AI Weekly #28/2026: xAI Fires Back with Grok 4.5 — Cheaper Than Opus, Faster Than GPT-5.6

TL;DR

This week in 30 seconds:

• Grok 4.5: xAI launches its first model since going public and acquiring Cursor, at $2/$6 per million tokens — significantly cheaper than Anthropic’s Opus 4.7 ($5/$25)
• China: Beijing consults with Alibaba, ByteDance, and Z.ai on a tiered export control system for frontier models
• Layoffs: 120,000 tech jobs already cut in 2026, with AI the most-cited reason in May — despite record profits at Oracle and others
• Capital question: The industry is planning $1.5 trillion in 2026 CapEx, but analysts say it needs $3 trillion in revenue to justify it

---

Audio Version

11:53 | Download MP3

Chapters

- 0:00 - TL;DR - 0:48 - Story of the Week - 2:52 - More Top Stories - 6:01 - Quick Hits - 7:17 - Tool of the Week - 8:14 - Fail of the Week - 9:41 - Number of the Week - 10:34 - Reading List - 11:12 - Next Week

Read aloud with edge-tts (en-US-AndrewNeural)

---

Story of the Week

Grok 4.5: xAI Pivots from Chatbot to Enterprise Tool

$2 per million input tokens, versus $5 for Anthropic’s Opus 4.7 — with this pricing, xAI sends its first model since its own IPO and the Cursor acquisition into the race [1]. Grok 4.5 launched on July 8, 2026, exactly one week ahead of OpenAI’s GPT-5.6 launch announced for July 15, which had previously been temporarily blocked by the Trump administration over safety concerns [1].

The focus is now explicitly on coding, office work, and agentic tasks rather than the consumer chatbot [1]. xAI markets the model as “twice as token-efficient” compared to the competition [2]. On cost, Grok 4.5 significantly undercuts both Opus 4.7 ($5/$25 per million tokens) and OpenAI’s Sol model ($5/$30) [1]. EU users, however, will need to wait until mid-July before the model becomes available there [1].

“An Opus-class model, but faster, more token-efficient and lower cost — roughly comparable to Opus 4.7, but much faster.”

— Elon Musk [2]

For developer and ops teams, this means: anyone who has been hesitant about coding agents due to cost now has a noticeably cheaper alternative at a comparable quality level with Grok 4.5 — provided the “Opus-class” label holds up in practice.

Critical voices: Musk’s comparison to Opus 4.7 comes directly from xAI itself — independent benchmark results for Grok 4.5 were not yet available at time of writing. Open questions remain: how the price advantage plays out under real agentic workloads with high tool-call overhead is still unclear.

Bottom Line: Grok 4.5 marks xAI’s strategic shift away from the consumer chatbot toward a price war for enterprise and coding workloads — at the expense of Anthropic’s and OpenAI’s margins.

---

More Top Stories

Beijing Weighs Export Restrictions on Chinese AI Models

China’s Ministry of Commerce is consulting with Alibaba (Qwen), ByteDance (Doubao), and Z.ai (GLM-5.2) on restricting foreign access to the country’s most advanced domestic AI models [3]. Under discussion: harsher penalties for leaks as violations of the National Security Law, plus a tiered control system — basic open-source tools subject to registration, advanced models requiring security review, and frontier models restricted to domestic use only or not released at all [3]. Beijing is thus increasingly treating AI — much like the US does — as “a critical national asset that requires controls” [3].

The timeline and final scope of the measures remain open [3]. So what? Should China actually restrict its open-weight models like Qwen or GLM, it would hit the Western open-source scene hard — Qwen derivatives in particular are currently among the most widely used base models outside the US.

---

UN Global Dialogue on AI Governance: All 193 States at the Table

On July 6 and 7, the first session of the UN Global Dialogue on AI Governance took place in Geneva — with representatives from all 193 UN member states as well as industry, civil society, and academia [4]. A 40-member international panel of experts released the first comprehensive global governance assessment during the session [4]. Discussions covered opportunities such as productivity gains and improved healthcare systems, but also risks like disinformation and the erosion of democratic processes [4].

“Science cannot guarantee that AI will not cause catastrophic harm.”

— Yoshua Bengio [4]

Particularly notable: Ambassador Rein Tammsaar pointed to the geopolitical concentration of frontier AI development in just two countries — the US and China [4]. So what? For every other state (and their companies), this means a structural dependence on models whose regulation they have no say in.

---

AI Agent Startup Lets Its Own Agent Run $100M Funding Round

Jersey City-based enterprise agent startup Lyzr had its AI agent, “SivaClaw,” run its own Series B funding round — as a living proof of product [5]. The agent answered questions from over 130 investors, drafted investment memos, and even tracked which slides investors spent the most time viewing [5]. The round closed at $100 million in Series B funding at roughly a $500 million valuation, with $400 million in total interest from Silicon Valley, the Middle East, and the financial sector [5].

“There’s so much capital chasing AI bets that startup founders with traction barely have to leave their desks to raise nine figures.”

[5]

So what? Whether this is a PR stunt or a genuine preview of agentic fundraising processes remains to be seen — but the sheer demand of $400 million against a $100 million target shows just how overheated the AI funding market remains.

---

Quick Hits

Briefly noted:

• Meta: New MTIA chip generation enters production in September to cut GPU costs amid an unprecedented component shortage — planned 2026 CapEx: $125-145B, 7 GW of compute capacity with planned doubling in 2027 [6]
• DeepSeek: Chinese company is developing its own inference chip to reduce dependence on Nvidia and Huawei — already ramping up hiring of chip design engineers and holding confidential foundry talks [7]
• Google: Now labels AI-generated ads in “My Ad Center” for Search, YouTube, and Discover — automatic for Google’s own AI ad tools, manual for externally created ads, with no independent verification [8]
• Ollama: Raises $65M Series B led by Theory Ventures, bringing total funding to $88M with 8.9M monthly users and just 14 employees [9]
• Tech Layoffs: Oracle cuts 21,000 jobs (-13%) despite a $3.7B quarterly profit (+27%), explicitly citing “adoption and deployment of AI technologies” as the reason [10]

---

Tool of the Week

Ollama — Local AI models made as easy as Docker

Ollama makes open-source LLMs run locally without users having to worry about CUDA versions, quantization, or model formats [9]. Following its fresh $65M Series B, the tool now offers a cloud service with subscription tiers billed by GPU time rather than token limits, alongside the free desktop version [9]. Particularly useful for developer teams who want to test models locally or avoid cloud APIs for privacy reasons: a model like Llama or Mistral is ready to run with a single ollama run command.

“Open models came out in 2023 but were very hard to use.”

— Jeff Morgan, CEO Ollama [9]

With 176,000 GitHub stars and adoption at 85% of Fortune 500 companies, Ollama has effectively established itself as the standard interface for local AI models [9].

Link to tool

---

Fail of the Week

“AI-related” as a layoff reason — despite record profits

TechCrunch maintains a running list of major tech layoffs in 2026 where employers explicitly cited AI as the reason [10]. Roughly 120,000 tech jobs have already been cut in 2026, with AI the most frequently cited reason for layoffs in May [10]. The paradox: several affected companies reported record revenues in the same period. Oracle cut 21,000 jobs (-13% of its workforce over 12 months) while posting a $3.7B quarterly profit (+27%) [10]. GitLab trimmed 350 jobs (-14%) despite $264M in quarterly revenue (+23% YoY), citing a demand for “100x growth” in agentic workloads as justification [10].

Root cause: Companies are using “AI efficiency” narratives to justify cost cuts to investors, regardless of whether the actual operational driver is automation, shifting demand, or simply margin pressure — a blending of PR messaging and genuine technical cause.

What we’re learning: Anyone reading corporate “AI replaced X jobs” statements should check the quarterly numbers in parallel — often, the stated reason for layoffs says more about investor-relations strategy than about the actual degree of automation.

---

Number of the Week

$3 Trillion

That’s how much revenue the AI industry needs to generate, according to analysts, to justify the $1.5 trillion in infrastructure CapEx planned for 2026 [11]. For comparison: Anthropic currently generates roughly $60B in ARR, while OpenAI stood at $20B ARR as of November 2025 — together, well under a tenth of the required sum [11]. The gap is widening further amid growing competition from cheaper open-source models and falling token prices [11].

“The required revenue per GW of capital expenditure has risen sharply due to these bottlenecks and rising construction costs.”

— David Cahn, Sequoia Capital [11]

Apollo chief economist Torsten Slok warns of recession risk should hyperscalers miss their 2028 profit projections [11].

---

Reading List

For the weekend:

1. Can AI answer the $3 trillion question? - An in-depth look at the industry-wide CapEx-vs-revenue gap, with assessments from Sequoia and Apollo (8 min)
2. The running list of major tech layoffs in 2026 - Continuously updated overview of all AI-related layoffs in 2026, useful as a reference (5 min)
3. UN Global Dialogue on AI Governance - Summary of the first UN session on AI governance with quotes from Bengio and Ressa, a good overview of the global regulatory discourse (6 min)

---

Next Week

What’s coming:

• Wed 7/15: OpenAI plans to launch GPT-5.6, after the Trump administration had temporarily blocked the release over safety concerns [1]
• Grok 4.5 officially becomes available in the EU mid-July — that’s when we’ll see whether its price advantage over Opus 4.7 holds up in the European enterprise market too [1]
• Still watching: whether Beijing’s export control plans acquire concrete deadlines should become clearer over the coming weeks via official statements from the Ministry of Commerce [3]

---

🤖 Behind This Newsletter

Generated in: ~15 minutes Sources scanned: 11 articles from the story selection Stories found: 11 → 11 used (1 story of the week, 3 top stories, 5 quick hits, tool, fail, number of the week) Validation: Pending (Phase 4 follows) Model: Claude Sonnet 5 Images: Pending (Phase 3.5 follows)

Full metrics

Phase	Metric	Value
Source collection	Prepared story selection	1 file
Selection	Stories presented	11
Selection	Stories selected	11
Draft	Sources cited	11

---

This newsletter was researched and written AI-assisted. Images generated with Pollinations.ai.