The most important news pieces from the AI world are here for you, whether you’re an AI enthusiast or just wish to keep an eye on what’s going on.
Focus Topic: One lab adds the watermark, the other hides it
Anthropic published a support page this week describing how it will embed invisible watermarks into text, code and files that come out of Claude. The mark survives copy-paste outside the platform, files carry the C2PA provenance label the industry already uses, and older models will be retrofitted while models shipped after August 2 have it built in. Anthropic is applying the EU AI Act’s transparency rules to its entire user base globally, not just in Europe. The company also plans to release its own detection tools, with one honest caveat: a mark means the content was “processed by Claude”, not necessarily written by it.
Four days later, Google went the other way. The Gemini app now has a Media watermark setting that lets you turn off the visible spark icon in the corner of images, videos and songs made with Nano Banana, Omni and Lyria. The option is hidden in countries where the law requires the mark, and the invisible SynthID watermark plus C2PA metadata stay in place either way. So the content is still traceable in principle. What changed is that it no longer looks like AI at a glance.
Put the two together and you get the actual state of AI provenance in 2026: the technical marking is getting better while the visible signal is getting weaker. A watermark that only a detection tool can read is not a label for the person reading the document. It is evidence you can retrieve later, if you know to look, if you still have the original file, and if you have a tool that reads that particular lab’s mark.
That distinction matters more in finance than in marketing. A watermark answers a forensic question, not a control question. If a board memo, a valuation model or a section of production code was drafted by Claude, an invisible mark can prove it after the fact. It does not tell your reviewer, at the moment of review, how much of this was machine-drafted and how much a human actually checked. And it does nothing at all for the AI tools that embed no mark, which is still most of them.
There is a second problem worth naming. Provenance is currently per-vendor. Anthropic marks Claude output with Anthropic’s scheme and plans Anthropic’s detector. Google embeds SynthID. C2PA is the one piece both share, and it lives in file metadata, which strips out the moment someone pastes text into an email or exports to a new format. Nobody is going to run four different detectors on every document that crosses their desk.
My simple advice:do not wait for the labs to solve this for you. If AI touched a deliverable, write it down yourself. One line at the bottom of the file, or one column in your document log: which tool, what it did, who reviewed it. It takes ten seconds and it survives copy-paste, format changes and vendor switches, which is more than any watermark can promise right now.
LLMs & AI Models
- 𝗦𝗽𝗮𝗰𝗲𝗫𝗔𝗜 launched 𝗚𝗿𝗼𝗸 𝟰.𝟲, scoring 61 on the Intelligence Index and beating GPT-5.6 Sol, at 𝟲𝟬% 𝗹𝗲𝘀𝘀 𝗰𝗼𝘀𝘁 than the frontier.
- 𝗢𝗽𝗲𝗻𝗔𝗜 previewed 𝗨𝗹𝘁𝗿𝗮𝗳𝗮𝘀𝘁, a Cerebras-powered API tier running GPT-5.6 Sol 𝘂𝗽 𝘁𝗼 𝟭𝟰𝘅 𝗳𝗮𝘀𝘁𝗲𝗿, at 750 tokens per second.
- 𝗚𝗼𝗼𝗴𝗹𝗲 released 𝗚𝗲𝗺𝗶𝗻𝗶 𝗙𝗹𝗮𝘀𝗵 𝟯.𝟳, with coding and knowledge-work upgrades at a price that undercuts Sonnet 5 and GPT-5.6 Terra.
- 𝗭𝗵𝗶𝗽𝘂 says 𝗚𝗟𝗠-𝟱.𝟯 is the strongest open-weights coding model, beating OpenAI on coding tests and finding 2,436 vulnerabilities.
- 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸 shipped an improved 𝗩𝟰 𝗣𝗿𝗼, open-sourced its agent software and raised API prices.
- 𝗠𝗲𝘁𝗮 released 𝗠𝘂𝘀𝗲 𝗚𝗹𝗶𝗺𝗺𝗲𝗿, a fully open model small enough to run agents on a laptop, with Muse Spark 1.2 weights promised next.
- 𝗢𝗽𝗲𝗻𝗔𝗜 launched 𝗚𝗣𝗧-𝟱.𝟲-𝗖𝘆𝗯𝗲𝗿, which answers 95% of advanced attack requests the standard model refuses, for vetted defenders only.
- 𝗔𝗽𝗽𝗹𝗲 built a China-specific model 𝘄𝗶𝘁𝗵 𝗔𝗹𝗶𝗯𝗮𝗯𝗮, bringing Apple Intelligence to Chinese iPhones within months.
- 𝗡𝘃𝗶𝗱𝗶𝗮 released 𝗡𝗲𝗺𝗼𝘁𝗿𝗼𝗻 𝟯.𝟱 𝗟𝗶𝗴𝗵𝘁𝗻𝗶𝗻𝗴, a small fast model, and expects Nemotron 4 to match the best open models available.
- 𝗟𝗧𝗫 released 𝗟𝗧𝗫-𝟮.𝟱, an open-weights world model that tops Gemini Omni Flash on internal video speed and quality tests.
- 𝗕𝘆𝘁𝗲𝗗𝗮𝗻𝗰𝗲 is reportedly pre-training a model with 𝘂𝗽 𝘁𝗼 𝟭𝟬𝗧 𝗽𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿𝘀, triple China’s largest to date.
- 𝗦𝗽𝗮𝗰𝗲𝗫𝗔𝗜 𝗮𝗻𝗱 𝗖𝘂𝗿𝘀𝗼𝗿 launched 𝗚𝗿𝗼𝗸 𝗕𝗼𝘁, turning Grok into a chat of autonomous teammates with their own computers and app access.
- 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 can now remember what you did on your 𝗠𝗮𝗰, without relying on screenshots.
New Tools
- 𝗪𝗵𝗮𝘁𝘀𝗔𝗽𝗽 is testing 𝗦𝗰𝗮𝗺 𝗗𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻, an on-device AI that flags suspicious chats from non-contacts without any message leaving the phone.
- 𝗚𝗼𝗼𝗴𝗹𝗲’𝘀 𝗚𝗲𝗺𝗶𝗻𝗶 𝗮𝗽𝗽 now lets you 𝘁𝘂𝗿𝗻 𝗼𝗳𝗳 𝘁𝗵𝗲 𝘃𝗶𝘀𝗶𝗯𝗹𝗲 𝘄𝗮𝘁𝗲𝗿𝗺𝗮𝗿𝗸 on images, videos and songs, while SynthID and C2PA stay embedded.
- 𝗚𝗮𝗺𝗴𝗲𝗲, a YC-backed startup, builds 𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘀𝗲𝗱 𝗺𝗥𝗡𝗔 𝗰𝗮𝗻𝗰𝗲𝗿 𝘀𝗵𝗼𝘁𝘀 𝗳𝗼𝗿 𝗱𝗼𝗴𝘀, after its founder made one for his own dog with AI tools.
- 𝗨𝗯𝗲𝗿 𝗮𝗻𝗱 𝗣𝗼𝗻𝘆.𝗮𝗶 plan to bring 𝟮,𝟬𝟬𝟬 𝗿𝗼𝗯𝗼𝘁𝗮𝘅𝗶𝘀 𝘁𝗼 𝗘𝘂𝗿𝗼𝗽𝗲.
- 𝗦𝗲𝗹𝗳-𝗱𝗿𝗶𝘃𝗶𝗻𝗴 𝘁𝗿𝘂𝗰𝗸𝘀 are officially 𝘁𝗲𝘀𝘁𝗶𝗻𝗴 𝗼𝗻 𝗖𝗮𝗹𝗶𝗳𝗼𝗿𝗻𝗶𝗮 𝗵𝗶𝗴𝗵𝘄𝗮𝘆𝘀.
- 𝗚𝗼𝗼𝗴𝗹𝗲 introduced the 𝗣𝗶𝘅𝗲𝗹 𝟭𝟭 𝗹𝗶𝗻𝗲𝘂𝗽, embedding Gemini deeper with voice upgrades, multistep tasks and proactive prompts.
Other Quick Picks
- 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 will embed 𝗶𝗻𝘃𝗶𝘀𝗶𝗯𝗹𝗲 𝘄𝗮𝘁𝗲𝗿𝗺𝗮𝗿𝗸𝘀 in Claude’s text, code and files, applying EU AI Act rules globally.
- 𝗧𝗵𝗿𝗲𝗲 𝗖𝗹𝗮𝘂𝗱𝗲 𝗮𝗴𝗲𝗻𝘁𝘀 sharing one server fell into a 𝗳𝗼𝘂𝗿-𝗵𝗼𝘂𝗿 𝘁𝘂𝗿𝗳 𝘄𝗮𝗿, sabotaging and locking each other out.
- 𝗢𝗽𝗲𝗻𝗔𝗜’𝘀 revenue run rate passed $𝟰𝟬𝗕, still behind Anthropic’s $47B ahead of both going public.
- 𝗦𝗽𝗮𝗰𝗲𝗫 closed its $𝟲𝟬𝗕 𝗽𝘂𝗿𝗰𝗵𝗮𝘀𝗲 𝗼𝗳 𝗖𝘂𝗿𝘀𝗼𝗿, with the Cursor brand likely to be phased out.
- An 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 𝗮𝗴𝗲𝗻𝘁 𝗵𝗮𝗰𝗸𝗲𝗱 𝗮 𝗴𝘆𝗺 booking system to bump another member off the waitlist. No undo.
- An unreleased 𝗖𝗹𝗮𝘂𝗱𝗲 𝗺𝗼𝗱𝗲𝗹 𝗺𝗮𝗱𝗲 𝗽𝗿𝗼𝗴𝗿𝗲𝘀𝘀 on the unsolved 𝗥𝗶𝗲𝗺𝗮𝗻𝗻 𝗵𝘆𝗽𝗼𝘁𝗵𝗲𝘀𝗶𝘀.
- 𝗢𝗽𝗲𝗻𝗔𝗜 𝘀𝘁𝗮𝗳𝗳 blame a 𝗿𝘂𝘀𝗵 𝘁𝗼 𝘀𝗵𝗶𝗽 for the rogue agent hack.
- 𝗞𝗶𝗺𝗶 𝗞𝟯 escaped a test sandbox to pull an 𝗮𝗻𝘀𝘄𝗲𝗿 𝗸𝗲𝘆 𝗼𝗳𝗳 𝗚𝗶𝘁𝗛𝘂𝗯, where rivals refused.
- 𝗜𝗴𝗼𝗿 𝗕𝗮𝗯𝘂𝘀𝗰𝗵𝗸𝗶𝗻 raised $𝟭.𝟭𝗕 𝗳𝗼𝗿 𝗥𝗶𝘃𝗲𝗿 𝗔𝗜, open models a person or business actually owns.
- 𝗭𝘂𝗰𝗸𝗲𝗿𝗯𝗲𝗿𝗴’𝘀 6,500-word superintelligence manifesto drew 𝗯𝗮𝗰𝗸𝗹𝗮𝘀𝗵 for glossing over AI harms.
- 𝗚𝗼𝗼𝗴𝗹𝗲’𝘀 𝗚𝗲𝗺𝗶𝗻𝗶 𝗮𝗽𝗽 hit 𝟭𝗕 𝘂𝘀𝗲𝗿𝘀, its fastest-growing product ever.
🇪🇪 AI News from Estonia
- 𝗡𝗲𝗯𝗶𝘂𝘀 is opening a 𝟮𝟮 𝗠𝗪 𝗔𝗜 𝗳𝗮𝗰𝘁𝗼𝗿𝘆 at Greenergy Data Centers in Hüüru near Tallinn, its first AI infrastructure site in the Baltics.
- The state funded the 𝗳𝗶𝗿𝘀𝘁 𝘀𝗲𝘃𝗲𝗻 𝗽𝘂𝗯𝗹𝗶𝗰-𝘀𝗲𝗰𝘁𝗼𝗿 𝗔𝗜 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 with 𝟭.𝟯𝟮𝗠 𝗲𝘂𝗿𝗼𝘀, including automated validation of PRIA applications.
- 𝗠ä𝗿𝘁 𝗥𝗶𝗱𝗮𝗹𝗮: in companies of 10 to 200 people, most data value is lost because decisions still run on 𝗵𝗮𝗻𝗱-𝗯𝘂𝗶𝗹𝘁 𝗘𝘅𝗰𝗲𝗹 𝘀𝗵𝗲𝗲𝘁𝘀.
- The biggest corporate cyber risk may be 𝗲𝗺𝗽𝗹𝗼𝘆𝗲𝗲𝘀’ 𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹 𝗔𝗜 𝗮𝗰𝗰𝗼𝘂𝗻𝘁𝘀, where work data leaves the employer’s control entirely.
- 𝗠𝗮𝗿𝗸𝗲𝗗𝗿𝗼𝗶𝗱 passed Elering’s 𝗮𝗙𝗥𝗥 𝗾𝘂𝗮𝗹𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝘁𝗲𝘀𝘁, opening a third revenue stream for battery storage owners.
- Estonian 𝗜𝗧 𝗷𝗼𝗯 𝗽𝗼𝘀𝘁𝗶𝗻𝗴𝘀 𝗴𝗿𝗲𝘄 𝟮𝟭.𝟮% in the first half, with 𝟳𝟳𝟵 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝗻𝘁𝘀 for a single entry-level AI data scientist role.
🎙️ AIPowerment Podcast episode 91 wraps up July’s key AI news: new flagship models, next-gen AI agents, copyright and security cases, plus Estonia’s growing AI adoption – from LHV’s new AI features to a court reminder that AI answers always need checking.
🎧 Listen to AIPowerment Podcast on Spotify, Apple Podcasts, and YouTube.
Want to stay in the loop?
Straight to your inbox: practical AI updates, finance use cases, tools to try, and upcoming trainings.
I cut the noise and send only what’s actually useful.
💼 Exploring how to make AI work in finance? Let’s talk practical use cases – connect with Gerlyn Tiigemäe for expert guidance.
📚 If you would like to participate in one of my trainings or listen to speaking engagements, here are the upcoming ones:
- Training “AI võimalused finantsvaldkonnas” @ Äripäeva Akadeemia, for intermediate level @ 21.08 – information, for beginners @ 08.09 – information;
- Training “AI võimalused projektijuhtidele” @ Äripäeva Akadeemia, next trainings @ 29.09 – information
- I will be doing couple of training together with Eesti.ai algatus , which are all free, so check out their website to find suitable trainings here.
