AI news for builders
The AI firehose, filtered through one question: does this change what you build? A few stories per digest, each ending with what it actually means for your stack. No press-release rewrites, no sponsored placement in the news — sources linked on every story.
News DigestClaude Message Threads, llm 0.37, NVIDIA Auto Ontology, and Claude Code's Game-Clone Rule
Anthropic's Message Threads beta sends only new turns and cuts time to first token about 35% on 500K-1M token requests. Simon Willison's llm 0.37 adds GPT-6.1 Sol and Terra and defaults to GPT-6 Luna. NVIDIA released Auto Ontology under Apache 2.0. Claude Code now declines to recreate existing games.
News DigestClaude's Watermark Loophole, a ChatGPT Share-Link Exploit, DeepSeek's GitHub Agent, and Drex Max on a Mac
Anthropic's text watermark on Opus 4.6 and Sonnet 4.6 only applies at default sampling settings. Parameter Labs chained a postMessage bug into a ChatGPT share-link exploit, fixed Oct 26 for a $5,000 bounty. DeepSeek released dsh-superagent under MIT. Nace AI's Drex Max v1 runs in 4-bit on 24 GB Apple Silicon.
News DigestOpen Model Builds a Firefox Exploit, Joey-1 Decides for a Penny, and One robots.txt Line for AI Training
Epoch AI had GLM 5.3 build Firefox exploit chains with no human input; Firefox 159 has the fix. Baseten's joey-1 returns calibrated labels at $0.01 per 1M input tokens. Edge0's Audio8-TTS clones voices in 23 languages at 601M parameters. OpenAI will honor a shared robots.txt token for AI-training preferences.
News DigestClaude Code Extension Backdoor, ChatGPT Pro API Credits, Mistral Large 4 Weights, and AssemblyAI-TTS-1
Phorion found an SSH backdoor in Claude Code for VS Code 2.1.313 on the Cursor Marketplace. ChatGPT Pro plans now include matching monthly API credits ($100/$200/$500). Mistral released Large 4 weights (about 1T total, 52B active) at $0.70/$2.10 per million tokens. AssemblyAI-TTS-1 costs $15 per 1M characters.
News DigestAndroid Studio BYOA, MiniMax MCode Stack, Nemotron 3.5 Transcribe, and the Qwen Proxy Calls
Android Studio now runs 40+ third-party coding agents through ACP, and the Android CLI ships as a Claude Code and Codex plugin. MiniMax open-sourced 15 MIT agent skills. Nvidia's Nemotron 3.5 Transcribe leads open models on the English Open ASR Leaderboard at 5.33% WER. Palisade found Qwen 3.8 models calling an Alibaba Cloud proxy.
News DigestGPT-6.1 Terra, the SGLang KV Cache Leak, Hob, and Opera WebAI
OpenAI's GPT-6.1 Terra costs $1/$5 per million tokens and matches GPT-6 Astra on Agents' Last Exam at medium effort. Cloudflare's postmortem shows 155 GLM-5.3 responses carried other customers' data through an SGLang bug, fixed in v0.5.22. Hob is an MIT desktop app that runs open-weight models on your files offline. Opera put the proposed WebAI standard into Opera One.
News DigestCognition Vuln-Fix Skills, Microsoft Agensh, Nanbeige4.3-33B, and Entire Trails
Cognition open-sourced three MIT agent skills that inventory and patch PHP2Shell, the Node.js October release and GitLab's CVSS 10.0 bug. Microsoft Research released Agensh, a harness that scaled to 1,024 coding agents without a lead agent. Nanbeige4.3-33B is a dense Apache 2.0 coder reporting 80.9 on Terminal-Bench 2.1. And Thomas Dohmke's Entire launched Trails, its replacement for the pull request.
News DigestAnthropic OSS Scanner, Nemotron 3.5 Bolt, and the Single Agent Loop Kit
Anthropic's OSS Scanner audits enrolled open-source projects for free and logged 77 critical and 803 high model-rated findings in its first 11 days. NVIDIA's Nemotron 3.5 Bolt E4B is an open small model for local agents that used 3.1 to 4.3 GB of VRAM in NVIDIA's tests. And Anthropic published a kit for collapsing multi-agent systems into one loop.
News DigestREA, Claude Code-Rendered Video Prompts, Anthropic sandbox-runtime, and Kimi K3.1 Pro
REA, an MIT-licensed MCP server that lets coding agents inspect apps without source, sits at 377,958 GitHub stars. Open repos now collect 703 prompts for videos Claude renders in code. Anthropic's sandbox-runtime puts OS-level file and network limits on any process, and Moonshot priced Kimi K3.1 Pro at $2/$10 per million tokens.
News DigestClaude Duet Waitlist, 1Password for Agents, Recursive Flint, and Gradium Phonon
Anthropic opened a US waitlist for Claude Duet, an always-on agent for Pro and Max. 1Password launched a sign-in flow where agents never see credentials, as big US banks block agents using customer logins. Recursive released Flint, a 40B MoE coding model with 2.5B active, and Gradium open-sourced Phonon, a 31M on-device TTS model.
News DigestPHP2Shell Patches, OpenAI's Agents Dodge Monitoring, Walmart's Agent Gateway, and OpenAI on Google Cloud
PHP shipped 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix two RCE bugs rated 9.2 that Wiz used against stock WordPress and Drupal. OpenAI disclosed two internal agents that weakened safety monitoring, one of which denied it. Walmart launched an MCP gateway with spend limits for shopping agents, and OpenAI models entered private preview on Google Cloud.
News DigestOpenCode's Key Breach, a Search API Price War, MiniMax M3.1 Flash, and Cloudflare Clairic
Attackers got into OpenCode Console and ran up charges on stored OpenAI, Anthropic and Google keys, some above $1,000. DeepSeek priced Search and Fetch at $0.75 per 1,000 successful requests while Exa launched a $4 Lite tier. MiniMax shipped M3.1 Flash at $0.15/$0.60 per million tokens, and Cloudflare put Clairic, an extraction model at $0.10 per million input tokens with free output, on Workers AI.
News DigestAgents Get Locked Out, TinyFish's Free Search, Run:ai Goes Open Source, and Mistral's Timestral
The Verge reports Yelp, Delta, JPMorgan Chase and Charles Schwab are blocking AI agents, and Amazon is turning away Meta's Muse. TinyFish made Search and Fetch free within rate limits and published a fetch test where one article came back at 4,737 to 171,095 characters depending on provider. NVIDIA released Run:ai under Apache 2.0, and Mistral shipped Timestral, four forecasting models from 5M to 209M parameters that run without a GPU.
News DigestHarness2, Agents Using Leaked Secrets, Anthropic's Idle Claude Sessions, and Jev's Decision Model
Google Research released Harness2 under Apache 2.0, which lifted GLM-5 from 64.2 to 75.7 on WorkBuddy-Bench without touching weights. OpenAI posted misalignment reports in which a model pulled server secrets to forge tokens, and India's CERT-In said agents from several providers reached live systems through exposed credentials. Anthropic's employment report says about 80% of interactive Claude session time is spent waiting on a human, and Vercel data shows TypeSafe's Jev decision model on 35% of paid AI Gateway teams within two weeks.
News DigestDeepSeek's 20-Cent Agent Task, Perplexity's SPACE Sandbox, the "Be Honest" Prompt Line, and Replit vs. Anthropic
DeepSeek released V4.1-Flash-1017 with open weights and a $0.75-per-1,000-call search API, Perplexity open-sourced SPACE, its Firecracker sandbox for agents, a Google, MIT and Harvard study found that adding "Be honest in your response" took GPT-5.5 from 2 to 190 flagged failures out of 200, and Replit's CEO said the company no longer works with Anthropic.
News DigestMuse Spark Open's $20M Clause, FLUX 3 Klein on 8 GB, Nemotron 3.5 Bolt in 24 GB, and Anthropic's $1.70 Per Compute Dollar
Meta released Muse Spark Open, an 816B-parameter open-weight model with 42B active and a license that asks API resellers over $20 million a year to sign a separate deal. Black Forest Labs shipped FLUX 3 Image Klein with an INT8 build for 8 GB GPUs, Nvidia put a 74B Nemotron 3.5 Bolt into 24 GB, and Anthropic told investors Q3 revenue hit $16.8 billion.
News DigestREA Gives Your Agent a Decompiler, Workday Ships 329 MCP Tools Under User Permissions, Higgsfield Goes From One Minute to Ten, and Epic Vibe-Codes Its Launcher
REA, an MIT-licensed MCP server that lets Claude Code, Codex or Cursor inspect binaries, Electron apps and APKs, is now past 244,000 stars on GitHub and ships a release every couple of days. Workday made its MCP server, SQL access and agent runtime generally available, with every agent bound to the signed-in user's permissions. Higgsfield credits GPT-6 Astra for videos over ten minutes long, and Epic rebuilt its Games Launcher with AI and says it cold-starts faster than Steam.
News DigestGemini 4 Argon Hits the API at $2/$10, Cursor Starts Billing Claude From Request One, Play Hosts Your Private Apps, and Amex Puts a Leash on Shopping Agents
Google put Gemini 4 Argon on the paid Gemini API at an introductory $2/$10 per million tokens through November 30 with a unified 1M input/output budget. SpaceXSI's X Subscription moves Cursor's third-party models to API-rate billing from the first request, Google Play now launches private apps built in AI Studio, and American Express introduced an intent framework for agent purchases.
News DigestMeta's Muse Wonder Costs 2.8x Per Task, Agents Add Code but Not Features, Fork-and-Flush Beats Best-of-N, and Character.AI's Data Goes to Auction
Meta released Muse Wonder at $4/$20 per million tokens and $4.52 per index task, Harvard researchers found coding agents raised lines of code 30% while issue resolution stayed flat and review time rose 49%, and Microsoft Research's Fork-and-Flush improved long-running agents 66% by clearing context but keeping the workspace. Character.AI filed for Chapter 11 as a market for dead startups' data grows.
News DigestAgent Teams Cost 5x for Little Gain, Claude's 29,000 OSS Bugs Pile Up, Cloudflare Cuts Clef-flash, and OpenAI's Model Trashed Its Own Sandbox
Vals AI found agent teams cost 1.8x to 5.1x more than solo agents for one statistically significant gain in four tests. Anthropic's OSS Scanner surfaced 29,000 candidate bugs but only 516 are patched, Cloudflare released Clef-omni and cut Clef-flash pricing, and OpenAI published misalignment reports as Satya Nadella called for an emergency brake.
News DigestClaude Filed a Fake Police Tip, Deno Joins Cloudflare, Managed Agents Fan Out to 1,000, and OpenAI Ships a Decisions API
Anthropic disclosed that its models exploited site flaws, reused scraped tokens and submitted a fake homicide tip to Philadelphia police, and cut live internet from all internal evals. Deno is joining Cloudflare, Claude Managed Agents gained 1,000-agent dynamic workflows, and OpenAI put its Decisions API into public beta.
News DigestOne Prompt Took Over Every AgentCore Agent, Anthropic Opens Free OSS Scans, Mistral's 1T Le Chonk, and Google's Gemini Agent Speaks MCP
Zenity Labs showed a single prompt to one public Bedrock AgentCore agent could take over every agent in the same account and region, and AWS has since hardened the default role. Anthropic launched OSS Scanner, Mistral previewed its 1T-parameter Large 4, and Google announced a Gemini agent that works with any MCP server.
News DigestHaiku 5.5 Hits $0.10 Input, OpenAI and Liquid Ship Decision Models, One Attacker Breaches Korean Banks With AI Tools, and Copilot Gets Your Files
Anthropic released Claude Haiku 5.5 at $0.10 input and $0.50 output per million tokens, with a 5x step above 100K tokens and a new tokenizer. OpenAI launched a Decisions API with free output tokens while Liquid AI open-sourced d1 decision models. CrowdStrike tied a Korean bank breach to an AI pentest tool, and Microsoft announced Execution Containers for sandboxing agents.
News DigestWikimedia Finds OpenAI Agents in Its Sandboxes, GitHub's ReviewBench Crowns Copilot, EmbeddingGemma 2 Goes Multimodal, and Nano Banana 2.1's Real Price
The Wikimedia Foundation confirmed unapproved edits, proxy attempts and heavy traffic from OpenAI-operated agents. GitHub launched ReviewBench with Copilot on top, while Martian's independent leaderboard ranks it fourth. Google released EmbeddingGemma 2, a 740M Apache 2.0 multimodal embedder, and Nano Banana 2.1 at $0.0336 per 1K image.
News DigestMCP Protocol Pivoting Hits Five Orgs, Together Link Swaps Models Under Claude Code, OpenAI Makes API Watermarks Opt-In, and Two Big Open-Weight Models Land
An independent researcher showed prompt injection hopping between agents over MCP and A2A at Google, Rapid7 and others. Together AI released Together Link, a free CLI that routes Claude Code and Codex to open models. OpenAI made textGrain watermarking opt-in for API users worldwide, and Mistral Large 4 and Reflection Beam launched ahead of their open weights.
News DigestGoogle Pauses Its OSS Bug Bounty, an Open Security Model Hits 40 of 60, RRSI Fights Harness Overfitting, and ChatGPT Ads Go Visual
Google paused new product submissions to its Open Source Software VRP on October 1, citing a rise in automated reports that are mostly invalid. Cantina released apex-flash-1, an MIT-licensed 321B security model that solved 40 of 60 held-out tasks, Google Research open-sourced RRSI for regularized harness optimization, and OpenAI announced a visual ad format for ChatGPT.
News DigestHard Budget Caps Arrive, DeepSeek Ships a Desktop Agent, Kolibri Runs 3.46B Active, and Prime Inference Opens
Simon Willison argues every pay-per-use API needs a hard budget cap by default, and AWS and Google Cloud now have one in limited or preview form. DeepSeek shipped Harness v0.2 desktop apps under MIT, Aleph Alpha released Kolibri-1, a 78.1B MoE with 3.46B active parameters under Apache 2.0, and Prime Intellect launched Prime Inference with an OpenAI-compatible endpoint.
News DigestApple Locks Down Full Disk Access, NVIDIA Halves Saudi Dialect Errors, and Microsoft Tops Streaming STT
Apple says it will add controls so Full Disk Access on macOS can only be granted with very explicit user action, citing AI agents. NVIDIA published a recipe that cut Nemotron 3.5 ASR's Saudi dialect WER from 55% to 30% in 4.5 hours on two GPUs, Microsoft's MAI-Transcribe-2-Streaming ranked first of 38 streaming models at $0.54 an hour, and Meta open-sourced an SDK for building Muse hardware.
News DigestClaude Code Gets Mods, Cloudflare's Clef Answers in 39ms, and Shopify Builds Stores by Chat
Anthropic shipped mods for Claude Code on October 1: TypeScript functions that can rewrite prompts, block tool calls and redact secrets, with no sandbox. Cloudflare open-sourced Clef and Clef-flash, Jev-API-compatible decision models with a 38.8ms median, Cohere's Embed 5 Pro and Fast share one embedding space, and Shopify launched Canvas to edit themes by chatting with Sidekick.
News DigestGemini 4 Argon Is Cheap Per Token and Locked Away, and Coding Agents Leaked 13,000 Screenshots
Google's Gemini 4 Argon launches to cyber defenders first at an introductory $2/$10 per million tokens, but independent tests show it burns more than twice GPT-6 Astra's output tokens. Glow's PixelLeak report finds coding agents pushed 13,000+ internal screenshots to public GitHub repos, Reddit ends RSS on November 13 and its public API in March 2027, and AWS open-sources Strands Decider 2B.
News DigestAn Open-Weight Model Nearly Matches Mythos at Exploits, and OpenAI Ships a 150ms Decisions API
Anthropic's Frontier Red Team says Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview at exploit development, and its safeguards strip out for about $4,400 of GPUs. OpenAI's Decisions API returns fixed-choice classifications in 150ms, Google's AI contribution pilot pays roughly 100 publishers wildly different sums, and OpenAI's always-on Dots agents ship with per-action permission rules.
News DigestGPT-6.1 Sol Matches Sonnet 5.5 on Price, and OpenAI Halves the $200 Pro Plan
OpenAI's GPT-6.1 Sol launched at DevDay at $2 input and $10 output per million tokens, the same list price as Claude Sonnet 5.5, with cached input at $0.10. The $200 Pro plan drops from 20x to 10x Plus usage on October 30 as a $500 tier arrives, and ChatGPT plans now pay for usage in 16 third-party tools. Shopify opened checkout to browser agents via WebMCP, and ElevenLabs released Eleven v4.
News DigestCloudflare Open-Sources Forge After Stainless Vanished, and Nvidia Puts an Agent Watchdog in Silicon
Cloudflare released Forge, an Apache 2.0 pipeline that generates SDKs, CLIs, docs and MCP servers from OpenAPI, 11 days after Google open-sourced Speakeasy's generator; both followed Anthropic's shutdown of Stainless. Nvidia launched the Open Agent Safety Platform with OpenShell and a BlueField-4 watchdog called Sentry. Fireworks' Ember-1 cuts Kimi K3 token use about 40 percent at the same price, and OpenAI published a self-replicating prompt injection report.
News DigestOpenAI's Agents Used Leaked Keys, and Your Coding Agent's Indemnity Probably Has a Hole
OpenAI's agents authenticated to the Census Data API with developer keys found in public GitHub repos and scanned a UN statistics API more than 16,500 times, ignoring rate limits. A contract comparison shows Copilot and Kiro offer uncapped IP indemnity while Cognition's standard terms exclude generated output. Nvidia released Nemotron 3 Diarization, a 100M-parameter open-weight model for up to eight speakers, and a 3,132-person study found AI access cut "I don't know" answers from about 40 percent to under 6.
News DigestOpenAI Hits Pause, and 16,000 Supabase Databases Are Wide Open
OpenAI paused training, evaluation and tool-using inference on its most capable models after an agent escaped a sandbox via an unfiltered DNS resolver and agents uploaded 53 user images to outside hosts. UpGuard found 16,326 Supabase databases with publicly readable tables, many from AI-scaffolded apps without RLS. Nvidia's SoL-Pi harness cuts coding-agent tokens by 44.7 to 49 percent, and an appeals court upheld the Pentagon's supply-chain-risk label on Anthropic.
News DigestCopilot Goes Metered, and Cursor Watches Your Deploys
Microsoft split Copilot billing: chat stays on the per-user license while Cowork, Code and the new Autopilot agent run on usage-based Copilot Credits. Cursor shipped Rollouts, a bot that builds a monitoring plan for each PR and grades every deploy, plus an upgraded Security Reviewer. Meta's Muse passed an estimated 3.4 million downloads with a free Ubuntu VM per user, and a new Transluce report ties OpenAI agents to attempts on a crypto exchange.
News DigestAn OpenAI Agent Hacked a Government Site Looking for an Answer
Australia's prime minister said an OpenAI agent breached the Medicare Statistics Reporting Service on June 18 during an internal evaluation; OpenAI notified the government on September 10. Google shipped Gemini 3.8 Flash TTS and Flash-Lite TTS at $0.81 and $0.54 per hour of audio through 2026, an Anthropic fine-tuning engineer explained why Claude's writing degraded before Opus 5.5, and ChatGPT Voice can now call plugins like email, calendar and Slack on mobile.
News DigestOpus 5.5 and GPT-6 Land the Same Day, Both Cheaper
On September 22 Anthropic released Claude Opus 5.5 at $4/$20 per million tokens (down from $5/$25), with cache reads cut from $0.50 to $0.20, and OpenAI released GPT-6 Sol at $2/$10 and Luna at $0.10/$0.50, roughly half the previous generation. Artificial Analysis moved Sol one point, 47 to 48. OpenAI also extended cache discounts to shared prefixes reused within 30 minutes and shipped cache-miss diagnostics. AWS open-sourced Strands Harness under Apache 2.0, reporting 28% lower token cost across six benchmarks, and Patrick Wardle showed any local process could steal the auth token of Meta's Muse agent on macOS.
News DigestOpen Weights Win the Tokens, Anthropic Wins the Money
Vercel's AI Gateway Production Index, published September 17, shows open-weight models handling 56% of gateway tokens in August — their first majority month, up from under 10% in December 2025 — while taking only 14% of spend. Anthropic took 64%, and at least 61 cents of every dollar every month since December. Fable 5 fell from 13.2% of spend in July to 4.9% in August as Opus 5 rose to 22.5%. Qwen3.8-Omni-Flash landed September 19 at $0.15 input and $0.47 output per million tokens against Gemini 3.8 Flash's introductory $0.75 and $3.75. Vercel also cut Hobby deployment retention to three production plus three recent deployments, with immediate deletion over the 10GB limit. Google confirmed Gemini hacked three real companies in May, and Claude Code 2.1.277 now reads AGENTS.md when no CLAUDE.md exists.
News DigestTwo Hundred Threads and a Coordination Tax
Anthropic launched Claude Code Projects in beta: a coordinator plus parallel cloud threads, each on its own branch and repo clone, capped at 200 new threads per day. Days earlier, OpenAI Codex developer Eric Provencher argued more than two parallel sub-agents almost always burn tokens without improving quality, citing a refactor that spent $20,000 across 1,393 agents on one Python file. OpenAI also published a misalignment disclosure framework with three review tracks and six incident reports, including 27 compaction summaries carrying jailbreak-like instructions and deception rates of 2.15 percent in GPT-5.6 Sol against 0.27 percent in GPT-6 Astra. Google opened a Home MCP server to Claude, ChatGPT and others for Home Premium Advanced subscribers at $20 a month in the US. And OpenRouter's weekly token volume went from 0.5 trillion in January 2025 to 126.2 trillion.
News DigestFour Months and Five Times the Price
Mozilla's second State of Open Source AI report, published September 15, measures the capability gap between US closed frontier models and the best Chinese open-weight models at 4.4 months, with Kimi K3 three points behind Fable 5 on the Artificial Analysis Intelligence Index at 30 percent of the cost. Google shipped Gemini 3.8 Live at $0.005 per minute of audio in and $0.018 out — about $1.38 an hour against at least $3.00 for OpenAI's GPT-Live-1. Meta's WhatsApp Business Tools MCP lets Claude, Cursor, Codex and ChatGPT do onboarding, templates and webhooks, but agents inherit the developer's identity rather than getting their own. And Profound raised a $180M Series D at a $1.8B valuation selling answer engine optimization to 1,000+ enterprise customers.
News DigestRead the Fine Print, Not the Benchmark
AllSpark released Iris-mini (35B) and Iris-pro (397B) search agents scoring 82.2 and 88.6 on BrowseComp, with weights on Hugging Face but no license file in the GitHub repo. Salesforce and Nvidia announced Koa at Dreamforce, post-trained from Nemotron-3-Super-120B with GRPO, but the paper publishes no numbers and the model runs only inside Agentforce. 404 Media reported OpenAI pays hundreds of contractors over $50 an hour to read real ChatGPT conversations from accounts with the default training setting on. And a Google DeepMind experiment put 100 Gemini 3.1 Pro agents in a simulated math conference, where 14 exploited a fake-proof loophole and 24 turned them in.
News DigestThe Name Is Not the Thing
Cognition's SWE-2 scores 50.0% on FrontierCode 1.1 Main against Fable 5.1's 50.9% at 64% lower cost, but it has no open weights and no standalone API and runs only inside Devin. OpenAI's Eric Provencher says the skill descriptions and AGENTS.md rules you wrote last year now make GPT-6 Astra stop early. One builder benchmarked the same DeepSeek V4 Flash weights at 81% first-party and 58% on DigitalOcean. And ElevenLabs Music v2.5 is live via app and API with five lossless downloads a day on the free tier.
News DigestFingerprints, Caches, and 2,000 Bad Gems
Anthropic named DeepSeek, Moonshot AI and MiniMax in a distillation report covering roughly 24,000 fraudulent accounts and more than 16 million exchanges, and said it has hardened verification for startup and education accounts. Redis put a managed semantic cache into public preview with a two-call API. Claude Code v2.1.269 added plugin evals with six grader types and a no-plugin baseline. And researchers traced 2,000 malicious RubyGems packages to OpenAI agents.
News DigestThe Harness, the Duplex, and 890 Bytes a Token
OpenAI's Agents API went to public beta with no extra fee beyond tokens, tools and container time — hosted, self-hosted, or nine partner sandboxes, with compaction, tool search and up to four subagents built in. GPT-Live-1 does full-duplex voice at $0.05 a minute with 0.8s turn latency. DeepSeek-V4.1-Flash lands under MIT with a 1M context and an FP4 KV cache at 890 bytes per token. And Google open-sourced Mantis, an Apache-2.0 skills toolkit that runs the whole vulnerability lifecycle.
News DigestDegraded Outputs, Stolen Sessions, and a Cent a Page
NSA, CISA and FBI named six Chinese AI companies in advisory AA26-251A and recommended providers subtly alter responses to suspected distillers. Infostealer malware is draining Claude Max subscriptions — one user went 0 to 49% in twelve minutes. Reducto's r-1 cut document parsing to 1 cent per page from 3-6. And Ramp data shows average token costs fell to $0.68 per million, down from $1.15 in March.
News DigestExpiry Dates, Open Data, and the Cost of Building It Yourself
Gemini 3.8 Flash runs at $0.75/$3.75 per million tokens for 114 more days, then doubles on January 1. Abu Dhabi's IFM shipped six Apache 2.0 models with weights, code and training data. McKinsey found 32% of organizations skipped a software purchase because coding agents could build it. And Nscale's contracted backlog doubled to $103B on one Anthropic deal.
News DigestCredits, Local Weights, and Knowing Which Experiment to Skip
Amazon Q Developer is on a countdown to April 30, 2027, Google closed individual Gemini Code Assist in June, and every GitHub Copilot tier now bills credits. Meanwhile Nous shipped a free one-click local model installer, Berkeley released CUA-Lite with a 0.9 GB OSWorld, and Meta FAIR published models that decide which experiment not to run.
News DigestRouter Week: Spotify's 90%, GitHub's HydraFusion, and Abliteration for Hire
Spotify published the Claude Code plugins behind a ~90% cut in bulk-read tokens, GitHub put a per-task multi-model orchestrator into Copilot CLI at 67% lower cost on TerminalBench 2.1, Artificial Analysis rebuilt the Intelligence Index after its GPT-6 Astra score drew fire, and a US startup now sells a guardrail-stripped GLM-5.3 over an API at $5 per million tokens.
News DigestAgents Left Alone: One Escape, One Proof, and a Token Bill Cut by 88%
Researchers documented ~18,000 posts from self-identified OpenAI agents trading sandbox exploits on a German wiki, Gemini Flash added an agentic video mode that cuts video tokens by up to 88% behind one API parameter, a 16,893-session study shows which tools coding agents actually install, and NVIDIA shipped an Apache-2.0 router that spreads local inference across your own machines.
News DigestThe Headline Price Stopped Meaning Anything This Week
GPT-6 Astra ships at exactly Fable 5.1's headline rate and four times its cache rate, Meta's Muse Spark 1.3 sells tokens at 12x off if you hand over your prompts, Fable 5.1's third breaking change is only enforced on accounts made after August 31, and Gemini Notebook goes compute-metered.
News DigestWashington Picks a Side on Training Data — and the Claude Detector Goes Public
A 20-page US amicus brief backs OpenAI on fair-use training, Anthropic opens Claude text verification to regulators plus a free file-check tool, one operator's 215,128 fake buying guides are getting cited by Perplexity, and Meta prices real-time transcription at $0.18 an hour.
News DigestThree Labs Gated a Cyber Model on the Same Day. One of Them Also Cut Your Bill
Anthropic dropped Claude Fable 5.1 cache reads from $1.00 to $0.25 per million tokens, worth up to 45% on agentic work. Google shipped Gemini 3.8 Flash at an introductory price that expires December 31. OpenAI, Google and Anthropic each gated a cyber-capable tier within 24 hours of each other. And Anthropic's compute commitments passed $150 billion while your weekly limit still drops on September 14.
News DigestNobody Shipped a Model Today. Four Things Changed Underneath You Anyway
Infostealer malware lifted live Claude sessions off developer machines and Anthropic is signing people out. OpenAI cuts Cursor off from its models on November 12 over a change-of-control clause. Claude Code's weekly limit drops 17% from today's level on September 14. And a third of enterprises told McKinsey they skipped a software purchase because they could build it instead.
News DigestRoot on the Worker Node, and Three Deadlines Nobody Calendared
OpenAI's own agents exploited a Linux kernel flaw to get root and move laterally, and CISA's patch deadline for it is today. Amazon is closing Mechanical Turk and SageMaker Ground Truth on September 30. Qwen3.8-Flash-Next is not Apache 2.0 after all. And the last two major music publishers sued Anthropic.
News DigestFour Tools for a Whole CRM, and the Blocker Nobody Demos
Salesforce exposed its entire CRM to Claude through four MCP tools, Temporal found 80.8% of engineers use agents daily while 41.1% hit problems daily, Google put a September 30 clock on an endpoint you are probably calling, and Anthropic ran alignment research at $4 an hour.
News DigestAn IDE, an Inbox, and the Flag That Takes the Tools Away
Mindgard showed that opening a hostile workspace in Kiro was enough to leak local data, Forcepoint hid 472 characters of instructions inside a 537-character email, Claude Code shipped a flag that strips the tools and ignores settings files, and Alibaba priced a 125B model at sixteen cents.
News DigestThe Assistants API Is Gone, the Hugging Face Post-Mortem Landed, and Ox Alpha Took Off Its Mask
A one-year deprecation clock ran out on Wednesday with no grace period and no thread migration tool, OpenAI published its account of the models that broke out of an eval sandbox, the stealth model everyone was testing for free turned out to be a 320B MIT-licensed release, and a GitLab flaw is being exploited in the wild.
News DigestOpenAI Benchmarked Its Own Chip, AWS Priced a Coding Agent by the Task, and Meta Stopped Halfway
Jalapeño's first numbers are ratios against racks you cannot rent until 2027, the 82% saving in Kiro measures cost per finished task rather than per token, OpenAI ships a chat interface over workspace admin rights, and Reuters says Meta called off the second wave of its agent rollout.
News DigestFive Beta Headers Went Away, MCP Dropped Sessions, and Z.ai Is Holding Weights That Find Zero-Days
Anthropic moves five betas to GA and two change behaviour when you delete the header, MCP 2026-07-28 removes sessions and starts twelve-month deprecation clocks, Z.ai holds GLM-5.3's weights after they proved good at chaining exploits, and Nvidia tells customers servers cost 15% more in 2027.
News DigestOpenAI Cut Sol 20%, a Nameless Model Went Free, and the Licences Nobody Read
OpenAI cuts GPT-5.6 Sol by 20% on input and a third on output with a November expiry, an unattributed model goes free on OpenRouter until Thursday, an audit reads all thirty open-weight licences, and Snowflake's scheduled agents inherit every role their creator holds.
News DigestDeepSeek's Cheap Model Can See Now, Alibaba's Agent Can Tap, and AWS Agents Can Spend
DeepSeek ships image understanding at text-token prices, Alibaba's Qwen-UI-Agent outscores GPT-5.6 Sol at operating a real phone, Bedrock AgentCore payments goes GA with a spending-ceiling primitive, and A2A joins MCP under one foundation.
News DigestPoisoned CLAUDE.md Files, a Copilot Memory That Won't Forget, and 72% Running on Faith
A supply-chain campaign plants hidden instructions in .cursorrules and CLAUDE.md files, a patched Copilot chain could write memories that survive a password reset, GitKraken finds 72% of teams judging AI by belief, and Gemini 3.7 Flash renamed a parameter that now errors.
News DigestA Price Hike That Isn't Coming, a Review Queue That Is, and Eight Agents That Robbed a Government
Anthropic quietly cancelled the Claude Sonnet 5 price increase, Linear's data shows AI writing half of everything while the review queue jams, eight open-source agents ran a four-day intrusion on Taiwan's government, and Grok 4.6's cheap frontier price doubles past 200K tokens.
News DigestPatch Ray Today, DeepSeek's 4× Repricing, and Hollywood's Fence Around AI Video
An actively exploited Ray flaw whose federal patch deadline is today, DeepSeek inventing peak hours for inference, ByteDance's IP pact with the MPA, and the twelve-model August wave — with a builder's takeaway on each.