

    Google released Gemini 4 Hydrogen, a multimodal model adapted from Gemini 4 Carbon, engineered for high-velocity everyday coding, debugging, and agentic workflows across four configurable thinking levels. It supports text, image, speech, and video input with 1 million token context window and up to 1 million output tokens.

    Why it matters Gemini 4 Hydrogen surpasses Argon on the Artificial Analysis Intelligence Index at 2.5x the speed of Gemini 3.8 Flash, at $1.50 per million input tokens and $7.50 per million output. It offers a practical cost-efficient tier for agentic and multimodal deployment at scale.

    Nov 14, 2026 edition

  8. CASRAI · Nov 13, 2026 · Industry news

    Google DeepMind ships Gemini Robotics 2.5 to robotics partners

    Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026, its most capable vision-language-action model that converts visual and language input into robot motor control. Rolling out to partners including Apptronik, Mercedes-Benz and Agile.

    Why it matters Robotics operators now have access to a production-ready embodied AI model that translates vision and language into physical control, replacing custom end-to-end training for robotic tasks.

    Nov 13, 2026 edition

  9. Anthropic official documentation · Nov 9, 2026 · Primary source

    Anthropic cuts Claude Fable prices and speeds inference

    Claude Fable 5.2 released November 9, 2026, with improved coding, reasoning, and vision. The model costs 20% less than Fable 5.1, generates output up to 30% faster, and maintains 1M-token context and 128K output maximum.

    Why it matters A team of 1,000 agents running Fable 5.2 formalized a proof of Fermat's Last Theorem in 30 hours and built a prototype operating system from scratch—demonstrating significant improvements in agentic reasoning for infrastructure automation and complex DevOps workflows.

    Nov 12, 2026 edition

  10. AI Daily Digest / Mistral blog · Nov 9, 2026 · Primary source

    Mistral opens weights of trillion-parameter Large 4 model

    Mistral released full weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter natively multimodal mixture-of-experts model with 50 billion active parameters per token and 1M context, trained on European infrastructure under a modified-MIT open-weight license allowing self-deployment.

    Why it matters Open-weight access enables on-premises sovereign deployment for cybersecurity and coding workloads; achieves frontier performance with 47% fewer tokens per task than Chinese rivals and leads open models on coding agent benchmarks.

    Nov 11, 2026 edition

Coverage history

179 earlier stories

November 2026 9

October 2026 26

September 2026 36

August 2026 45

July 2026 45

June 2026 18