AI News: October 11, 2026

1. a16z Data: Only 4.5 Percent of US Consumers Pay for an AI Chatbot, but the Top 1 Percent Spend $900 a Month

Andreessen Horowitz. The seventh edition of a16z’s Top 100 Gen AI Consumer Apps list added observed US card spending (YipitData panels) for the first time, finding that nearly half of US consumers use AI but only 4.5 percent held a paid personal ChatGPT, Gemini, or Claude subscription in August. The top 1 percent of payers average about $900 a month, up 79 percent since early 2025, and account for nearly a fifth of observed spending, mostly on development, automation, and creative tools, while the typical payer spends about $25. In a TechCrunch interview, a16z partner Olivia Moore said social, dating, retail, travel, finance, and health still have no entrants in the top 100. Source

2. Sakana AI’s Multi-Agent Reviewer Caught 73 Percent of Planted Core-Claim Errors

Sakana AI. Sakana published “Beyond Imitation,” a TMLR paper that grades AI peer reviewers on whether they find planted mistakes rather than how closely they mimic human reviews, using a Contradiction Benchmark of 1,164 inserted contradictions across 257 papers. Its Multi-Layered Review system (appendix, literature, and three-pass review agents on off-the-shelf Claude models, about $0.47 per review) caught 73.43 percent of core-claim errors with four reviews versus 14.81 percent for the best baseline. On real retracted arXiv papers the gains shrank to 16.11 percent exact matches, and the system remained vulnerable to hidden prompt injection. Source

3. Nace AI Open-Sourced Drex 1.5, a 9B Decision Model That Ties Jev on Decision Index

Nace AI. Drex 1.5 is an 8.95B-parameter decision model that returns a probability for each supplied option in one forward pass instead of generating text, built on a distilled Qwen 3.5 9B backbone with a separate pointer head. Nace reports 58.08 on the public Decision Index 0.3.1, inside the tie band with TypeSafe’s closed Jev 1.13.0 (57.96), with 93.4 percent accuracy on 32K to 128K token documents but weak knowledge and fine-grained sentiment scores. Weights are on Hugging Face, a Q8_0 GGUF runs on Apple silicon and CPU, a hosted version is on OpenRouter, and the API mirrors Jev’s request format so existing clients can switch with a few environment variables. Source

4. Alibaba Released Qwen-Image-2.1-Turbo for 8-Step 2K Generation and Editing

Alibaba Qwen. Qwen-Image-2.1-Turbo is an accelerated checkpoint of the open-weight Qwen-Image-2.1 that generates and edits images in 8 denoising steps instead of 40, keeping the same 7B single-stream DiT with a Qwen3-VL 8B text encoder and a 64-channel RGBA VAE for native transparency. Block-causal attention lets prefix KV caching compute text and reference-image context once and reuse it across steps. It loads in Diffusers (from source), is also hosted on Alibaba Cloud Model Studio, and ships under a research-only license, so commercial self-hosting requires separate permission. Source

5. Saluki 27B Squeezed Qwen3.8-27B Into 7.89 GB and Beat It at Parallel Tool Calls

Conway Research. The team behind the on-device assistant Underdog released Saluki 27B under Apache 2.0, a 2-bit mixed-precision GGUF of Qwen3.8-27B that is 7.89 GB versus 54 GB in BF16 and runs in stock llama.cpp. Building on ISTA-DASLab’s GSQ-RCO quantization, it was tuned to protect tool calling: it scored 42 versus 35 for the full model on 100 BFCL v4 parallel tasks and 88 versus 84 on a 120-task BFCL-based bench, retaining 96 percent on average across nine benchmarks. Competition math dropped the most (AIME 2025 fell from 96.7 to 79.2), and the final compression recipe has not been published. Source

6. Cockroach Labs Ran a “Teaching Hospital” of Coding Agents for Five Months

Cockroach Labs. Cockroach Labs detailed MOLT Sinai, a pipeline that treats GitHub issues as patients and assigns agents roles such as Fellows that implement fixes, Review Attendings that hunt for faults, and Discharge Nurses that audit reviews before merge, with humans making final calls. From April to September it merged about 1,238 PRs and over a million lines with 7 reverts, for just over $135,000 in tokens (about $84 per issue), and added Db2 support in under two days for $4,172 versus roughly nine months for Oracle support in 2024. The team reported that the process sometimes became bureaucratic, required a three-round rework circuit breaker, and still needs human verification of the agent-written Db2 code. Source

7. A Developer Used Over 500 Billion Tokens of Agent Work to Decompile a Commercial Shooter

momo5502. A developer described running fleets of Claude Code and Codex CLI agents (up to 14 Luna and 2 Opus 5.5 agents at once) with Hex-Rays’ ida-mcp to perform byte-matching decompilation of an unnamed first-person shooter, coordinated through GitHub issues and a shared Discord channel. About 99 percent of functions were reconstructed and 83 percent are byte-exact, verified by a comparison script that CI hashes against a secret so agents cannot alter it. Lessons included that reviewer agents accepted workers’ excuses for deviations, that correctness must be machine-checkable, and that discarding bad output was faster than salvaging it. Source

8. Prime Intellect Rewrote Its Prime Agent Coding Harness in Rust

Prime Intellect. Prime Intellect rewrote Prime Agent, its open-source coding agent harness launched in August with more than 300,000 downloads and 8 trillion processed tokens, from TypeScript to Rust across nine crates with per-session worker isolation and Windows support. It reports cold-start time to input of 55.8 ms versus 737.8 ms and post-startup memory of 106 MB versus 607 MB. The company said the underlying sandbox and inference infrastructure is intended to power agent swarms and reinforcement learning on its platform. Source

9. Probes Beat an Opus 5.5 Text Monitor at Catching Agent Sabotage

Research. A paper by Oskar Hollinsworth, Adam Gleave, Chris Cundy, and coauthors trains white-box deception probes on what it calls the largest deception dataset to date, using an architecture that aggregates signals across many layers and tokens. The probes reached 98.8 percent AUC on SHADE-Arena, surpassing an Opus 5.5 text-monitoring baseline, and up to 99.7 percent AUC at identifying a model’s true hidden goal in “introspective deception” cases that cannot be judged from the transcript alone. The FIBS training dataset is released for teams building activation-based agent monitors. Source

10. Statistical Reanalysis Finds METR Time-Horizon Jumps Are Not Equally Hard

Research. Drew Nguyen and William Fithian refit METR’s 50 percent time horizon on 228 tasks and 26 AI systems using splines and item-response theory, dropping the assumption that AI difficulty scales linearly with the log of human completion time. The fitted curve is nearly flat between 2 and 30 minutes, so going from 3 to 30 minutes is much easier than going from 30 minutes to 5 hours despite the same 10x multiplier. The authors provide better-calibrated point estimates and diagnostic plots, and recommend reading time horizons alongside them as longer-task benchmarks appear. Source

11. “Structure Tax” Paper: JSON Output Hurts Accuracy Only When Answers Come Before Reasoning

Research. A study across multiple models, datasets, and schemas found that the accuracy loss from constrained JSON or XML output depends on schema design rather than structure itself. Reasoning-first field ordering matched or exceeded free-form accuracy, while answer-first ordering caused steep drops, particularly in smaller models, and reasoning-preserving schemas also improved calibration. For practitioners using structured outputs, field order in the schema is a direct lever on task accuracy. Source

12. Split-Key Scheme Proposes Making LLM Watermark Detectors Public

Research. Pierre Fernandez, Tom Sander, and coauthors argue that LLM watermark detectors could be released publicly, measuring how much a public detector helps attackers at access levels from token scores to a binary verdict. They propose a split-key method that exposes one key through a public detector while keeping a second key private, plus a statistical test that flags the resulting imbalance when an attacker tampers with only the public signal. Public access helped removal only at small edit budgets but enabled forgery, which the private pipeline could still detect. Source

13. Meta-Analysis: LLMs Have Not Broken the 75 Percent Ceiling in Automated Deception Detection

Research. A systematic review of 25 years of automated verbal deception detection (289 reports, 6,136 models) meta-analyzed 3,653 models across 97 datasets and found pooled accuracy of 74.4 percent. Accuracy tracked methodological quality, such as ground truth and evaluation on independent data, rather than model complexity, and embeddings and LLMs did not improve results. Only 12.46 percent of reports used verifiable ground truth, and the authors conclude that a 70 to 75 percent ceiling is unlikely to move under current research practices. Source

14. Tutorial Shows How to Build a Jev-Style Decision Model From Qwen3-1.7B

Community. A walkthrough by Nish Tahir builds a “System one” decision model by restricting Qwen3-1.7B’s output logits to option letters and taking the top option in a single forward pass, evaluated on CommonsenseQA. Zero-shot accuracy was 59.4 percent and rose to 62.4 percent after a quick finetune, and temperature scaling (T of about 3.8) corrected the raw model’s overconfidence. Code for dataset building, evaluation, finetuning, and calibration is on GitHub as build-your-own-jev. Source

15. REA Gives Coding Agents Reverse-Engineering Tools for Binaries and Electron Apps

Open source. REA (Reverse Engineer Anything) is an MIT-licensed toolkit, installed with npx rea-agents@latest setup, that lets coding agents inspect functions, strings, references, and call graphs in native binaries, map modules, routes, and IPC in JavaScript and Electron apps, and capture browser or process activity for comparison across runs. It reached the Hacker News front page with more than 700 points. Source

16. Talorys Runs a Self-Hosted Personal Agent on Cloudflare’s Free Tier

Open source. Talorys is an MIT-licensed, single-user AI assistant that deploys to a user’s own Cloudflare account using Pages, a private Worker, a Durable Object built on the Cloudflare Agents SDK with SQLite storage, and Workers AI running GLM-4.7-Flash. It handles chat with persistent memory, tasks, notes, projects, and recurring reminders via Durable Object alarms, avoids paid services such as D1, R2, and KV, and sends no telemetry. Source

17. Danu Robotics Is Pitching a Pincer-Claw Recycling Robot

Danu Robotics. Edinburgh-based Danu Robotics builds H.E.R.O., an AI-driven recycling sorter that grips items with a pincer claw rather than the suction arms used by competitors such as Glacier and Recycleye. The company has raised $5 million in late-seed funding, holds $500,000 in signed contracts with more than 200 prospects in its pipeline, and plans a portable standalone unit for venues such as airports and hospitals. Source

18. Report: Google Is Testing a Gemini 4 “Carbon” Checkpoint Said to Rival Opus 5.5 at Coding

Google. Business Insider, citing internal documents, screenshots, and chats, reported that Google employees are testing Gemini 4 variants code-named Argon, Barium, and Carbon, with Carbon deployed on the internal coding platform Jetski in recent days. One employee described Carbon as the “Gemini pro next model” and compared its coding performance to Anthropic’s Opus 5.5, though no benchmarks were shared and Google has not confirmed Carbon or a release date. Carbon and Barium appear to be checkpoints within the Argon family rather than new model tiers. Source