AI Product Updates Daily — June 15, 2026

Three days after Commerce Secretary Howard Lutnick's export control directive took Fable 5 and Mythos 5 offline, senior Anthropic staff were in Washington trying to negotiate their way back on. The models remain down. Meanwhile, OpenAI is fighting a subpoena coalition from 14 US state attorneys general, Google dropped a new open-weight multimodal model small enough to run on a laptop, Canva's Magic Layers went live inside Gemini and ChatGPT, and the G7 summit opened in Évian with all three frontier AI CEOs at the table. Here is everything that moved today.

Anthropic: Fable 5 still offline, negotiations ongoing

Fable 5 and Mythos 5 remain suspended on day three of the export control shutdown.1 Anthropic published its disagreement publicly but has not announced a resolution timeline.

The negotiation now reportedly centers on whether a tiered access structure would satisfy the government's national security requirements. Three options are under discussion: a complete block for foreign nationals, a geofenced version routing high-risk query categories to Opus 4.8's guardrails, or enhanced monitoring and logging requirements giving the government near-real-time audit capability.2 The Commerce Department's original directive required suspension of access for all foreign nationals everywhere; the gap between that and Anthropic's global commercial footprint is what makes a tiered solution attractive to both sides.

The case is drawing attention from AI policy lawyers. If the Commerce Department can suspend commercial distribution of an AI model by citing a claimed jailbreak — without providing technical details to the developer and without standard export control notice-and-comment procedures — it establishes a regulatory mechanism for AI capabilities with no formal statutory basis.3 Whether Congress treats the Fable 5 action as a precedent to codify or constrain is one of the biggest AI policy questions of the second half of 2026.

Mythos Preview context: Ben Thompson's June 15 Stratechery piece offers useful framing. Anthropic announced Mythos Preview two months ago as too dangerous to make publicly available, then released Fable 5 (Mythos with guardrails), and now has the guardrails themselves challenged by a jailbreak report. Thompson notes that Fable reads subjectively like a new generation of base model — "it made other models, including GPT 5.5 and Opus 4.8, feel small and dumb" — which may explain why the government was alarmed rather than placated by the safety framing.

All other Anthropic models remain available. The Claude Code /fork command and updated CLI shipped June 13 and are unaffected.

OpenAI: state AG subpoena, GPT-5.6 watch

Enterprise AI adoption snapshot — Logicalis 2026 CIO Report

Survey of 1,000+ CIOs globally, published March 3, 2026

Increased AI investment (past year)

Say adoption is moving too fast

"Learning as we go" approach

No contingency plan if AI provider goes offline

통계 카드를 불러오는 중…

A coalition of US state attorneys general, led by New York, issued OpenAI a sweeping subpoena on Friday, June 13, 2026.4 The subpoena covers advertising practices, user engagement mechanics, handling of consumer and health data, and policies toward minors and older adults. OpenAI said it takes the concerns "seriously" and has built protections into its product, but has not confirmed how many states are participating.

The subpoena lands while OpenAI is preparing its IPO, which analysts expect will value the company at roughly $1 trillion. Legal pressure has been building: a Canadian woman sued OpenAI in June alleging ChatGPT contributed to her daughter's suicide, and the Florida AG separately sued OpenAI and CEO Sam Altman after two shootings in which alleged attackers reportedly used the chatbot to plan the crimes.

On the model side: GPT-5.6 has a 94% probability of releasing by July 31 per Polymarket prediction markets,5 with social media reports placing a tentative date of June 23. A Sohu report (June 15) suggests the Fable 5 shutdown may push OpenAI to delay the release, though this is unconfirmed.6 GPT-5.2 Instant, Thinking, and Pro were fully removed from ChatGPT as of June 12.7

Google: Gemma 4 12B open release

Gemma 4 12B at a glance

Google DeepMind open-weight multimodal model, released early June 2026

Parameters

12B

Min RAM to run locally

16 GB

Modalities

Text + Image + Audio

License

Open-weight

통계 카드를 불러오는 중…

Google DeepMind released Gemma 4 12B in early June 2026 — a free, open-weight multimodal model that handles text, images, and audio, and runs locally on a 16GB laptop.8 The model is available on Hugging Face with a gguf quantization suitable for consumer hardware.

Performance claims: Gemma 4 12B is reported to match 26B-class models on standard benchmarks.9 MATLAB published a walkthrough on June 15 demonstrating agentic AI workflows running on a 16GB MacBook Pro with Gemma 4 in LM Studio.10 For developer teams evaluating open-weight alternatives after the Fable 5 shutdown, this is the most capable new option that requires no API keys or per-token cost.

Canva Magic Layers inside Gemini turns generated images into fully editable layered designs. — Canva Magic Layers live inside Gemini and ChatGPT — AI-generated images become editable Canva layers 11

Canva: Magic Layers inside Gemini and ChatGPT

Canva's Magic Layers feature went live globally inside Google Gemini and OpenAI's ChatGPT on June 8, 2026.11 The integration lets users generate images with Gemini's Nano Banana models or ChatGPT's image generation, then immediately pass them through Magic Layers to unlock every element as a separate, editable layer inside Canva. The old generate → download → re-upload cycle is replaced by a single @Canva prompt.

Magic Layers was used 9 million times in its first four weeks after the March 2026 standalone launch. The Canva Connected App now works inside Gemini, ChatGPT, Anthropic's Claude, and Microsoft Copilot, all powered by Canva's MCP Server. The June 14 confirmation that the Gemini web version of the connected app is now visible to users (spotted by TestingCatalog) is the latest milestone in the rollout.12

Meta: Edits app gets desktop version and AI production assistant

Meta previewed two updates to its Edits video editing app on June 14:13 a desktop version (enabling Edits to slot into professional post-production workflows) and an AI-powered production assistant embedded inside the editing interface. Details on the production assistant's specific capabilities were not fully disclosed, but the announcement was previewed at an internal Meta event. Availability timelines have not been published.

G7 Évian: first summit with all three AI CEOs present

The 52nd G7 Summit opened June 15 in Évian-les-Bains, France, running through June 17.14 Sam Altman (OpenAI), Demis Hassabis (Google DeepMind), and Dario Amodei (Anthropic) are all confirmed attendees — the first time all three have appeared simultaneously at a G7. French President Macron explicitly invited them as part of a strategy to include private-sector AI leaders in governance discussions.

The structural tension the summit will not resolve: the US wants voluntary governance commitments that preserve American competitive advantage; the EU wants binding data center energy disclosure, AI incident reporting timelines, and coordination on foundation model export controls.

What the US is blocking specifically:

Proposed provision	US position
Mandatory data center energy/water disclosure	Rejected — would disadvantage US hyperscalers vs. China
AI incident reporting with mandatory timelines	Domestic regulatory matter only
Export control coordination on foundation models	Resisting sharing enforcement methodology with G7 partners
Common compute thresholds for frontier model definitions	Wants to maintain its own standards (EO-based)

The expected communiqué will restate OECD AI Principles, use non-binding language, and include a separate EU member-state declaration (not full G7 consensus) pushing for data center disclosure.

For developers building AI applications in the EU: the summit changes no immediate rules, but confirms that EU AI Act enforcement proceeds on schedule with high-risk AI system requirements applying from August 2026.14

OpenAI Sora: API deprecation September 24

A context note worth tracking: OpenAI shut down the Sora web and app experiences on April 26, 2026, with the API scheduled to close on September 24, 2026.2 Any teams still using the Sora API have roughly 100 days to migrate. Google Veo 3.1, Runway Gen-4.5, and Kling 2.5 are the current migration targets for video generation workflows.

Quick dispatches

HHS AERO program: The Department of Health and Human Services announced on May 21, 2026 that it is using ChatGPT to perform rolling analysis of at least five years of Medicaid audit reports from all 50 states, targeting between $100 billion and $200 billion in estimated annual fraud.2 Critics have noted AI misread risks on complex financial documents and that early enforcement patterns have disproportionately flagged Democratic-administered states.
NAVER + NVIDIA gigawatt AI factory: NAVER and NVIDIA announced on June 7, 2026 that NAVER will expand sovereign AI infrastructure at its GAK Sejong data center in South Korea, starting at 55 MW in H1 2027 and scaling to gigawatt capacity. The plan includes NAVER's next-generation HyperCLOVA X models and a "Seoul World Model" combining NVIDIA Cosmos with NAVER's street-view spatial data.15
Meta $27B Nebius compute deal: Meta signed a five-year, $27 billion infrastructure agreement with Nebius, including one of the first large-scale deployments of NVIDIA's Vera Rubin platform (projected 3.3× improvement over Blackwell for large-scale training). The deal gives Meta GDPR-compliant European compute capacity.2
Logicalis 2026 CIO Report: Of 1,000+ CIOs surveyed, 94% report increased AI investment, 51% say adoption is moving too fast, 89% describe their approach as "learning as we go," and 16% have no contingency plan if a key AI provider goes offline. The Fable 5 shutdown made that last number concrete.16
EU AI Act enforcement confirmed August 2026: High-risk AI system requirements under the EU AI Act begin applying in August 2026. The Évian summit produces no language to delay this. Teams serving EU users should review Annex III of the Act for high-risk system classification.
Coursera Admin AI — Conversation Analytics: Coursera launched Conversation Analytics with Admin AI on June 15, a feature for learning leaders that turns learner activity data into decision signals without requiring manual data analysis.17