The Week in AI: June 22-28, 2026 (Week 26)

News of the week: June 22-28, 2026 (Week 26)

The week was dominated by the question of who gets to use the most powerful AI models. The US government asked OpenAI to release the new GPT-5.6 only to a small group of partners for now, while partially freeing up Anthropic’s Mythos 5 for more than a hundred American organizations. OpenAI also unveiled Jalapeño, its first in-house chip. Meanwhile, companies are starting to cut back on tokens, and money keeps pouring into chips, data centers and agent startups.

New Models

OpenAI introduces GPT-5.6 in three versions, but only for a select few

On June 26, OpenAI announced a new family of models, GPT-5.6, in three versions: Sol for the hardest tasks, Terra for everyday business work and Luna as a fast, low-cost option. For now only about twenty organizations have access, with a wider release promised “in the coming weeks” (VentureBeat).

Prices per million tokens (a token is a small piece of text, roughly part of a word) are 5 USD for input and 30 USD for output for Sol, 2.50 and 15 USD for Terra, and 1 and 6 USD for Luna. On the terminal coding benchmark Terminal-Bench 2.1, Sol scored 91.9% in its highest mode, compared with 88% for Claude Mythos 5.

The limited release was not the company’s decision alone. According to TechCrunch, the White House Office of the National Cyber Director and Office of Science and Technology Policy requested it on June 25, reportedly over safety concerns (TechCrunch).

Sam Altman reportedly told employees that the government will approve access “customer by customer” during the preview. OpenAI added on its blog that such an approach should not become the long-term norm (TechCrunch).

Asian companies release Mythos-style models

While the US export ban keeps Mythos 5 out of reach of users abroad, Asian companies are offering their own alternatives. Japanese startup Sakana AI launched its Fugu model, and on June 24 Chinese security firm 360 unveiled Tulongfeng, a tool for finding software vulnerabilities (TechCrunch).

Sakana co-founder David Ha summed it up by saying that access to the best models “can disappear overnight”. 360 also showed off Yitianzhen, a tool for cyber defense and incident response.

Regulation and Law

US government partially releases Claude Mythos 5

On June 26, the US government allowed Anthropic to make its Mythos 5 model available again to roughly a hundred American companies and federal agencies. This comes from a Commerce Department letter confirmed by CNBC (CNBC).

Anthropic had previously shut down both of its newest models, Mythos 5 and Fable 5, to comply with an export directive citing national security. The directive required cutting off access for all foreign nationals, including the company’s own employees.

Commerce Secretary Howard Lutnick wrote in the letter that safeguards are in place for “certain trusted partners”. Anthropic posted on X that the model can be deployed at organizations that operate and defend critical infrastructure (TechCrunch).

The release does not cover Fable 5, so that model remains unavailable to the public.
The court dispute between Anthropic and the Pentagon, which earlier labeled the company a supply chain risk, is still ongoing.

Claude may now ask for your ID

On June 22, TechCrunch pointed out a new section of Anthropic’s privacy policy, effective July 8. In some cases, the company may request age or identity verification using a scan of a passport or driver’s license and a selfie (TechCrunch).

According to the company, verification should only apply to a small group of flagged accounts. The checks are handled by Persona, which also stores a facial geometry template.

Products and Agents

Claude Tag as a permanent colleague in Slack

On June 23, Anthropic introduced a beta of Claude Tag for Claude Enterprise and Claude Team customers. It is an always-present assistant in Slack that people tag like a team member and assign tasks to (TechCrunch).

The agent remembers company context and breaks tasks down into steps. Admins decide which tools and channels it can access, and in what is called ambient mode it can join a conversation on its own.

OpenAI and Trail of Bits fix open source

On June 22, OpenAI launched the Patch the Planet initiative with security firm Trail of Bits. Experts use tools such as Codex Security to find bugs in open source software (code that is freely available to everyone), verify them and help maintainers with fixes (TechCrunch).

June 23 – Email company Superhuman acquired GPTZero, a detector of AI-written text. According to its CEO Edward Tian, it has over 19 million users and 30 million USD in annual recurring revenue (TechCrunch).
June 24 – Meta is turning Creator Studio into a standalone app with an AI assistant for creators on Facebook. It advises when to post and suggests replies to comments, which the creator approves (TechCrunch).
June 25 – According to card payment data from Indagari, the number of paying Claude consumers and their spending have both grown by about 75% since January. ChatGPT, however, still has far more paying users (TechCrunch).

Chips and Infrastructure

OpenAI shows its own Jalapeño chip

On June 24, OpenAI unveiled Jalapeño, its first in-house chip, designed and built with Broadcom. It is an inference chip, meaning it runs already-trained models when they answer users (TechCrunch).

The chip is still in testing, and OpenAI reports significantly better performance per watt than current alternatives. The company says its own models helped develop it, another step by the big players toward less dependence on Nvidia.

Micron cashes in on the memory shortage

Memory chip maker Micron announced on June 24 that its third-quarter revenue quadrupled year over year to 41.45 billion USD, with profit of 28.2 billion USD. The stock rose more than 13% after the results (TechCrunch).

The company also signed a deal to supply memory chips and storage to Anthropic and took part in its Series H funding round. It expects revenue of 49 to 51 billion USD next quarter.

June 22 – SpaceX signed a compute deal with open source lab Reflection AI at the Colossus 2 data center near Memphis. Starting July 1, Reflection will pay 150 million USD a month, with a total value of up to 6.3 billion USD (TechCrunch).
June 22 – AI chipmaker Groq confirmed a 650 million USD investment. After Nvidia licensed its technology and hired its CEO Jonathan Ross, the company is focusing on its own cloud with 13 data centers (TechCrunch).
June 25 – After a meeting between Andy Jassy and Prime Minister Narendra Modi, Amazon announced another 13 billion USD for India through 2030, mainly for AWS data centers in Mumbai and Hyderabad (TechCrunch).

Business and Investment

The end of “tokenmaxxing”: companies start saving

Companies that until recently pushed employees to use AI as much as possible are now dealing with the bills. According to TechCrunch, on June 24 Accenture was trying to stop token spending on trivial tasks such as converting a PDF into a presentation (TechCrunch).

CNBC reported on June 26 that Uber introduced monthly limits starting at 1,500 USD per employee after using up its annual AI budget in four months. Flo Crivello, CEO of startup Lindy, moved all traffic from Claude models to the cheaper DeepSeek (CNBC).

This is a sensitive topic for both big labs. According to CNBC, Anthropic had an annual revenue run rate of 47 billion USD in May, and both companies confidentially filed for IPOs in early June.

June 22 – Google DeepMind is investing 75 million USD in independent film studio A24. Together they plan to develop tools for filmmakers with feedback from artists (TechCrunch).
June 23 – Menlo Ventures raised a 3 billion USD fund, the largest in its 50-year history. According to Bloomberg, its stake in Anthropic is worth around 14 billion USD (TechCrunch).
June 25 – Startup General Intuition, which trains AI agents on video game footage, raised 320 million USD at a 2.3 billion USD valuation. The round was led by Khosla Ventures (TechCrunch).
June 25 – Patronus AI raised 50 million USD for simulated “digital worlds” used to test AI agents. The round was led by Greenfield Partners, and the company’s revenue grew fifteenfold in a year (TechCrunch).

Research

Anthropic Economic Index: when and what people use AI for

On June 26, Anthropic published the sixth Economic Index report, subtitled Cadences. For the first time it tracks Claude usage by day and hour, with data from April 10 to June 10, 2026 (Anthropic).

The share of personal use rises from about 35% on weekdays to just under 50% on weekends.
People look for recipes most often at 6 p.m., and for sleep advice around 5 a.m.
In total, 93% of conversations end with some kind of output, most often an explanation, a document or a how-to guide.
In the developer tool Claude Code, 54% of sessions run on the most powerful model, Opus, compared with just 10% in regular chat.
A survey of about 9,700 people shows that over a third of respondents expect AI to handle most of their work within a year. Yet only about 10% consider losing their own job likely.