Anthropic Export Ban Eased, GPT-5.6 Previews, and Gemini Deep Think

The artificial intelligence industry has experienced a massive tectonic shift over the past 48 hours. Between unprecedented government interventions and a fierce battle for benchmark supremacy, the landscape of frontier AI models has fundamentally changed this weekend. Here are the most critical updates for Saturday, June 27, 2026.

The Standoff Ends: Anthropic's Mythos 5 Gets a Cybersecurity Exemption

The highest-stakes story of the week revolves around Anthropic and the U.S. government. Two weeks ago, the Commerce Department enacted a sweeping, emergency export freeze on Anthropic, forcing the company to pull its most advanced models—Mythos 5 and the public-facing Fable 5—offline globally due to national security concerns regarding their elite hacking capabilities.

Today, the standoff broke. U.S. Commerce Secretary Howard Lutnick announced a partial reversal of the ban. Anthropic has been authorized to redeploy Mythos 5, but strictly within a tightly controlled "safe harbor."

The model is now available exclusively to a highly vetted group of roughly 100 American cybersecurity firms and critical infrastructure providers to bolster domestic defense. However, the concession comes with strict caveats: the government’s total ban on Anthropic’s general-public model, Fable 5, remains firmly in place pending further safety reviews by the Trump administration.

OpenAI Staggers the Release of GPT-5.6

Operating under the shadow of the same regulatory pressure, OpenAI has taken a highly cautious approach to its next major release. Following a direct request from the administration, CEO Sam Altman announced a highly restricted, limited preview of the GPT-5.6 family.

Rather than a massive public launch, access to GPT-5.6 is currently being validated on a strict, client-by-client basis. The move echoes the intense safety scrutiny placed on Anthropic and highlights a new reality for the industry: the scarcest input for frontier intelligence is no longer compute power, but government clearance.

Google Drops the Benchmark King: Gemini 2.5 Pro with Deep Think

While regulatory battles consumed Anthropic and OpenAI, Google seized the opportunity to redefine model capabilities. Google quietly shipped Gemini 2.5 Pro, and early benchmark results suggest it may be the most capable model currently available to the public.

1.Massive Ingestion:2 Million Token Context.

Gemini 2.5 Pro doubles the effective working context of its predecessors. It can ingest and reason over entire codebases, full-length books, or months of conversation history in a single session.

2.Extended Inference:Deep Think Mode.

Google introduced "Deep Think," an extended reasoning mode that allows the model to spend significantly more compute on complex problems, explore multiple reasoning paths, and verify its logic before responding.

3.Shattering Records:Benchmark Dominance.

The model has set new highs across the board, scoring an unprecedented 89.8% on the MMLU-Pro and a near-perfect 97.2% on the MATH-500 benchmark.

The Strategic Shift: Deep Think mode is Google's direct answer to the complex, agentic reasoning models previously dominated by Anthropic's Claude Code. While Deep Think usage costs roughly 4x the standard rate and takes longer to respond, Google claims it improves accuracy on complex multi-step problems by up to 25%.

As June comes to a close, the AI market is split into two distinct arenas: the heavily regulated, government-vetted security models, and the fierce commercial race to deliver the deepest reasoning capabilities to developers.