Three levers cut AI cost-per-outcome in production. Pick the right size model. Measure tokens at the application layer. Route cheap-by-default with premium escalation. Stop reading the pricing page. https://ailab.ee/2026/06/intelligence-per-dollar-is-the-new-buyer-question/

Article drafting in Publisher is now run end to end by an autonomous agent. Decides what the piece needs. Researches only when called for. Pulls from your knowledge base. Close the tab, come back to a finished draft. https://ailab.ee/2026/06/new-how-publisher-turns-your-performance-data-into-be...

Privacy in Play Text is the simplest kind. Your text never leaves your device. No server for it to go to. No telemetry. No accounts. Voices generated on-device. https://ailab.ee/2026/06/we-built-a-free-text-to-speech-app-in-two-evenings-and-here-is-how-it-works/

Centralised enforcement at the EU AI Office is the quietly important part of the AI Omnibus deal. Less member-state fragmentation, more consistent reads. Full take: https://ailab.ee/?p=1093

The best first use cases for coding agents are not giant rewrites. Start with regression fixes, tests, dependency upgrades, docs drift, migrations, and narrow refactors with clear checks. Agents need rails before freedom. https://ailab.ee/2026/06/the-next-ai-development-bottleneck-is-judgment/

WordPress sync in the new Publisher: fix a typo in Publisher, it lands on the live WordPress post in place. No new draft, no -2 duplicate. Reschedules and retries update the existing post. Self-heals stale references. https://ailab.ee/2026/06/whats-new-in-publisher-an-mcp-server-automatic-pipeli...

Seven sectors where your AI defaults to high-risk under EU rules: - Education - Employment - Essential services - Law enforcement - Justice - Biometrics - Critical infra https://ailab.ee/2026/06/is-your-ai-system-high-risk-the-eus-new-draft-guidelines-explained/

Play Text ships with Kokoro-82M baked in. 327 MB of neural voice, running locally on Apple silicon via MLX. Natural-sounding speech. No API key, no cloud bill, works offline. https://ailab.ee/2026/06/we-built-a-free-text-to-speech-app-in-two-evenings-and-here-is-how-it-works/

Microsoft's new coding model uses up to 60% fewer tokens than Claude Haiku 4.5. The trick is adaptive solution length. Short answers for easy tasks, longer ones only when needed. https://ailab.ee/2026/06/intelligence-per-dollar-is-the-new-buyer-question/

Lines of code are becoming a worse productivity metric. Anthropic says its engineers merge about 8x more code per day than in 2024. The real signal is pressure moving to review, tests, architecture boundaries, and release judgment. https://ailab.ee/2026/06/the-next-ai-development-bottleneck-is...

Publisher pipeline: drop a link in the queue, it drafts on-brand, asks for approval, publishes to WordPress, stitches the final URL into social copy automatically, and lands them in your schedule's next slots. https://ailab.ee/2026/06/whats-new-in-publisher-an-mcp-server-automatic-pipelines-and-...

If your firm is under 750 people and €150M revenue, the AI Act just got simpler. SMC compliance relief is real. What changed and why it matters: https://ailab.ee/?p=1093

Select text in any Mac app, Mail, Slack, a PDF, anywhere. Press the Play Text shortcut. A floating player opens with the full text and starts reading. Always-on-top so you can move on to the next thing. https://ailab.ee/2026/06/we-built-a-free-text-to-speech-app-in-two-evenings-and-here-is-how-i...

The AI in Publisher now consults your real numbers before drafting. Your top posts. Best channels. Winning formats. Strongest timing. The agent shapes each draft to match what works for your audience. https://ailab.ee/2026/06/new-how-publisher-turns-your-performance-data-into-better-drafts/

A weak ticket now creates a large vague diff. A strong specification creates a controlled agent task. As coding agents take on longer work, specs become more valuable than tickets. Clear goals, constraints, and acceptance checks beat better prompts. https://ailab.ee/2026/06/the-next-ai-develop...

Publisher shipped today. It now reads your real performance data and feeds it back into the AI that drafts and schedules. A content tool that gets measurably smarter from the posts you publish. https://ailab.ee/2026/06/new-how-publisher-turns-your-performance-data-into-better-drafts/

The industry has a name for AI waste now: tokenmaxxing. Amazon staff were caught burning tokens to climb internal leaderboards. Models pad answers to game benchmarks. Both versions hit the same invoice. https://ailab.ee/2026/06/intelligence-per-dollar-is-the-new-buyer-question/

Tuesday morning workflow with the new Publisher MCP server: tell Claude to draft a post from your notes, schedule it on WordPress, queue 4 social variants. It happens inside Publisher. Approvals still required. Token-scoped. https://ailab.ee/2026/06/whats-new-in-publisher-an-mcp-server-automatic...

Two ways your AI gets classified high-risk under Article 6: 1. Safety component of an Annex I regulated product 2. Use case from Annex III Full breakdown: https://ailab.ee/2026/06/is-your-ai-system-high-risk-the-eus-new-draft-guidelines-explained/

Open a long Safari article. Click Play Text. Choose "Read article" or "Pick a section". A pink pill sweeps word by word on the live page as it speaks. You stay in context, you just stop reading. https://ailab.ee/2026/06/we-built-a-free-text-to-speech-app-in-two-evenings-and-here-is-how-it-works/

This week's LLM charts, explained simply. Three different leaders: Claude Fable 5 is smartest, MiniMax M3 is now most-used on OpenRouter, and Claude is hardest to trick (F5). No single model wins all three. https://ailab.ee/2026/06/llm-charts-explained-week-of-16-june-2026/

The thing about the new Publisher release: MCP server, pipelines, and WordPress rebuild all share the same publishing code underneath. Agent, human, or pipeline, the result is the same. Mixed workflows finally feel safe. https://ailab.ee/2026/06/whats-new-in-publisher-an-mcp-server-automatic-pip...

Publisher just shipped its biggest update yet. 58 changes in two weeks. Highlights: MCP server (Claude works inside your workspace), end-to-end pipelines, and a rebuilt WordPress integration with two-way sync. https://ailab.ee/2026/06/whats-new-in-publisher-an-mcp-server-automatic-pipelines-and...

EU just published the missing manual on classifying high-risk AI. The Commission's draft guidelines walk through Article 6 in plain terms. https://ailab.ee/2026/06/is-your-ai-system-high-risk-the-eus-new-draft-guidelines-explained/

Behind Play Text. Our founder wanted to listen to articles like a podcast. Existing apps cost too much for what they did. Two evenings with Claude Code on Opus 4.8 later, Mac + iOS apps existed. Free. https://ailab.ee/we-built-a-free-text-to-speech-app-in-two-evenings-and-here-is-how-it-works/

We just shipped Play Text. A free text-to-speech app for Mac and iPhone with live word-by-word highlighting. Built in two evenings. Free. No cloud, no subscription, no tracking. https://ailab.ee/we-built-a-free-text-to-speech-app-in-two-evenings-and-here-is-how-it-works/

The next AI development bottleneck is not writing code. It is judgment. Anthropic says 80%+ of code merged into its own codebase in May was Claude-authored. If agents produce code faster than teams can review it, review becomes the system. https://ailab.ee/the-next-ai-development-bottleneck-is...

Anthropic just released Claude Fable 5, making a Mythos-class model safe for general use. It hits state-of-the-art on nearly all tested benchmarks and beats their past public models. We are testing if it justifies the commotion. https://www.anthropic.com/news/claude-fable-5-mythos-5

GPT-5.5 and Claude Opus 4.8 score within a point on the Intelligence Index. Running it costs $3,357 on one, $4,685 on the other. Same answer, 40% more expensive. https://ailab.ee/intelligence-per-dollar-is-the-new-buyer-question/

Microsoft cancelled Claude Code licenses across its biggest engineering division. Two weeks later they shipped their own coding model that uses 60% fewer tokens. The two announcements are the same story. https://ailab.ee/intelligence-per-dollar-is-the-new-buyer-question/

Yes! This 👇

You finally have the headroom to test less-technical AI agents. Claude Cowork just doubled its 5-hour usage limits for Pro, Max, and Team users. Time to get hands-on and learn what actually works for you. https://support.claude.com/en/articles/15400594-claude-cowork-june-2026-usage-promotion

EU just delayed high-risk AI obligations. Annex III to 2 Dec 2027, Annex I to 2 Aug 2028. The new timeline and what to do with the extra 12-16 months: ailab.ee/what-the-eu-...

Needed a tool to track office listings. Our AI assistant built one in minutes. Needed competitor monitoring? Built that too. We gave it the ability to create custom micro apps on the fly. https://ailab.ee/our-ai-executive-assistant-creates-purpose-built-apps-on-the-fly/

We gave our AI agent its own M365 account. It reads emails, detects promises, and once apologized on our CEO's behalf while he was stuck in traffic. Works ~80% of the time. The other 20% keeps us humble. https://ailab.ee/giving-an-ai-agent-its-own-microsoft-365-account/

What does it take to give an AI agent its own email? Dedicated sub-agents, a communication etiquette system, and a lot of trial and error. Here's what we built and what broke. https://ailab.ee/giving-an-ai-agent-its-own-microsoft-365-account/

Together with @heysec.com we built a fully autonomous cybersecurity research factory. It researches prompt injection vulnerabilities, builds research cases with PoC exploits, creates firewall rules to mitigate these threats, and iteratively tests those against active exploits.

We quietly released a new product a week ago. Originally developed as an internal AI agent, it has since evolved into a full platform with a visual studio and advanced agent orchestration. https://ailab.ee/support-agent/

These agents' confusion moments are a great example of this other 20-30%. 👉 https://ailab.ee/we-turned-openclaw-into-a-fully-functioning-executive-assistant/

We’re putting the OpenClaw EA to the real test and asking it to handle the move of our offices. Let’s see how that goes 🤞 👉 We published more about what this AI EA really is all about: https://ailab.ee/we-turned-openclaw-into-a-fully-functioning-executive-assistant/

You can now also listen to all our blog posts 🎧 Give it a try on the latest blog post about how we built an EA on OpenClaw: https://ailab.ee/we-turned-openclaw-into-a-fully-functioning-executive-assistant/

We turned OpenClaw into an AI Executive Assistant that handles real email, calendar, and communications every day. It works about 70-80% of the time. The other 20-30% has been... educational. https://ailab.ee/we-turned-openclaw-into-a-fully-functioning-executive-assistant/

We’ve now built agents that autonomously optimize and improve other teams of agents. This includes optimizing not only data or prompts but the entire codebase. Of course, it includes guardrails and easy rollbacks to maintain stability and recover from possible mistakes.

👀

We’ve just launched our new website ✨ We feel this design represents us much better. As always, the design and code were 100% generated by AI. https://ailab.ee/

Version 2 of the Support Agent we built a while back is coming along really well.

It doesn’t always have to be a large and complicated application to improve the user experience with AI. Sometimes just a tiny addition will add the most value. Like this integrated spellchecker using Gemini 3 Flash.

Happy holidays! 🎄

Here’s how we built our AI-powered customer support agent: https://ailab.ee/projects/intelligent-customer-support-agent/

We gave our AI support agent trust issues. Best decision we made. https://ailab.ee/2025/09/how-we-solved-a-16-year-customer-support-bottleneck-with-ai/