This issue contains affiliate links and two labeled sponsor slots. Editorial rankings are never sold.
Hey,
Strange week. The big tools got dramatically more generous, and quietly more interested in your data. Both at once, often from the same companies.
THE WEEK IN ONE LINE
Free: ChatGPT gave its free tier unlimited text chats. Meta gave away a genuinely capable agent that runs on your own laptop, no account, no cloud, no tokens.
Not free: two deadlines land before Monday. One disconnects your ChatGPT integrations, the other turns your Jira into training data.
Here is all of it, in order of what actually needs your attention.
01 · Two deadlines before Monday
⏳ FRIDAY, AUGUST 14 · CHATGPT ENTERPRISE AND EDU
OpenAI is retiring individual-user sync for connected apps. From August 10 no new individually authorized sync connections can be created, and on August 14 existing ones are disabled, with deletion of the associated synced data beginning. Admin-managed sync is unaffected.
If your workspace connects Google Drive or similar, an admin needs to enable the corresponding plugin before Friday or those connections simply stop. Check your workspace connector settings today.
⏳ MONDAY, AUGUST 17 · ATLASSIAN
Content from Jira, Confluence, and Jira Service Management becomes eligible to train Atlassian's AI products, including Rovo, Rovo Chat, and Rovo Dev. That covers Confluence page titles and body content, Jira issue titles, descriptions and comments, custom workflow names, and custom field names.
It is opt-out, not opt-in. Around 300,000 organizations are enrolled by default.
Where the toggle lives: admin.atlassian.com → Security → Data contribution. Some accounts see it under AI features → Data usage, and it can apply per product, so check each. Confirm the path in Atlassian's own docs, because admin UIs move.
The Atlassian tiering is the sting. In-app content can be switched off by an admin. Metadata collection cannot be turned off at all on Free, Standard, or Premium.
Plan | In-app content | Metadata |
|---|---|---|
Free / Standard / Premium | Admin can opt out | Locked on |
Enterprise | Off by default | Off by default |
Opt out and Atlassian says it removes in-app data within 30 days, content attributes within 90, and retrains affected models. De-identified aggregate data may be kept up to seven years. Automatically exempt: HIPAA accounts, customer-managed encryption keys, Government Cloud, and Isolated Cloud.
02 · The great giveaway
ChatGPT's free tier stopped being a trial. Luna becomes the default model for Free and Go users this week, and those tiers now get unlimited text chats plus a new Think button for harder questions, subject to abuse guardrails. Limits still apply to file uploads, images, and other tools, so this is not unlimited everything. But "unlimited text chat, free" is a line nobody expected to cross this year, and it lands two weeks after the price cut that made Luna 80% cheaper to serve.
✦ MUSE GLIMMER · META · AUGUST 10
30 billion parameters. Apache 2.0. One consumer GPU. Meta put the weights on Hugging Face under a permissive licence, built for always-on local agents. Works with or without an internet connection.
The hardware math is the whole story. At full precision it needs over 55GB of memory, beyond any consumer card. Meta shipped 4-bit quantized builds that bring it under 20GB, plus a speculative-decoding drafter, so it fits a 24GB GPU or a high-end Mac and still runs fast enough for a real agent loop. Ollama, LM Studio, and llama.cpp all support it. Apache 2.0 means you can use it commercially, modify it, and ship it inside a product, which is notably more permissive than Llama ever was.
✓ Strong at: agentic orchestration and reasoning. Meta reports it leading its size class on MCP Atlas at 75.5 and DeepSearch QA at 74.6. The demo that captures the point: it finds a Home Assistant instance on the local network, queries device APIs, writes a dashboard from scratch, then deploys a local server to check its own work.
✗ Weak at: computer-use and terminal tasks, where Qwen3.6-27B stays ahead (OSWorld-Verified 65.9 against 75.6). And on CI Memories, a privacy benchmark where lower is better, Glimmer scores 26.4 against Gemma's 12.1. Worth knowing when the entire pitch is keeping data local.
Who should care. If you have a 24GB GPU or an M4 or M5 Max Mac, you can now run a capable agent on work you would rather not send anywhere, with no per-token bill. The sensible pattern is hybrid: repetitive and sensitive work stays local, genuinely hard reasoning still goes to a frontier model.
◆ SPONSORED
Miro: where the thinking happens before the ticket exists.
If this week has you auditing where your team's knowledge actually lives, Miro is the canvas half of that question: diagrams, planning, workshops, and messy early thinking, with AI that turns a sticky-note dump into something structured. The free plan is permanent and covers three editable boards with unlimited members, and Starter runs $8 per user per month on annual billing.
The ToolChase site has no ads, no pop-ups, and no cookie wall. Two labeled sponsors here are what pay for that, which also means our scores stay unbuyable and our reviews stay independent. If the partner above is useful to you, a click costs you nothing and keeps the whole thing running. Want one of the two slots? Here's how it works.
03 · Tool by tool: what the big names shipped
💬 ChatGPT
Beyond the free-tier change: it can now find and book restaurant reservations through OpenTable, Resy, and Yelp, rolling out across all plans on mobile, web, and desktop. Ask with a location, time, party size, and preferences, and available slots appear in the chat. New education plugins also landed for ChatGPT Edu and Teachers deployments, with schools controlling which tools are available.
🧠 Claude
The Compliance API now covers Cowork and Claude Code across desktop, web, mobile, and CLI, in beta for Enterprise. Security teams can pull session content and metadata for audits and eDiscovery through the same interface they already use for chats. Nothing changes about existing endpoints. Claude Code itself is shipping fixes at a relentless pace, including fullscreen terminal rendering and mobile push notifications when Remote Control is enabled.
✨ Gemini
Now past 900 million monthly users across 230 countries and 70+ languages. Google's Made by Google event is today in New York with the Pixel 11 family expected. The interesting question is not the hardware, it is whether Gemini does something materially useful on-device rather than another demo reel.
🎨 Who currently holds the crowns
On the independent leaderboards: ChatGPT Images 2.0 leads text-to-image at 1385 Elo and image editing at 1463, and remains the best at rendering readable multilingual text. Gemini Omni Flash tops text-to-video at 1527 Elo, a clear 45 points ahead of the field. Reve 2.1 and Meta's Muse Image round out the image podium.
⌨️ Codex
Portable Agent Plugins landed, searchable across local, personal, workspace, and remote catalogs. Also support for the new MCP spec, a --approve-for-me flag for automatically reviewed approvals, and the ability to import Cursor-managed skills without creating duplicates.
04 · The security footnote worth reading
Two firms published separate ways to make Rovo, the assistant about to be trained on all that Jira data, hand it to someone else.
PATH 1 · FIXED
Varonis "RovoBlast," August 7. A crafted link with a hidden prompt. A signed-in user clicks, Rovo treats the text as their query, searches what they can access, and ships results to an outside server. The proof of concept pulled a private API key from Confluence. Atlassian fixed this server-side on July 8 and the reporter validated it.
PATH 2 · UNRESOLVED
PromptArmor, August 5. Instructions hidden in content Rovo reads. An uploaded file was enough. It reportedly still worked with organization-wide web search disabled, because that setting does not remove the tool that opens results. PromptArmor says it disclosed on May 23 and got no substantive reply for over two months.
The nuance matters. Neither breaks Jira permissions. Both borrow the permissions of whoever is signed in, which for most employees still means confidential tickets and API keys. The link path is fixed. The content path is single-sourced and unconfirmed as remediated. No published evidence either has been used against a real organization.
The transferable lesson has nothing to do with Atlassian. Any assistant with broad read access and the ability to fetch a URL can be talked into combining those two powers by text it merely reads. Connect an agent to your company's knowledge and the attack surface becomes every document anyone points it at.
05 · Also worth knowing
⏸ OpenAI hit pause on its own model. The company said it is pausing some internal activities involving Astra, an upcoming model, after an internal evaluation found significant advances in agentic coding and cybersecurity, and that it is adding security controls for higher-capability models. A lab voluntarily slowing itself down is rare enough to note, though the disclosure is thin on specifics.
💰 Nvidia lined up roughly $500 billion in financing partnerships to help customers fund AI infrastructure. Morgan Stanley estimates hyperscalers could spend around $3.5 trillion between 2026 and 2028. Watch the circular financing, where the company selling the infrastructure also helps fund the buyers.
Full running log at AI Tools News 2026.
◆ SPONSORED
Netlify: ship it yourself.
The other half of owning your stack is owning where it runs. Netlify deploys sites and apps straight from Git or a drag-and-drop folder, with a free tier that covers real projects and no server to babysit. Disclosure worth making: ToolChase itself runs on it, which is why it is here.
06 · What we updated on ToolChase
Every tool page now carries the official product video, so you can watch a tool work before spending a trial on it. The pricing changes from the last three issues are reflected across the directory.
Given this week's theme, the guides worth your time: Best Free AI Tools, Best AI Coding Assistants, and the flagship 50 Best AI Tools in 2026.
07 · Prompt of the week: the training-data audit
Atlassian is not unusual. Opt-out-by-default is becoming the norm, and these changes arrive as a paragraph in a terms update nobody reads. Paste this in and find out where you stand:
You're a data governance analyst. Here are the SaaS and AI tools my team uses, and roughly what lives inside each: [LIST: tool, what data it holds].
1. For each tool, what is its current policy on using customer content to train AI models? Is it opt-in, opt-out, or not used at all?
2. Which ones have changed that policy in the last 12 months, or announced a change taking effect soon?
3. For any that are opt-out, tell me exactly where the setting lives and whether it is available on my plan tier.
4. Rank my tools by exposure: which holds the most sensitive content combined with the weakest opt-out?
Cite your sources with dates so I can verify, and flag clearly where you are uncertain rather than guessing.
Why it works: the answers drift, so treat the output as a checklist to verify rather than gospel. Point 4 is the one that changes behaviour. Most teams discover their most sensitive content sits in the tool with the least control, and that is a decision worth making deliberately instead of by default.
THE TAKEAWAY
The models are getting free and the defaults are getting hungry. Both trends reward the same ten-minute habit: know what you have switched on.
Reply and tell me one thing: are you switching off the Atlassian toggle, or leaving it? I read every reply, and I am curious how this splits.
See you next Wednesday,
Emre
Founder, ToolChase
Building an AI tool? Submit it for a free editorial review. Read The State of AI Tools 2026, our study of the full directory.