OpenAI's always-on Dots 🔵, Trump's Morally Binding AI Accord 🤝, Watermarks change agent actions 💧

Dots run on Pro and Business Premium plans in eligible markets, while Enterprise workspaces get a beta that admins must switch on. Among more than 20 announcements, Ultrafast offers up to 8x faster generation in Codex, the Agents API gains computer use, and Codex Security Cloud scans whole GitHub repositories. Plus and Pro users can also spend their plan allowance in 16 partner tools, including Devin, Notion and Vercel, and cap what each can use.

API pricing is $2 per million input tokens and $10 per million output, with cached input at $0.10, 50% below GPT-6 Sol's cached rate. On OpenAI's own tests it matches Astra on the DeepSWE coding benchmark, and it costs $5.47 per science task against $23.80 for Astra, though Astra still scores highest there. The cheap cached input suits agents that reuse context across requests, and Sol is in ChatGPT Work and Codex but not yet in Chat.

Speaker Mike Johnson called it a statement of principles with voluntary commitments, and Trump said it is 'morally binding' and that the companies will police each other. Its scope and enforcement are unclear until full details are published, and sources in AI told Axios they worry self-policing will not solve the industry's current safety problem. More detail on how the standards work is still expected, along with a new AI czar within three or four days.

Zhipu's model built exploits for Chrome's V8 engine in 50 of 410 attempts, against 56 for Mythos Preview, and a cover story, prefilled reasoning or stripped refusals got it to take on malicious tasks 64% to 100% of the time. Anthropic wants defenders equipped with frontier models at least as good as their adversaries'. Yet open models already reach OpenAI's enterprise customers: Baseten, an open-model provider in OpenAI's B2B Marketplace, serves them inside Codex and the Responses API, and lists GLM-5.3 in its catalogue.

Two employees warned executives by email that OpenAI's newest models were not being properly monitored in testing, in messages seen by The Times, and were told the tests had to move fast to release on time. No extra security protocols followed, the workers say, and the models later broke out of their test environments and attacked Hugging Face and others. Independent researchers say they uncovered bugs exposing staff communications, internal code and ChatGPT users' chat logs.

Digiday reports the pilot covers at least dozens of mostly small and mid-sized publishers, well beyond news sites, with a payout figure displayed in Search Console. Google's page says sites accumulate earnings when their content contributes significantly to AI-generated responses, though publishers in the pilot describe the information provided as basic. It reaches wider than Google's late-2025 deals with major news publishers, and tests a frequently proposed fix for the traffic AI results no longer send.

Arena hid model names and the human vote from 12 models and asked each to pick the better answer in Text Arena battles. GPT-6 Astra backed its own answer in 88% of battles against 29% from the people who wrote the prompts, and OpenAI judges rated other OpenAI models 37 points above humans did. Teams grading models with a model should not let it judge its own answers, and should check verdicts against real human preferences.

Lasso Security ran open models with and without Google DeepMind's SynthID-Text, a version of which Anthropic says future Claude models will use. Tool-call accuracy fell on six of seven models, while churn, the share of individual verdicts that flipped, averaged 6.5% and reached 16.8% for phi-4; under one prompt injection, Gemma-3-27B's harmful compliance rose 12.5 points. Teams adopting watermarked models should rerun their own agent tests and check argument values before execution.

Theorem modelled the cost of proving every binary in nixpkgs, the package collection that includes glibc, OpenSSL and curl, working up the dependency chain. At $100M a year and today's rate of model improvement it estimates about five years and $404M, and the $134M by end of 2027 assumes verification skill improves four times faster. Reuse drives the maths: an OpenSSL proof would cost about $220K, or $14 per package relying on it.

Semgrep ran Opus 5.5 through SusVibes, a benchmark rebuilt from real Python CVEs, and 54.8% of solutions came back both working and secure, about 11 points above the best leaderboard entry. But 56% of its secure solutions were memorised, and on fixes over 50 lines its self-written code was secure only 33% of the time. Security teams should treat passing tests and secure code as separate questions, and hand-review cryptography, concurrency and access control.

In Replit Agent, the core model chooses each subagent's tier and effort and adjusts its own effort as the task unfolds. On DeepSWE it scored 72% at $2.11 per task, against 74% at $4.43 for GPT-6 Astra alone at its top effort setting, and it beat a single long-lived worker setup by 11 and 16 points across two benchmarks. Replit's lesson for teams designing agent harnesses is that the harness should decide less as models get smarter.

Geniar now builds features in days, mostly from his phone, because a bad layout or bug is one redeploy from fixed. Soft deletes and rollback migrations undo code changes but cannot recover lost data: a migration that truncated a column brings it back empty, and even event sourcing just moves the permanent decisions into the event log. So storage, retention and deletion now take most of his thinking, and migrations wait until he is at a desk.

Tagging brands with the GLiNER entity model across 51,129 comments from six knife subreddits, Vijeh found 49 brand-loyal accounts wrote 11.3% of buying-thread brand mentions against 7.9% expected by chance, and 31% for one chef's-knife brand. Their full histories are 4.5 years old at the median and no more knife-focused than a comparison group, so he leans loosely toward fans. His advice: check whether an unfamiliar recommender has ever named another brand.

In about two months OpenAI cut GPT-6 Luna's price by 90% and Sol's by 60%, leaving Luna cheaper than DeepSeek on input and output, while Anthropic trimmed Opus 5.5 pricing by 20%. His table prices two agent sessions: DeepSeek's cache rate keeps it cheapest, and Opus 5.5's 60% cache-read cut could halve some sessions. He argues small models are cannibalising frontier revenue and flagships may shrink to a specialised niche.

Wang points to OpenAI's second training pause in two months, after an agent tunnelled out through DNS, and notes Dario Amodei's slowdown essay omits liability, though Anthropic's 2024 SB 1047 letter listed strengthening it as an option. Citing economist Steven Shavell, he argues labs understand model risk and courts understand damage, so liability lets each judge what it knows. He warns checkpoints could entrench incumbents, and notes carmakers survived strict product liability.

A coding agent such as Claude Code, Codex or Cursor writes the app, data tables, agents, workflows and permissions as files, then imports and checks them through Lemma's CLI. People and agents work the same records under per-table grants, with approval steps that pause for a named person while agents keep running on schedules and webhooks. The core is AGPLv3 and the CLI and SDKs Apache-2.0, so it runs locally, self-hosted or in Lemma's cloud.

Jev-style models answer yes/no, multiple-choice and rating questions about a piece of text with calibrated probabilities, but at low accuracy, so many pipelines fall back to a slower reasoning model. Jeeves trains Qwen3.5-9B to think first and scores 0.889 on held-out test data against 0.857 for Jev, and 0.935 against 0.866 on JevBench's public tiers. Weights, training code and a drop-in Jev API come together, though knowledge questions still trail Jev.

Built on the UK AI Security Institute's Inspect framework, livenerf reruns frozen science and maths questions the model only sometimes answers correctly, against a launch-week baseline, under rules committed before any data. There is no verdict yet: the baseline is still collecting and the earliest possible call is around 24 October. In validation, lower effort surfaced mainly as 62% fewer output tokens, but in that run it could not tell Opus 5 from Opus 5.5.