Dakarda AI Newsletter · 10 July 2026
Friday Edition
Three AI Fronts – 24 Days to the Deadline
OpenAI releases GPT-5.6 in three variants, SpaceXAI (xAI + SpaceX) strikes with Grok 4.5 after acquiring Cursor for $60B, and Meta breaks a long-standing rule by introducing a paid model for the first time. Meanwhile, the EU counts down to mandatory AI content labeling, and Uber tests a novel agent deployment model in the enterprise. It's been an intense week.
Intro · Alex
This week, the AI market exploded on three fronts simultaneously. OpenAI showed it can still surprise — Sol beats competitors on agent tests at a fraction of the cost. SpaceXAI with Grok 4.5 proves that merging with Cursor wasn't just a Musk gesture, but a real enterprise entry strategy. And Meta? After years of giving away models for free, Meta finally says: 'nothing in life is free.' In today's edition, I've traced five events that will reshape the power dynamics for months to come. From benchmarks and prices, through EU regulations, to a practical case study of agent deployment in a corporation. Plus a tech section, tool of the week, and a tip of the day that will be useful for anyone working with AI agents.
What's worth knowing
OpenAI releases GPT-5.6: Sol beats competitors, ChatGPT Work enters offices
The GPT-5.6 model family in three variants — Sol (flagship, most powerful), Terra (balanced), and Luna (fastest, cheapest). Sol surpasses Claude Fable 5 by 13.1 points on the Agents' Last Exam test, doing so with 61% less time and at half the cost. At the same time, OpenAI presented ChatGPT Work — an agent integrated with Codex that performs tasks in spreadsheets, calendars, and emails, with an oversight option for companies.
SpaceXAI strikes with Grok 4.5 — model after acquiring Cursor for $60B
SpaceXAI (formerly xAI) released Grok 4.5 — a model co-trained with the acquired Cursor, designed primarily for coding and agent tasks. Pricing is aggressive: starting at $2.5/M tokens compared to $10-15 for competitors. The model is positioned as 'Opus-class, faster, cheaper' — a direct threat to Claude Code and OpenAI Codex.
Meta Muse Spark 1.1 — the first paid model in Meta AI history
Meta released Muse Spark 1.1 with 1M token context, surpassing older versions of OpenAI and Anthropic on benchmarks. For the first time, Meta offers paid access through the Meta Model API ($1.25/M input, $4.25/M output). This marks the end of the free AI era from Meta and signals a new phase of monetization.
EU: 24 days to mandatory AI content labeling — Action Plan published
The European Commission published the Action Plan on Cybersecurity and AI and a formal opinion on the Code of Practice on Transparency of AI-Generated Content. This triggers a 24-day countdown to mandatory AI content labeling (Art. 50 AI Act, effective August 2, 2026). At the same time, the EU is building a pre-release evaluation system for advanced models, which will impact the release strategies of all players in the European market.
Uber tests 'Agentic Pods' — embedded AI in HR, finance, and legal departments
Uber's CTO deployed 30 top AI engineers directly into HR, finance, and legal departments for 2 weeks. Instead of a central AI team, engineers observed work, identified tedious processes, and built agents that automate tasks (e.g., financial reports requiring access to multiple systems). The company honestly admits that agent ROI is still unproven.
From the tech world
GitHub 'Verified' Commits can be rewritten without breaking the signature
Researchers discovered that a signed commit with a green 'Verified' badge can be transformed into a new hash while retaining a valid signature. This means systems that block malicious commits by hashes are useless — an attacker can push the same content under a new 'trusted' hash.
Windows Defender 0-day patch could clog your disk
Microsoft patched the RoguePlanet zero-day in Windows Defender, but the patch introduces a new problem — writing Zone.Identifier ADS without size limits, which can completely fill up the disk. A classic case where a security patch creates a new DoS vulnerability.
Tip of the day
How to test AI agents before enterprise deployment
Inspired by Uber's Agentic Pods model, before deploying an AI agent in your company, conduct a 'shadow audit' — for a week, record (in text form, with team consent) repetitive tasks performed by humans. Then map which ones require access to multiple systems and which are purely cognitive. This will help distinguish tasks the agent can actually automate from those requiring human judgment. Practical step: create a simple decision matrix. X-axis — number of systems the agent needs to access (1-5). Y-axis — risk of error (low/high). Deploy agents only in the 'low risk, many systems' quadrant — that's where ROI is immediate. For high-risk tasks, introduce a 'human-in-the-loop' mode where the agent prepares a draft and a human approves it. Uber does the same — and as they admit, agent ROI is still not obvious, which is why they start with safe areas.
Tool of the issue
ChatGPT Work — an office agent that works with spreadsheets, calendars, and emails
A new AI agent from OpenAI integrated with Codex, working on web, desktop, and mobile. It performs tasks in office applications (spreadsheets, calendar management, email handling) with an oversight option for companies. Codex has been fully absorbed into the new ChatGPT desktop app.
Reading list
OpenAI GPT-5.6: Sol, Terra, and Luna — what do they mean for the market?
TechCrunch analyzes in detail the architecture of the three variants, prices, and market positioning. Worth reading to understand why OpenAI chose a price segmentation strategy instead of a single model.
SpaceXAI Grok 4.5 — first enterprise entry after fusion with SpaceX
A detailed description of the model, pricing, and strategy. Why Musk is targeting enterprise coding and how Grok 4.5 at $2.5/M tokens could destabilize the entire API market.
Three major model launches in two days, but it's the regulations and practical deployments (Uber, EU) that may have a bigger impact on how AI actually changes our daily work — than yet another percentage point on a benchmark.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.