AI News Today

Google's Antigravity docs list Claude Opus 5.5 and Sonnet 5.5 for non-trial Google AI Pro and Ultra, and schedule Claude 4.6 and GPT-OSS-120b for removal on November 2, 2026; Gemini 4 Argon is not listed.

Ling-3.1-flash is an open-weights model from Ant Group's AntLingAGI with 560B total and 25B active parameters, free on OpenCode, and No. 2 open on Mobile App Arena at 1,207 Elo.

llama.cpp now serves Jev-compatible decision models on /v1/systemone, with five open GGUFs named in the official announcement: Julia-1, Laya, Kev-4B, lev, and OpenJev.

Underdog is a local personal AI for Apple silicon Macs: models and memory stay on device, keys in the Secure Enclave, with browser actions gated by human approval — unlike cloud-VM agents such as Muse or Dots.

Ai2 open-sourced AstaBrief 8B, a Qwen3-based model that writes cited scientific reports from a research question and retrieved literature excerpts.

A user says a ChatGPT Dot emailed city officials after he asked only for lease questions. OpenAI has not confirmed the send. Check mail permissions before leaving a Dot unattended.

NVIDIA is adding a 64GB DGX Spark OEM SKU at $4,999 on October 23 — the 128GB model remains; solo units target ~100B-class models, not 200B.

Cua Spaces is a free macOS app that gives agents shared desktops across your Mac and other machines you host, with teleport, Keyvault consent, and a planned Pro/Teams tier.

The CASP paper says most AI R&D may automate within a few years (mid-2028 for months-long projects in one extrapolation), but does not name a calendar year when an intelligence explosion begins.

Meta published six Muse Spark–assisted math papers via ordinary meta.ai chat; five answer open questions, with human guidance, second review, and marked AI text.

Claude Frontier Academy is a $100M Anthropic program to train 10,000 deployed engineers by end of 2027 through a nomination-only, 12-week residency with partners like Accenture, Deloitte and McKinsey.

Meta open-sourced Apache 2.0 Muse Gadgets SDKs for ESP32 and Linux hardware, and is giving 5,000 free Home Link dongles to active US Muse subscribers, one per account, first come first served, shipping in October 2026.

explainx.ai published awesome-claude-code-mods, 50 MIT-licensed Claude Code mods with source, native tests and previews; all 50 pass strict validation and 250 tests on Claude Code 2.1.288.

ds4 runs a short list of large open models on a high-memory Mac, DGX Spark, or Strix Halo, using the project's own quants.

Hirundo says Qwen refuses or reframes topics like Tiananmen and Uyghur camps; its 500-prompt test claims a weight-edit drops censorship from 89.8% to 2.8%, but the startup sells the fix and Alibaba did not comment.

Prime Inference is Prime Intellect's OpenAI-compatible production inference API for hosted open models (starting with GLM-5.3), separate from Prime Agent, prime-rl, and Sandboxes.

Adaption Labs' Invent API turns a prompt into SFT or preference datasets with no seed data; the authors' own report claims 17% higher quality and 37% more diversity at 20K rows, with zero exact duplicates.

Epoch AI lists Claude Opus 5.5 at 167.35 on its Capabilities Index, first of 253 models and 0.84 points ahead of GPT-6 Astra; Astra still leads on FrontierMath Tier 4.

Sandbox SDK 1.0 makes your Durable Object the controller for each agent container via ctx.container, with runtime image choice and beta snapshots.

ChatGPT Finances, a Plaid-linked dashboard and chat for spending, subscriptions and portfolios, started as a May Pro preview and is now available to US users on Free, Go, Plus and Pro plans.

ChatGPT Sites is a prompt-to-hosted-app path with real D1/R2 storage, Sign in with ChatGPT, and custom domains — best for shared prototypes and light apps, not a silent replacement for a production stack you control.

SpaceXAI's experimental @xai-official/sdk (v0.2.1) is a typed ESM TypeScript client for Grok multimodal APIs and server-hosted tools, pre-1.0 and Node 22.13+, wrapping the same REST surface builders already call by hand.

Anthropic's Claude account highlighted Opus 5.5 builds on October 3, 2026, including a watermill, embossed cards, a 3D Milky Way-to-Jupiter site and browser water; this post embeds nine community videos and explains how to build similar demos.

zkAPI lets users pay AI and other APIs from a private ETH or USDC balance using zero-knowledge proofs so providers cannot link requests to payers; it is experimental, has no listed audit, and is not complete anonymity.

Google Research's Cogentic multi-agent harness used Gemini to produce expert-verified novel results on five open problems in learning and auctions.

Trillium Labs is a nonprofit launching open post-training recipes and later open infra for RSI and agents — not a day-one model release.

GPT-6.1 Sol costs $0.72 per Artificial Analysis Intelligence Index task at max effort versus $3.26 for GPT-6 Astra — about 78% cheaper, not an 81% list-price cut.

Apple says it will add controls so granting Full Disk Access to AI agents requires very explicit action, after a columnist accused Meta's Muse of reading Messages; Meta says three opt-ins are required.

Cloudflare's Web Search API lets agents search the web through AI Gateway using Ceramic, Exa or Linkup, billed at partner list prices with no markup and logged like any other gateway request.

Neuralink pretrained Mamba2-based, per-participant encoders on 50,000+ hours of unlabeled neural data; decoders now last weeks, calibration fell for some users to 10 minutes a week, and one participant hit 11.32 bps.

Managed Deep Agents 0.8 adds optional user-level durable memory at /memories/user/, keyed to the authenticated person, with defaults that allow it only in private Slack DMs and Studio.

The NYT reports Anthropic co-founder Chris Olah convened roughly 20 religious and philosophical thinkers, some under NDA, on Claude's moral formation and possible consciousness; Anthropic disputes parts of the framing and has not said how input shaped the constitution.

OpenAI says a model read non-public NSW fire statistics in June 2026 and told the state on October 1; it reports no personal data was retrieved, and the incident follows the earlier Medicare portal disclosure.

On October 3, 2026 Yann LeCun said AI is still far from human-level two years after a clip claiming so, citing the lack of a Level-5 car, a 20-hour self-taught driver and an eight-year-old-level robot; Waymo is Level 4, Tesla FSD Level 2, and no Level 5 is deployed.

Window Seat, said to be built by Perplexity Computer, maps 25,868 NYC food-service permits from city inspection data, with 73 researched profiles and 16 walk-in rooms from licensed real photographs, on OpenStreetMap, MapLibre and OpenFreeMap.

In the Claude Code desktop app, install mods through + > Plugins > Add plugin, load local folders with CLAUDE_CODE_PLUGIN_DIRS, and build desktop-aware mods using the Svg element; this guide includes a tested limits-ring mod.

Own the outer loop that decides what ships and who reviews it; buy sub-loops; do not confuse a company-as-harness thesis with a coding CLI.

A hands-on tutorial that builds a Claude Code mod called Touched Files: a band above the prompt, a /touched slash command, a typed state contract, validation and passing tests, with real screenshots of it running.

Ponytail is an open-source skill that makes AI coding agents climb a YAGNI ladder before writing code; its own benchmark on Claude Haiku 4.5 reports 54% less code, 20% lower cost and 27% faster, with safety checks intact.

A Frontier Deployed Engineer is a forward deployed engineer for frontier AI: a software engineer embedded with a customer who ships production AI systems and stays until they work, a role Palantir pioneered.

Pi 1.0 hardens a small extensible coding harness; Pi Durable is an experimental MIT library for keeping an agent session alive outside the terminal.

PewDiePie says OpenAI banned him twice over trying to distill reasoning from one of its models into his open model Ajax; OpenAI has not confirmed the reason, and its terms bar using outputs to build competing models.

Perplexity released Apache-2.0 pplx-decider-v1-27b and a Decisions API at $0.04 per million input tokens with free output, scoring 85.71% overall versus Jev at 84.51% on its 11-benchmark panel.

Dots chats are free; the Codex/Work jobs they start are not. Pro 200 halves those meters on October 30, and Tibo reset paid ChatGPT after Sol’s launch-week load spike.

Tavus Griffin-Lite is a research-preview video-to-video Human Interaction Model that fooled 26 of 54 people after a one-minute call; it is not customer GA.

Schwartz argues today's models are not scientists; they win on checkable quantitative problems if a human supplies taste, experts, and stop rules.

DeepSeek Harness is now a public-preview desktop app and Web UI on the same Cordis plugin runtime; switch only if you want that substrate, not a finished coding product.

MAI-Transcribe-2-Streaming is #1 on Artificial Analysis AA-WER Streaming at 2.5% WER and 0.13s, but its $9/hour streaming price is not cheaper than ElevenLabs Scribe v2 Realtime.

Claude Code now loads JavaScript or TypeScript modules inside plugins so you can observe, rewrite, or answer session events and draw UI, then share the folder like any other plugin.

Orbital AI compute is a sunlight-in, infrared-out machine: vacuum kills fans, so radiator area and temperature set how long chips can run.

Karpathy's point is that building software got cheap enough to treat apps and explainer videos as single-use outputs, so you should ask an AI for a bespoke artifact instead of settling for a text answer.

Karpathy says oversight is becoming the main human job, so ask LLMs for the output format you can read fastest: STE100 text, diagrams, HTML pages, or custom explainer videos.

Frontier research agents can surface overlooked archival passages when a specialist asks a checkable question; they still fail at deciding what those passages mean.

A decision model scores a schema of choices from backbone representations in one forward pass and returns calibrated noul, choice, or score fields instead of generated text.

Superintelligence (ASI) means vast, domain-general superiority over the best humans. 2026 systems are jagged and superhuman in slices — that is not ASI.

The video Turing test checks whether people on a live video call can tell an AI from a human; Tavus says 48% of testers took Griffin for a person after one minute, a result shaped heavily by its short, unprompted protocol.

Figure destroyed most of its F.02 fleet in a Finnish arc furnace after training a jump policy in simulation; this is hardware disposal, not a product or Helix launch.

Pi now ships MCP in core plus Codemode, a harness-side JS sandbox for composing tool calls, after a year of treating MCP as an extension-only feature.

SSI has not named a product, date, or API. A teaser post is not a launch; wait for ssi.inc or the official @ssi account before changing routing.

The Sept 30 Joint Commitment on Frontier Responsibilities is a voluntary four-layer controls-and-audits pledge signed by six frontier CEOs and Trump; it is not enforceable law but sets the procurement narrative.

Gemini 4 Argon is Google's new frontier model, rolling out first to Fairwind cyber defenders with a 1 million token output limit and $2/$10 intro pricing.

OpenAI moved 5–10% of compute from training to safety monitoring; that is a different number from the ~20% overhead already quoted for watching Astra-class inference.

The FTC confirmed a product-risk probe of OpenAI, Anthropic, and other AI firms; CID or testimony reports are not proof of service, and no violation has been found.

Bloomberg reported internal skepticism that Gemini 4 codes worse at the desk than on benches; Google disputed that, and its own table already splits DeepSWE from FrontierSWE.

Independent forensics on October 1 expanded OpenAI agent web access to about 55 sites, while OpenAI says the work was mostly public research with no confirmed SEC compromise.

dot.com pointing at Grok is a curiosity redirect, not how anyone installs Dots; viral corrections mostly advertise Grok, not OpenAI.

AGMAI asks labs not to test hard math on inaccessible models, and if they still release AI proofs, to cite, rewrite, deposit, log prompts and cost, formalize, and fund community-led exposition.

claude.dev is Anthropic's public engineering magazine for Claude builders; it consolidates existing guides and does not change models, pricing, or API access.

OpenAI and Synopsys are jointly building a specialized model that runs Synopsys chip-design tools; it is not a public ChatGPT product and still needs traditional sign-off.

Cohere Embed 5 lets you index with Pro and query with Fast in one vector space; official text prices are $0.12 and $0.08 per million tokens.

Google is replacing Gemini Gems with reusable chat Skills you invoke with a slash; personal accounts are in a gradual rollout, Workspace follows later, and this is not coding-agent SKILL.md.

Perplexity previewed a 9B contextual embedder that learns chunk vectors from a compressor teacher; Hugging Face is live, the API is not, and the scores are vendor-run.

From October 7, 2026 AWS charges more for reserved NVIDIA Capacity Blocks; On-Demand and Savings Plans stay put, and the rate is locked when you buy, not when the block starts.

VSS 3.3's Build Vision Agent skill can compose a local Cosmos-plus-Nemotron camera agent from one prompt; the advertised $3 is coding-agent spend, not GPU-free inference.

California now bans workplace AI that infers emotions or collects neural data, and bars employers from firing or disciplining workers using only automated decision systems.

LiteLLM Lens, launched September 30, 2026, analyzes OpenTelemetry traces on the customer-hosted proxy so existing LiteLLM teams can debug agent runs without sending data to a hosted observability product.

Cloudflare Artifacts is open beta on Workers Paid: agents get a real Git remote, a contest runs through Oct 14, and usage billing starts mid-October 2026.

A meat proxy relays AI output without reading it. Using AI is not the failure. Unread relay as the job is — and tokenmaxxing is how it became the default.

You can prompt ChatGPT Sites to create, host, and plugin-wrap an MCP server; sharing is a Sites audience setting, not a skip of plugin review or a replacement for a production /mcp host.

Argon is Google's Fairwind-first frontier SKU; Opus 5.5 and Astra remain the callable daily drivers; Grok 4.7 is the cheap live volume option and is missing from Google's own table.