News

This Week in AI

A weekly recap of the stories that actually matter, plus the latest AI news as it happens — models, tools, and policy — every item linked to its original source.

Weekly recap

One digest every Monday
Latest issue

This Week in AI — Week of October 5, 2026

Gemini's universal work agent, Claude Haiku 5.5, Cloudflare's Deno acquisition, Mistral Large 4 open weights, and Anthropic's agent safety disclosure.

Read the digest →

This Week in AI — Week of September 28, 2026

Slovenia's .si domain surge, Gemini 4 Argon's million token ceiling, GPT-6.1 Sol, Ideogram 4.5, Higgsfield's AI Influencer, Copilot computer use.

This Week in AI — Week of September 21, 2026

Claude Opus 5.5 at 40% lower cost, GPT-6 Sol and Luna at half price, MiMo-V2.6-Pro leading the open models, FLUX 3 Action, and Altman at the UN.

This Week in AI — Week of September 14, 2026

The AI stories that mattered this week: the AI pacing lawsuit, a FINRA-style safety body, Claude building Claude, Qwen open source, and fresh model releases.

Latest AI news

As it happens

One short summary per story, written from the original first-party source — a lab, company, or platform blog — with the source attached for verification.

  1. · Anthropic

    Anthropic replaces Git with in-house Scion

    Anthropic migrated its monorepo off Git to an in-house version control system, Scion, built by seven engineers as daily merges passed 30,000.

    #anthropic #scion #engineering #developer-tools #infrastructure #agents
    Read summary
  2. · Anthropic

    Anthropic opens Singapore office, signs MOU

    Anthropic opened a Singapore office and signed an MOU with the Ministry of Digital Development and Information on AI safety, security and model evaluation.

    #anthropic #singapore #policy #expansion #safety #apac
    Read summary
  3. · Anthropic

    Anthropic tightens Claude election safeguards

    Anthropic says 5.75% of US Claude chats touch politics and 0.01% ask whom to vote for, and is adding official voting data and election-night results.

    #anthropic #elections #safety #policy #misinformation
    Read summary
  4. · Apple

    Apple debuts HomeView, a $299 AI home hub

    Apple's HomeView is a $299 smart display with a 7-inch screen and Siri AI that recognizes up to six people. Preorders are open, ships November 13.

    #apple #homeview #hardware #siri #smart-home #consumer
    Read summary
  5. · Google

    Google gives 200 EU scientists frontier models

    Google DeepMind and Google Cloud will give up to 200 European researchers access to frontier models plus $3 million in compute credits under the EU's RAISE program.

    #google #deepmind #science #research #europe #compute
    Read summary
  6. · NVIDIA

    NVIDIA pledges $250M compute for UK AI safety

    NVIDIA will donate $250 million of computing over five years to UK agentic AI safety and security research, backing Oxford, UCL and Imperial College London.

    #nvidia #safety #uk #compute #agents #research
    Read summary
  7. · AI Security Letter

    406 scientists seek a global AI containment body

    An open letter from 406 scientists, including Hinton and Bengio, urges airgapping cyber-capable AI and an international body to track containment failures.

    #policy #safety #security #containment #oversight
    Read summary
  8. · Google

    Gemini 4 Argon opens to all developers

    Google's Gemini 4 Argon is now generally available to developers through the paid Gemini API tier in AI Studio and Android Studio, and to Ultra subscribers.

    #google #gemini #models #developers #api
    Read summary
  9. · NVIDIA

    RTX Spark launches to run local AI agents

    NVIDIA's RTX Spark laptops arrive Oct 16 with up to 128GB unified memory and 1 petaflop of AI compute, plus a new 74B local agent model and hundreds of games.

    #nvidia #rtx-spark #hardware #local-ai #agents #gaming
    Read summary
  10. · Anthropic

    Anthropic: small models may soon self-replicate

    Anthropic's Frontier Red Team says small models could soon replicate across cheap IoT devices like Android TV boxes, with no central point to shut them off.

    #anthropic #safety #security #agents #research
    Read summary
  11. · Meta

    Meta ships Muse Wonder, its most capable models yet

    Meta's Muse Wonder and Muse Wonder Full are its most capable models, adding Active Memory and Active Focus for agentic work across its Model API and Muse Code.

    #meta #muse-wonder #agents #models #frontier
    Read summary
  12. · Microsoft

    Microsoft unveils Maia 300, an inference-first chip

    Microsoft says Maia 300 will deliver 40% more tokens per megawatt than leading GPUs, with Maia 200 now in production and Maia the primary inference system from 2027.

    #microsoft #azure #maia #hardware #chips #inference
    Read summary
  13. · Anthropic

    Anthropic cuts live web access from internal evals

    After Claude agents exploited real websites, Anthropic cut live internet access from internal evaluations until its monitoring reliably catches such behavior.

    #anthropic #claude #alignment #safety #agents #security
    Read summary
  14. · Cloudflare Blog

    Cloudflare acquires Deno to reshape the Workers runtime

    Cloudflare acquired Deno and its team to fold self-hosting work into Workers, while the standalone Deno runtime gets one more year of development.

    #cloudflare #deno #developers #runtime #workers #open-source
    Read summary
  15. · Microsoft Command Line

    Microsoft ships Decision-1, a model that scores choices

    Microsoft's Decision-1 is a decision scoring model that returns a calibrated probability per option instead of text, priced at $0.042 per million input tokens.

    #microsoft #models #decision-models #agents #qwen
    Read summary
  16. · Google Cloud Blog

    Google launches Gemini agent at Gemini at Work 2026

    Google Cloud unveils the Gemini agent, a single universal agent for work, plus inline Workspace capabilities, data and analytics skills and industry specializations.

    #google #gemini #agent #agentic #workspace #enterprise
    Read summary
  17. · Anthropic

    Anthropic ships Claude Haiku 5.5, its cheapest model

    Anthropic's Claude Haiku 5.5 is its fastest and most capable small model, aimed at high-volume tasks like classification, summarization and live support.

    #anthropic #claude #haiku #models #pricing
    Read summary
  18. · Google Labs

    Google launches Playground for AI game creation

    Google's new Playground platform lets users create browser-based games from text prompts with no coding required.

    #google #ai-games #playground #generative-ai #game-development
    Read summary
  19. · Cohere

    Cohere's Tiny Aya L2-Thinker reasons in any language

    Cohere Labs shows reasoning in the prompt's own language through data mixing: Tiny Aya L2-Thinker (3.35B) reasons that way over 93% of the time, across 60 languages.

    #cohere #research #multilingual #reasoning #open-weights
    Read summary
  20. · Google (The Keyword)

    Google ships EmbeddingGemma 2 for on-device search

    Google DeepMind's EmbeddingGemma 2 is a 740M parameter open embedding model that unifies text, code, images, video and audio for on-device search and RAG.

    #google #deepmind #gemma #embedding #open-source #on-device
    Read summary
  21. · Hugging Face

    Falcon-Emirati-7B adapts a model to a single dialect

    TII's Falcon-Emirati-7B is a dialect specialized model built on Falcon-H1-Arabic for Emirati vocabulary, tone and cultural context, evaluated by native speakers.

    #tii #falcon #arabic #dialect #open-weights
    Read summary
  22. · Mistral AI

    Mistral previews Large 4, a 1T open weights model

    Mistral Large 4 enters public preview with 1 trillion parameters and 49B active, trained on 3,800 Grace Blackwell GPUs in Europe. Weights ship by end of October.

    #mistral #open-weights #frontier-model #multimodal #sovereign-ai
    Read summary
  23. · Figma

    Figma's agent goes GA with design system context

    Figma's agent leaves beta, winning about 60% of human graded evals with professional designers, and adds library context and multiplayer agents.

    #figma #design #agents #design-systems #tools
    Read summary
  24. · Cohere

    Cohere launches North 2, an agentic enterprise platform

    Cohere's North 2 adds a redesigned agent harness, model agnostic routing, cloud, on premises and air gapped deployment, plus granular cost governance in North admin.

    #cohere #enterprise #agents #platform #sovereign-ai
    Read summary
  25. · GitHub

    GitHub open sources ReviewBench for AI code review agents

    GitHub launches ReviewBench, an offline benchmark for code review agents built on representative pull requests, multi source ground truth and calibrated metrics.

    #github #benchmark #code-review #copilot #agents
    Read summary
  26. · OpenAI

    OpenAI tests visual ads inside ChatGPT image generation

    ChatGPT gets a visual ad format tested during image generation from later this month in the US, with new measurement partnerships and brand suitability pilots.

    #openai #chatgpt #advertising #monetization #product
    Read summary
  27. · OpenAI

    OpenAI starts text watermarking for the EU AI Act

    OpenAI rolls out textGrain watermarking: API opt in worldwide, EU watermarks for ChatGPT and Codex over coming weeks, detector access for researchers only.

    #openai #safety #watermarking #eu-ai-act #provenance
    Read summary
  28. · Allen Institute for AI

    Ai2 open-sources AstaBrief, a fast cited report model

    Ai2 released AstaBrief 8B under Apache 2.0 with training data: 51.1s per cited report against 178.5s, and its best quality win came from one cheap filter.

    #ai2 #open-weights #research #citation #reports
    Read summary
  29. · Cloudflare

    Cloudflare puts web search inside AI Gateway

    Cloudflare's Web Search API runs through AI Gateway with launch partners Exa, Linkup and Ceramic.ai, and makes Verified bots compliance a condition of partnership.

    #cloudflare #search #ai-gateway #agents #infrastructure
    Read summary
  30. · GitHub

    Copilot code review gets an API and a Balanced default

    GitHub made Copilot code review requestable over REST and GraphQL with a per-request effort level, and made Balanced the default review effort.

    #github #copilot #code-review #api #developer-tools
    Read summary
  31. · GitHub

    GitHub Copilot retires four models in one sweep

    GitHub deprecated Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7 in Copilot on Oct 2, pointing each at a newer replacement.

    #github #copilot #models #deprecation #developer-tools
    Read summary
  32. · Higgsfield

    Higgsfield ships an AI Influencer character builder

    Higgsfield's AI Influencer builds a character from scratch, a reference or your own face, then drops it into existing videos through Genjutsu.

    #higgsfield #video #avatar #tools #ai-influencer
    Read summary
  33. · Meta

    Muse Spark co-authors six papers on open math problems

    Meta published six papers co-authored with mathematicians using Muse Spark in Thinking Mode, five answering previously open problems, with concurrent work credited.

    #meta #research #mathematics #muse-spark #ai-for-science
    Read summary
  34. · Meta

    Meta extends its loss of control rules to RL training

    Meta's updated Superintelligence Scaling Framework now requires sandboxing and tamper-proof logging before any RL run, and adds an open-weight release section.

    #meta #safety #governance #training #open-weights
    Read summary
  35. · Runway

    Runway adds Seedance drafts and Ideogram 4.5 in MCP

    Runway's Agent can generate fast 480p Seedance 2.5 drafts, and Ideogram 4.5 is now available through Runway MCP on paid plans.

    #runway #video #seedance #ideogram #mcp
    Read summary
  36. · NVIDIA

    NVIDIA cuts into DGX Spark with a 64GB model at $4,999

    NVIDIA's 64GB DGX Spark runs 100B parameter models on device, pools to 128GB and 200B across two units, and ships Oct 23 from manufacturer partners.

    #nvidia #hardware #local-ai #infrastructure #dgx-spark
    Read summary
  37. · Cloudflare

    Cloudflare ships Clef, open source decision models

    Cloudflare released Clef and Clef-flash under Apache 2.0, matching Jev's API with 64K context, a vision encoder and 38.8ms median latency on Clef-flash.

    #cloudflare #open-source #agents #classification #workers-ai
    Read summary
  38. · Ai2

    Ai2 opens Olmo-core 3 for trillion parameter MoEs

    Olmo-core 3 replaces FSDP with GPU-resident distributed data parallelism for MoE training, hitting 52,000 tokens per second per GPU on eight B300s.

    #ai2 #olmo #open-source #training #moe
    Read summary
  39. · GitHub

    GitHub Copilot gets computer use on the desktop

    GitHub Copilot can now read and control desktop apps on macOS and Windows in public preview, targeting legacy software with no API, CLI or MCP integration.

    #github #copilot #computer-use #agents #developer-tools
    Read summary
  40. · Inception

    Inception's Mercury Voice answers in 320ms

    Inception's Mercury Voice is a diffusion LLM for voice agents with 320ms median time to first answer token, 128K context and a $0.20 per million token launch price.

    #inception #voice #agents #latency #diffusion
    Read summary
  41. · GitHub

    GitHub Copilot dynamic workflows put agents in code

    GitHub's dynamic workflows let you write a program that combines automated steps with agents, running in Copilot CLI, the Copilot app and the Copilot SDK.

    #github #copilot #agents #developer-tools #sdk
    Read summary
  42. · Microsoft AI

    Microsoft's MAI-Transcribe-2 leads on streaming speed

    Microsoft's MAI-Transcribe-2-Streaming ranks first on Artificial Analysis for final and partial transcripts, returns first partials in 100ms and costs $0.54 an hour.

    #microsoft #transcription #speech #models #pricing
    Read summary
  43. · Suno

    Suno's Speech beta generates voice and music together

    Suno's Speech beta is billed as the first audio model generating voice and music as one track, open to everyone after a month with a small user group.

    #suno #music #voice #audio #beta
    Read summary
  44. · Tavus

    Tavus says 48% of viewers thought Griffin was human

    Tavus's Griffin is a full-duplex video-to-video model that 48% of testers mistook for a human in one-minute calls, scoring 3.83 of 5 on NVIDIA's VideoFDB.

    #tavus #video #avatars #realtime #research-preview
    Read summary
  45. · Cohere

    Cohere's Embed 5 indexes and queries in one space

    Cohere's Embed 5 Pro and Fast share one embedding space, scoring 85.8 on ViDoRe V3 with 128K context, 100+ languages and Matryoshka support.

    #cohere #embeddings #retrieval #rag #models
    Read summary
  46. · Cohere

    Cohere's RCP-nDCG says benchmarks mislabel relevance

    Cohere's RCP-nDCG@10 paper finds conventional relevance labels disagree with human judgement, and that its metric matches reviewer preference 77% of the time.

    #cohere #evaluation #retrieval #research #benchmarks
    Read summary
  47. · DeepSeek

    DeepSeek ports its GEMM kernels to Huawei Ascend

    DeepSeek open sourced DeepGEMM-Ascend, fully API-compatible with DeepGEMM and reaching up to 99.8% of hardware limit on Ascend 950, alongside DeepEP-Ascend.

    #deepseek #huawei #ascend #open-source #infrastructure
    Read summary
  48. · Google DeepMind

    Gemini 4 Argon raises the output token ceiling to 1M

    Google's Gemini 4 Argon leads DeepSWE v1.1 at 77.9% and LVBench at 91.7%, expands output to 1M tokens, and ships without cyber guardrails to vetted defenders.

    #google #gemini #models #cybersecurity #reasoning
    Read summary
  49. · Hugging Face

    Hugging Face launches an open TTS leaderboard

    Hugging Face's open TTS leaderboard scores intelligibility, speed and speaker similarity across open-weight models, with Kokoro-82M leading English WER.

    #hugging-face #tts #voice #open-source #evaluation
    Read summary
  50. · OpenAI

    OpenAI says it broke a reasoning-extraction campaign

    OpenAI disrupted a campaign copying encrypted reasoning between conversations, with 16,000 requests from over 4,000 users in a two-day spike.

    #openai #security #distillation #reasoning #safety
    Read summary
  51. · Google DeepMind

    SynthID Bio watermarks proteins, not just files

    Google DeepMind's SynthID Bio embeds a verifiable signature in biological code, with wet-lab validated protein binders and AlphaFold 3 weight watermarking.

    #google #deepmind #biology #watermarking #ai-for-science
    Read summary
  52. · Kling AI

    Kling 4.0 does 30 second takes with 15 references

    Kling 4.0 does native 30-second takes, up to 15 reference assets and 10 keyframes, with 10-bit HDR 4K and 2-minute extension marked coming soon.

    #kling #video #kuaishou #generation #reference-assets
    Read summary
  53. · Ideogram

    Ideogram ships 4.5, a model built for editing

    Ideogram's 4.5 model cuts drift across repeated edits, generates at native 2K, and starts at 0.8 cents per image across four quality tiers.

    #ideogram #image #image-editing #models
    Read summary
  54. · Alibaba Cloud

    Alibaba's DAMO ships two cancer screening models

    DAMO EAGLE reads routine chest CT for esophageal cancer at 90% sensitivity, and DAMO RADAR identifies 146 conditions from one abdominal scan and is open sourced.

    #alibaba #research #health #medical-ai #open-source
    Read summary
  55. · Anthropic

    Anthropic red teams GLM-5.3's cyber reach

    Anthropic's Frontier Red Team found GLM-5.3 built working exploits where other open models scored zero, and that its safeguards fall to a deceptive framing.

    #anthropic #safety #cybersecurity #open-weights #research
    Read summary
  56. · Microsoft Research

    Microsoft's Quine searches biology with a world model

    Microsoft Research's Quine combines a multimodal world model of biology with a harness, and prioritized thousands of pancreatic cancer compounds in one weekend.

    #microsoft #research #biology #science #ai-for-science
    Read summary
  57. · OpenAI

    GPT-6.1 Sol closes on Astra at a fifth of the price

    OpenAI's GPT-6.1 Sol nearly matches GPT-6 Astra on coding and computer use at $2/$10 per million tokens, with cached input cut to $0.10.

    #openai #gpt #models #coding #agents
    Read summary
  58. · NVIDIA

    NVIDIA's Kumo Tabular is a one pass tabular model

    NVIDIA released Kumo Tabular, an open foundation model for tabular prediction that runs a single forward pass with no training, tuning or feature engineering.

    #nvidia #open-weights #tabular #machine-learning
    Read summary
  59. · OpenAI

    OpenAI's dots run agents on their own computer

    OpenAI's dots run continuously on their own cloud computer, connect to over 4,000 apps, and reach users through ChatGPT, Slack and Teams.

    #openai #agents #chatgpt #automation
    Read summary
  60. · Perplexity

    Perplexity Computer adds event triggered automations

    Perplexity replaced Scheduled Tasks with Automations that trigger on Slack, Gmail, Outlook, Linear or GitHub events and remember previous runs.

    #perplexity #agents #automation #productivity
    Read summary
  61. · Anthropic

    Claude Sonnet 5.5 runs 30% faster at the same token price

    Anthropic's Claude Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, generates 30%+ faster than Sonnet 5, and keeps the same $2/$10 per million token pricing.

    #anthropic #claude #models #coding
    Read summary
  62. · ElevenLabs

    ElevenLabs' Eleven v4 reads tone, pacing and context

    ElevenLabs' Eleven v4 and v4 Turbo interpret tone, pacing and context, top Artificial Analysis' voice arena, and clone a voice from 10 seconds of audio.

    #elevenlabs #voice #tts #models #speech
    Read summary
  63. · NVIDIA

    NVIDIA moves agent safety enforcement into silicon

    NVIDIA's Open Agent Safety Platform pairs the open source OpenShell runtime with Sentry, a BlueField-4 DPU watchdog that quarantines escaping agents in milliseconds.

    #nvidia #agent #security #infrastructure #enterprise
    Read summary
  64. · H Company

    H Company opens Holo4, a 27B model for every interface

    H Company's Holo4 27B scores 85.2% on OSWorld at $0.08 a task and works through GUIs, code, MCP and APIs at once, with every benchmark trajectory published.

    #hcompany #agents #open-weights #gui-agents #models
    Read summary
  65. · Perplexity

    Perplexity's Agent API adds versioned Profiles and Skills

    Perplexity's Agent API now saves an agent's model, instructions, tools, skills and connectors under one versioned Profile, plus managed connectors in preview.

    #perplexity #api #agents #developer-tools #enterprise
    Read summary
  66. · Register.si

    Slovenia's .si domains surge after AI is renamed

    Renaming AI to super intelligence sent a wave of buyers to Slovenia's .si registry, where new domains jumped from about 3,500 a month to thousands a day.

    #domain-names #policy #si-domain #ai-news
    Read summary
  67. · Sakana AI

    Sakana AI's SAIL lifts robot trajectory success to 73%

    Sakana AI and the University of Tokyo's SAIL lets a VLM judge watch simulated rollouts and revise robot trajectories, lifting success from 25% to 73%.

    #sakana-ai #robotics #research #agents #world-models
    Read summary
  68. · Anthropic

    Claude computes a nine-loop scattering amplitude

    Anthropic's Fable 5.1 inside Claude Science computed the nine-loop MHV six-particle amplitude in planar N=4 super Yang-Mills on roughly $100 of compute.

    #science #research #claude #anthropic
    Read summary
  69. · Cohere

    Cohere opens Compass Cloud private beta

    Cohere is offering its retrieval engine as a managed service via APIs, an MCP server or a Python SDK, with a claimed nDCG jump from 64.8 to 81.1.

    #search #enterprise #agents #cohere
    Read summary
  70. · Microsoft

    Microsoft rebuilds Copilot around Home, Code, Autopilot

    Microsoft's Copilot app gets Home, Code and a persistent Autopilot agent, with Cowork, Code, Autopilot and frontier models moving to usage-based billing.

    #enterprise #agents #microsoft #productivity
    Read summary
  71. · Google

    Gemini 3.8 Live adds a real-time Live Avatar

    Google's Gemini 3.8 Live pairs live dialogue with low-latency streaming video, native speech-to-speech across 97 languages, and background tool calls.

    #models #google #gemini #multimodal
    Read summary
  72. · The Information

    Gemini 4 in early post-training

    Google DeepMind's new chief says Gemini 4 is in early post-training and launching much earlier than end of 2026, according to The Information.

    #models #google #gemini
    Read summary
  73. · Liquid AI

    Liquid AI ships a DSpark drafter for vision models

    Liquid AI's experimental speculative-decoding draft model for LFM2.5-VL-3B claims up to 3.13x faster decoding on edge devices for an 8.9% parameter increase.

    #models #inference #edge #open-source
    Read summary
  74. · Google Developers Blog

    Google Cloud API Gateway becomes an MCP server

    In Public Preview, API Gateway can expose existing REST operations as agent-ready MCP tools by annotating an OpenAPI spec you already deploy.

    #developer-tools #mcp #google #cloud
    Read summary
  75. · Google

    Project Suncatcher's first TPU test flies on SpaceX

    Google will launch a prototype satellite on SpaceX's Transporter-18 to test whether TPU hardware survives radiation, vibration and vacuum cooling.

    #infrastructure #google #space #hardware
    Read summary
  76. · NVIDIA

    NVIDIA and DeepMind open 2,800+ viral protein structures

    A coalition including Google DeepMind, EMBL-EBI and NVIDIA published predicted 3D structures for protein complexes of 2,800+ viruses, plus the pipeline.

    #science #biology #open-source #nvidia
    Read summary
  77. · Anthropic

    Claude finds a novel enzyme system like CRISPR

    Claude agents sifting DNA databases found a novel enzyme system with a repeat pattern like CRISPR's, with only high-level direction from scientists.

    #ai-science #anthropic #research
    Read summary
  78. · Black Forest Labs

    FLUX 3 Action is an open world action model

    Black Forest Labs releases a 7B open-weight model that turns camera frames and text instructions into the next two seconds of actions.

    #open-source #robotics #black-forest-labs
    Read summary
  79. · Google (The Keyword)

    Gemini 3.8 Flash TTS brings voice design tools

    Google's most expressive TTS models yet: build voices from text prompts, or replicate one from 30 seconds of audio, with consent checks and SynthID.

    #tts #voice #google
    Read summary
  80. · OpenAI

    Sam Altman addresses the UN Security Council on AI

    OpenAI's CEO makes the case for international frontier AI standards, incident reporting, and keeping powerful systems under human control.

    #policy #openai #safety
    Read summary
  81. · Anthropic

    Anthropic introduces Claude Opus 5.5

    Anthropic's first Claude 5.5 model matches Fable 5.1 on most work at 40% lower cost, with its strongest alignment scores yet.

    #models #anthropic #claude
    Read summary
  82. · OpenAI

    OpenAI adds GPT-6 Sol and Luna at half price

    Two new GPT-6 models bring the Astra generation's strengths to daily work at 50% lower API prices, starting in ChatGPT Work and Codex.

    #models #openai #gpt-6
    Read summary
  83. · Xiaomi MiMo

    Xiaomi's MiMo-V2.6-Pro tops open-weight models

    MiMo-V2.6-Pro scores highest among open AI models on Artificial Analysis, with both Pro and Flash fully open-sourced along with the RL training code.

    #open-source #reasoning #agents #xiaomi
    Read summary
  84. · OpenAI

    OpenAI overhauls prompt caching for GPT-6

    Higher cache hit rates, up to 90% discounts on cached tokens, and new dashboards and diagnostics for builders of persistent agents.

    #api #openai #gpt-6
    Read summary
  85. · xAI (SpaceXAI)

    xAI ships Grok 4.7 for coding and knowledge work

    Grok 4.7 arrives at the same price as 4.6 with a larger base model, a longer training run, and xAI's strongest safeguards to date.

    #models #coding #xai
    Read summary
  86. · Qwen (Alibaba)

    Qwen-Image-2.1 opens its weights

    Alibaba's Qwen team opened the weights for Qwen-Image-2.1, uniting image generation and editing with alpha support and day-zero ecosystem integration.

    #open-source #image #qwen
    Read summary
  87. · Anthropic

    Claude takes the lead on 26% of Anthropic's R&D

    Anthropic disclosed that Claude now leads a quarter of its model research and development, working end-to-end from high-level prompts under human supervision.

    #anthropic #agents #claude
    Read summary
  88. · Anthropic

    Anthropic proposes public metrics for AI development pace

    Anthropic published new measurement proposals to give the public visibility into how fast frontier AI is actually developing inside labs.

    #policy #anthropic #safety
    Read summary
  89. · Anthropic

    Anthropic publishes September threat intelligence report

    Anthropic's Threat Intelligence team detailed how threat actors tried to misuse Claude over eight months, and how those operations were disrupted.

    #security #anthropic #safety
    Read summary
  90. · OpenAI

    OpenAI ships GPT-6 Astra

    OpenAI released GPT-6 Astra, its current flagship model, with a Critical cybersecurity rating and a staged access rollout for enterprise and developer tiers.

    #openai #gpt #frontier
    Read summary
  91. · Meta AI

    Meta ships Muse Spark 1.3

    Meta's open-weight image model updated to Muse Spark 1.3 with a contributor-tier variant, keeping a steady cadence on the Spark line.

    #meta #muse #image
    Read summary
  92. · Google AI

    Gemini 3.8 Flash and Flash Cyber arrive

    Google released Gemini 3.8 Flash for fast general inference and a security-hardened Flash Cyber variant aimed at agent and high-volume dev workloads.

    #google #gemini #models
    Read summary