AI Daily · 2026-10-05

OpenAI moved on both compliance and monetization: it introduced textGrain, an invisible text watermark for EU AI Act provenance rules, with API opt-in, EU rollouts, and researcher-only detector access, though short, constrained, or synonym-substituted text weakens detection. It also began testing visual ads in ChatGPT, explicitly labeled and kept separate from answers, while expanding attribution and incrementality partners. Elsewhere, Together Link lets teams route existing coding agents to open models with a single command and claims over 50% cost savings, while Simon Willison found Qwen3.8-27B nearly perfect on arithmetic when reasoning is enabled but much weaker without it on large numbers. Verified provenance and AI commercialization are emerging as twin battlegrounds.

North America · First-hand

OpenAI

⭐⭐⭐ [Product Update] Our approach to EU text provenance rules

OpenAI outlined its approach to text provenance in response to the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine-readable way. Starting today, API customers globally can opt in to text watermarking for select models (off by default); in the coming weeks an invisible watermark will be added to eligible ChatGPT and Codex text output in the EU; and applications are opening for a text watermark detector, initially limited to approved researchers and expert organizations. The technology, called textGrain, adds an invisible statistical signal to the model's word choices that a detector looks for. OpenAI says textGrain matched or exceeded other tested approaches including SynthID for text, but detection is substantially lower for shorter or more constrained text such as mathematics, and editing weakens it: in 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%, and 25% replacement cut it to 17%. Verification tools for audio and images, including the openai.com/verify web tool and the Content Provenance API, remain publicly accessible.

⭐⭐ [Product Update] Building advertising for the way people use AI

OpenAI is introducing a new visual ad format in ChatGPT, initially tested during image generation, with ads clearly labeled, kept separate from the image being created, and not influencing the answers ChatGPT provides. Testing begins later this month in the US with an initial group of advertisers. OpenAI is also expanding its measurement tools and partner ecosystem, including integrations with Hightouch, Tealium, and LiveRamp for sending conversion data into ChatGPT Ads, attribution partners such as AppsFlyer, Triple Whale, Adjust, Branch, and Singular, and incrementality work with Haus, Measured, and WorkMagic on geo-based experiments. OpenAI says ChatGPT reaches 1.2 billion people each week, and cites early partner results: per DV Rockerbox, WeightWatchers' attributed cost per acquisition on ChatGPT Ads was 15.3% lower than its blended paid-search benchmark; WorkMagic reported statistically significant lift for Dose with 67% of incremental purchases from net-new customers; and per Triple Whale, 93% of Portland Leather's visitors from ChatGPT Ads were new.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐ [Product Update] Together Link: open models in the harness you already use. Start with one command today.

Together AI launched Together Link, which connects the coding harnesses teams already use to open models on Together AI and claims to cut spending by over 50%. It supports Claude Code, Claude Desktop, Codex (in the ChatGPT app and the CLI), OpenCode, and Pi, is installed with a single curl command, keeps existing settings and logins unchanged, and can be reverted to native closed models with one command. Its Auto mode uses the Together AI Router to read each session's first task and pick a model: with an Anthropic key it routes between Opus 5.5 and GLM 5.3, and without one between GLM 5.3 and GLM 5.3 Flash, routing once per session so prompt caching still works. A per-session tracker shows what was spent next to what the same session would have cost on Opus 5.5. Billing runs on the existing Together API key via serverless pay-as-you-go or credit packs, and Together AI says it serves the largest share of OpenRouter tokens for DeepSeek V4.1 Flash (40.8%), GLM 5.3 Flash (28.2%), and Kimi K3 (23.1%) as of 9/30/2026.

Research

⭐⭐ [Research] Qwen3.8 27B addition in words

Simon Willison reran an experiment originally done by Colin Frasier two years ago with GPT-4o: having a model add positive integers and express the exact results purely in English words. Running the local Qwen3.8-27B-Q4_K_M.gguf model on a DGX Spark with reasoning disabled, it reached 23.57% numeric accuracy across 5,070 cases with 96.17% format compliance, dropping from 97.04% for one- to three-digit operands to 6.44% for ten- to thirteen-digit operands. With reasoning enabled on a single run per combination (169 attempts), it answered 167 correctly, and the author notes a second run would likely differ. The report also includes reasoning traces from some of the larger calculations.