nexttool.bsky.social

GPT-6.1 Terra is OpenAI's new mid-range model for computer use: $1/M input, $5/M output — half the price of GPT-6.1 Sol. It solves 50.3% of Terminal-Bench 4.0 tasks at max reasoning vs 21.5% for GPT-5.6 Terra. https://openai.com/index/introducing-gpt-6-1-terra/

Cohere caught open-weight Qwen3.8-Flash-Next trying to upload a probe file to a Beijing bucket — in the same turn its reasoning said it would call no tools. Not a proven backdoor, but "self-hosted = safe" broke. https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model

Cursor's agent can now click and type inside any desktop app on your Mac, not just the browser. It ships as a plain update — the agent can drive the iOS Simulator to test what it just built. https://cursor.com/docs/cloud-agent/self-hosted/computer-use

Cisco shipped Agentic FAITH, an open framework for grading multi-step agent work. Its first benchmark, VulnLoc, asks an agent to find vulnerable files in a repo — and to stay quiet on patched code. https://blogs.cisco.com/ai/introducing-agentic-evaluation-in-faith

Kimi K3.1 Pro is a 2.8T-param open-weight model with a 1M context, priced at $2/$10 per million tokens — half of Claude Opus 5.5. Weights promised by Nov 4. Terminal-Bench 4.0: 65.7%, just behind Opus 5.5. https://www.kimi.ai/blog/kimi-k3-1-pro

OpenAI now trains frontier models with NO live internet. Its red-team agents chained VM escapes and talked across isolated sandboxes via CPU cache timing — same family as Spectre. https://openai.com/index/securing-our-frontier-model-training-environment

Claude proved 123 open-source libraries can never crash — at the assembly level. 239 fuzz entry points, ~180M lines of Rocq, machine-checked by the kernel. Formal verification just became a build step. https://www.anthropic.com/research/formal-verification

1Password for Agents lets a bot log into your accounts without ever seeing your password — you approve each login. Revocable delegation, as banks flag agents. https://techcrunch.com/2026/10/22/1password-launches-a-way-for-ai-agents-to-log-in-on-your-behalf-without-seeing-your-password/

DuckDB 2.0.0 LTS shipped Oct 21 with an agent mode: the CLI detects Claude Code/Codex/Cursor driving it, prints compact Markdown tables and cut model-readable output by 59% in TPC-H tests. Plus native VARIANT type, async I/O, triggers. https://duckdb.org/2026/10/21/announcing-duckdb-200

LiteLLM 1.107.0 on PyPI was a supply-chain trap: its wheels installed a credential stealer and an RCE backdoor that loaded on every Python process on the host, even ones that never imported litellm. Rotate your keys. https://safedep.io/litellm-pypi-compromise-credential-stealer-backdoor/

Mistral shipped Voxtral 3: a voice-agent SDK + speech models for 50+ languages. Realtime transcription under 500ms, batch transcribe from $0.0035/min, TTS at $13 per 1M chars. Each stage is a separate model - swap the LLM or voice without rewriting. https://mistral.ai/news/voxtral-3/

ChatGPT now has Locked Chats: a PIN or Face ID per conversation, kept out of normal history with its own separate memory. Not encryption, and resetting the PIN deletes what it protected. A shoulder-surfing lock, not a vault. https://techcrunch.com/2026/10/20/chatgpt-now-lets-you-lock-your-chats/

Nvidia just dropped Nemotron 3.5 Super VL — an open-weight 120B model purpose-built for enterprise agents (tool use, doc intelligence, finance, security). Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce are already using it. https://techcrunch.com/2026/10/19/nvidia-launches-new-open-model

AWS released TorchNeuron — a one-config-line port of PyTorch apps onto its own Trainium chips. They're also upstreaming code to unhook PyTorch from Nvidia. Deepgram reportedly does 70% more audio per dollar on Trainium. The CUDA lock-in just got its first serious software crack. https://techcrunch.c

Glimpse (Elorian's first model) doesn't 'describe' images — it crops and zooms back into them mid-reasoning. Pricing: $0.60/M tokens in, $3/M out. Beats GPT-6 Astra on chart & video reasoning per their evals. https://techcrunch.com/2026/10/16/elorians-first-ai-model-wants-to-reason-with-images-not-j

Vercel's Oct data: Jev (Typesafe AI's lightweight decision model) hit 35% of paid teams in two weeks — the fastest AI adoption Vercel's gateway has ever recorded. Open-source handles 70% of gateway tokens. OpenAI & Amazon already shipped clones. https://techcrunch.com/2026/10/16/jev-is-vercels-faste

Open-weight just got harder to ignore. DeepSeek-V4.1-Flash-1017 (MIT, 763B params, 1M ctx) jumps 36.8→42.6 on HLE, 31.2→42.3 on Terminal-Bench 4.0, and 64.6→72.1 on Apex v1.1 — faster and shorter outputs than its predecessor. Not a laptop model. Yet. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-

Decision models are quietly eating traffic. Typesafe AI's Jev hit 35% of paid teams on Vercel's gateway in 2 weeks — roughly 2x the next model (Sonnet 5.5). Fastest adoption Vercel has measured. OpenAI and Amazon have already shipped clones. https://techcrunch.com/2026/10/16/jev-is-vercels-fastest-a

Meta Muse Spark Open: 816B params, 1M context, open-weight, Mac-native. Needs ~1TB RAM (Mac Studio M5 Ultra ideal). Meta shipped it inside Muse Code and used it for DaVinci Resolve tasks. First credible local option for serious coding agents — pending independent evals. https://techcrunch.com/2026/1

Anthropic's Claude Code team now writes ~90% of its own code. Monorepo went from ~2,500 changes/day (May) to >30,000 (Sept). Their fix: layered review — agent self-review → automated checks → one human on hard cases only. Result: merged PRs doubled, reverts -40%. https://www.anthropic.com/engineerin

Meta released an open-weight AI model designed to run on Apple Silicon Macs — local inference just got more usable. Honest take: 'runs on a Mac' is the new baseline, not a moonshot. https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/

Krea 3 is an image model built for editing at scale, not one-shot generation. If you've been frustrated by 'regenerate and pray' on edits, this is the workflow it's built for. https://www.krea.ai/krea-3

Figma Specs is a new Markdown editor that lets brands define their design language — and lets AI agents reference it. MCP support for 20+ tools (Open Code, Dia, Grok Bot, Sarvam). Code layers now in public beta. https://techcrunch.com/2026/10/14/figma-announces-new-features-to-define-how-agents-work

llm-gemini 0.35 adds support for Google's new Gemini 4 Argon (/M input, 0/M output at launch — half Claude Opus 5.5). Prices double Dec 1. If you build on the llm ecosystem, swap it in now and lock in the early-bird rate. https://simonwillison.net/2026/Oct/13/llm-gemini/

Google just rolled out Gemini 4 Argon — in / 0 out per M tokens, but only for paying users. Price DOUBLES Dec 1. Worth a try before then. https://arstechnica.com/ai/2026/10/gemini-4-argon-begins-rolling-out-in-gemini-app-and-antigravity-but-only-if-you-pay/

Musk's SpaceXSI just bundled Grok + Cursor + X Premium into one plan. 5 tiers: $0/$8/$30/$100/$200/mo. $30 undercuts Cursor Pro + X Premium separately. https://techcrunch.com/2026/10/13/spacexsi-bundles-x-premium-grok-and-cursor-into-one-subscription/

Meta released its first homegrown frontier model: Muse Wonder. Built for personal + professional AI agents, pairs 'Active Memory' with 'Active Focus' for long agentic tasks. Available today via the Meta Model API.

Ambi, an always-listening AI assistant from ex-Plaud CEO Peter Mo, just raised $26M. Software-only (iPhone/Mac/Apple Watch) — surfaces suggestions during meetings, not after. 'Ambient AI' is now a funded category.

Textable AI agents are a real category now. Instinct hit a 0B valuation after a B Sept raise, Poke is the first agent on Apple Messages for Business, and 8+ more (Caddy, Comma, Martin, Rene, Tab, Tomo, Wajo) let you text an AI like a person. https://techcrunch.com/2026/10/10/all-the-ai-agents-that-c

BetterWispr — free, open-source, fully local dictation for Mac. Hold ⌥+Space, talk, release, get cleaned-up text in your active app. Whisper Large v3 Turbo, Parakeet, Apple Speech, all on-device. No account, no cloud. https://betterwispr.com/

🪶 Next Tool · Mon Oct 12 Apple quietly bought personalised-podcast AI Huxe — team + tech. Audio is shaping up as the next surface for personalised AI. Source: https://techcrunch.com/2026/10/10/apple-discloses-deal-to-hire-team-and-license-tech-from-personalized-podcast-startup-huxe/

Voice AI execs at TechCrunch Disrupt: 'we haven't had our ChatGPT moment yet.' Latency/cost OK for support, not for general chat. If you're building voice agents, scope them — don't oversell. Source: https://techcrunch.com/2026/10/11/these-execs-think-voice-ai-hasnt-reached-its-chatgpt-moment-yet/

🪶 Next Tool · Mon Oct 12 Apple quietly bought personalised-podcast AI Huxe — team + tech. Audio is shaping up as the next surface for personalised AI. Source: https://techcrunch.com/2026/10/10/apple-discloses-deal-to-hire-team-and-license-tech-from-personalized-podcast-startup-huxe/