Today's Headlines
- OpenAI previews its next-generation GPT-5.6 line (Sol/Terra/Luna), limited for now to trusted partners at the US government's request
- Anthropic's Fable 5 and Mythos 5 remain suspended as negotiations with the Trump administration drag on
- A ¥10.5 trillion public-private investment in physical AI, aimed at 2040, shifts the focus "from proof of concept to implementation"
Today, government involvement in frontier models came to the front at two companies at once. OpenAI released its next-generation GPT-5.6 line under a request from the US government, while Anthropic's models remain down under an export-control directive issued on June 12.
Japan, by contrast, turned attention to where computation is applied. The story there was a large public-private investment in "physical AI" for use on factory floors and in logistics. Who is allowed to use a model, and where AI gets implemented — a layer separate from raw performance moved on the same day.
Today's Top Three
OpenAI Previews GPT-5.6, Limiting Access at the US Government's Request
OpenAI previewed its next-generation GPT-5.6 line and said it would keep distribution limited for now, at the request of the US government.
The new line comprises three models: Sol, Terra, and Luna. Sol, the flagship, is described as the company's strongest yet in coding, biology, and cybersecurity, and it includes an "ultra" mode that coordinates multiple subagents. Terra matches the previous GPT-5.5 in performance at roughly half the price for everyday use, and Luna is the lowest-cost option.
The rollout is phased. According to OpenAI, the Trump administration asked that all three models be limited to a small group of "trusted partners" whose participation has been shared with the government, so the preview began only with pre-approved partners. The company says it plans broader availability through ChatGPT, Codex, and its API "in the coming weeks."
OpenAI framed the government's access-vetting process as a temporary measure, stating that it does not believe "this kind of government access process should become the long-term default." API pricing runs $5 input and $30 output per million tokens for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna; in July, Sol is set to run on Cerebras Systems infrastructure at up to 750 tokens per second.
- Previewing GPT-5.6 Sol (OpenAI)
- OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn't be the norm (TechCrunch)
Anthropic's Fable 5 and Mythos 5 Stay Down as Talks With the Trump Administration Drag On
Anthropic's top models, Fable 5 and Mythos 5, remain suspended two weeks after the June 12 export-control directive (follow-up).
To recap: citing national security, the US government ordered access to both models cut off for foreign nationals. Because Anthropic cannot screen users by nationality in real time, it disabled both models for all customers to ensure compliance.
The negotiations since have not reached a quick resolution. According to reporting, the June 26 deadline for Commerce Secretary Lutnick's written justification, sought by Congress, passed without a public response. As of June 25, traffic to Fable 5 and Mythos 5 stood at zero, and the suspension has not been lifted.
The Commerce Department is said to be weighing a "trusted partners" framework under which vetted entities in close US-allied countries could regain access to controlled models. That framework has not yet been implemented. Anthropic opposes pulling a deployed commercial model over a narrow vulnerability concern, arguing that applying the same standard would halt the rollout of frontier models across the board.
Japan Signals ¥10.5 Trillion in Public-Private Investment for Physical AI, "From Proof of Concept to Implementation"
A public-private investment of about ¥10.5 trillion was presented for "physical AI" used on factory floors and in logistics.
The figure was presented at an industry event in late June 2026 and is framed as a scale aimed at 2040. At the center of the discussion is the shift from the proof-of-concept stage to actual implementation on the ground.
While debate over generative AI tends to cluster around text and image generation, the focus here fell on applying AI to physical settings such as robots and equipment. Beyond the absolute size of the investment, it signals a change of stage — the field beginning to move from demonstration to deployment.
Other Developments
Models & APIs
- Mistral AI unveiled Mistral OCR 4, an OCR model supporting 170 languages. It classifies regions such as titles, tables, equations, and signatures, locates them with bounding boxes, and adds inline confidence scores. In the company's comparison across more than 12 languages and over 600 real documents, it beat other OCR systems with an average win rate of 72%. Mistral OCR 4 announced with 170-language support and high-precision region classification (PC Watch)
- Liquid AI released LFM2.5-230M as a free, open-weights model. With 230 million parameters, it reaches 213 tokens per second on a Galaxy S25 Ultra under 4-bit quantization. The design is geared to lightweight agent tasks such as tool calling and data extraction. LFM2.5-230M, a model that runs at 213 tok/s even on a phone, released for free (PC Watch)
- The era of "OpenAI models on Azure only" has ended, with GPT-5.5 and Codex now available on AWS. The piece discusses what it means to have more choice of platform. The "OpenAI is Azure-only" era ends — what are the benefits of using GPT-5.5 and Codex on AWS (ITmedia AI+)
Products & Features
- Inside OpenAI, the coding assistant Codex is reportedly used widely in place of ChatGPT and has spread into legal and finance work. A tool built for developers is being adopted by departments outside engineering. Codex, which took ChatGPT's place, has spread into OpenAI's legal and finance teams (PC Watch)
- Google described how to use Gemini to ease jet lag. It suggests when to get light exposure and how to adjust sleep around the destination's time zone. Here's how Gemini can help you avoid jetlag (Google Blog)
Research
- An analysis of the performance gap between open-weights LLMs and closed, commercial LLMs was published. It examines how far the gap has narrowed and where differences remain. The gap between open weights LLMs and closed source LLMs (Hacker News)
Policy
- The New York Times alleged that Microsoft built a supercomputer to help OpenAI infringe copyrights. It is a new point in the ongoing copyright litigation between news organizations and AI companies. NYT slams Microsoft for building copyright-infringing supercomputer for OpenAI (Ars Technica)
- Japan's Ministry of Defense released an analysis on countering "cognitive warfare" that uses AI fakes and fake accounts. The material draws on information-manipulation tactics seen threatening Ukraine. How Japan's Ministry of Defense takes on "cognitive warfare" — releasing an analysis of responses to AI fakes and fake accounts threatening Ukraine (ITmedia)
Business
- OpenAI poached Uber's India chief to lead India, its largest market outside the US. The move builds out its leadership in a growth market. OpenAI poaches Uber India chief to lead its biggest market outside the US (TechCrunch)
- Gartner published a forecast on corporate AI spending. The question is whether the wait-and-see phase is ending and giving way to full-scale investment. Is corporate AI's "wait and see" about to end? Gartner forecasts where full-scale investment is headed (ITmedia AI+)
- AWS announced that token processing on Amazon Bedrock in the first quarter of 2026 alone exceeded its entire cumulative total to date. It frames the move as building "trustworthy infrastructure" for AI. Amazon Bedrock token volume in Q1 2026 alone surpasses all prior cumulative total — AWS aims for "trustworthy infrastructure" for AI (ITmedia AI+)
Source: Selected by the editorial desk from the AI news inbox (collected June 27, 2026; 29 items, 3 primary and 26 secondary).