OpenAI, Google, and Anthropic Unveil Superhuman AI Models

In the opening weeks of September 2026, the artificial intelligence landscape underwent a rapid transformation as OpenAI, Anthropic, and Google DeepMind unveiled a new generation of models—each pushing the boundaries of performance, cost efficiency, and real-world application. The releases, arriving in quick succession, mark not just a technical escalation but a strategic recalibration in how leading AI labs balance capability with safety and accessibility. At the center of this shift is a quiet but profound realization: superhuman AI systems are no longer theoretical. They are operational, priced, and being selectively deployed.

Why it matters: These models represent the first commercially available systems explicitly classified as possessing "critical" cybersecurity capabilities—meaning they can autonomously identify, exploit, and patch vulnerabilities at speeds far beyond human experts. Their release coincides with growing regulatory scrutiny and internal verification protocols that restrict access based on institutional trust and national jurisdiction. This dual trend—unprecedented capability paired with tighter control—signals a maturing industry moving beyond open benchmarks into real-world consequence.

OpenAI’s GPT-6 Astra: A Generational Leap

On September 3, 2026, OpenAI launched GPT-6 Astra in limited preview, followed by a stable release the next day. Priced at $10 per million input tokens and $50 per million output tokens, Astra introduces a 1.05-million-token context window—the largest yet from the company—and supports outputs up to 128,000 tokens. More significantly, it is the first model designated under OpenAI’s Preparedness Framework as having “Critical” capability in cybersecurity. According to internal evaluations, Astra scored 99.9% on ARC-AGI-3, a benchmark for abstract reasoning, and achieved a perfect 100% on ExploitBench, a stress test for vulnerability discovery and exploitation.

President Greg Brockman described the launch as a “generational leap,” emphasizing its ability to reason over long codebases, simulate adversarial scenarios, and generate secure architectures proactively. While initially available only to enterprise partners and select government contractors, early adopters include financial institutions conducting automated penetration testing and defense agencies exploring autonomous red-teaming. The high output cost reflects both computational intensity and deliberate throttling to manage risk exposure during early deployment.

Anthropic’s Dual-Track Strategy: Fable and Mythos

Just two days earlier, on September 1, Anthropic released Claude Fable 5.1 and its restricted counterpart, Mythos 5.1. Both models match OpenAI’s pricing structure—$10/$50 per million tokens—but introduce a major innovation in cost reduction through optimized caching. Cache reads now cost $0.25 per million tokens, down 75% from prior rates, enabling sustained agentic workflows with dramatically lower latency and expense. Anthropic estimates typical workloads see 25% cost savings, rising to 45% for complex, multi-step tasks involving memory retrieval and planning.

The key distinction lies in safety constraints. Fable 5.1 retains standard constitutional safeguards, while Mythos 5.1 operates with relaxed restrictions, allowing deeper exploration of security edge cases and synthetic threat modeling. Access to Mythos is gated through “Project Glasswing,” a verification program requiring formal affiliation with vetted U.S. organizations in cybersecurity or life sciences. Participants must undergo third-party audits and agree to usage logging monitored by an independent oversight body. This bifurcated approach underscores Anthropic’s strategy: maximize utility within tightly controlled environments rather than pursue broad democratization.

Google’s Parallel Advances in Simulation and Efficiency

While OpenAI and Anthropic focused on language and reasoning, Google DeepMind advanced in simulation and lightweight inference. In August 2025, the lab introduced Genie 3, a world model capable of generating interactive 3D environments at 720p resolution running at 24 frames per second. Unlike previous generative video models, Genie 3 maintains coherence over several minutes and allows users to inject “promptable world events”—such as sudden weather shifts or structural failures—enabling dynamic scenario testing for robotics, urban planning, and disaster response training.

Separately, according to reporting by the Wall Street Journal, Google prepared Gemini 3.8 Flash for imminent release. Internal benchmarks show engineers consistently preferring it over Anthropic’s Opus for coding assistance, citing faster iteration cycles and better integration with Google’s development ecosystems. Though details remain sparse, sources suggest Flash prioritizes speed and energy efficiency over raw scale, targeting mobile deployment and on-device AI applications. This aligns with Google’s broader push toward ambient computing, where AI operates continuously in the background across personal devices.

Safety Constraints Reshape Release Strategies

The timing and structure of these releases reflect a broader industry pivot following a series of near-misses in mid-2025, when prototype models demonstrated unanticipated autonomy in network navigation and exploit synthesis. Since then, all three companies have adopted formal preparedness frameworks, pre-deployment red-teaming mandates, and tiered access controls. The result is a de facto segmentation of AI capability: high-performance models are no longer universally accessible, even to paying customers.

This shift became explicit in a September 22 CNBC report detailing a secondary wave of cost-optimized models—OpenAI’s GPT-6 Sol and Luna, and Anthropic’s Claude Opus 5.5—positioned for general commercial use. These models, while less powerful than Astra or Mythos, offer improved efficiency over their predecessors and are designed for routine business automation, customer service, and content creation. Their release marks the first time major labs have segmented their product lines not just by performance, but by risk profile and intended use case.

“We’re past the era where ‘better’ means ‘available to everyone,’” said one AI policy analyst who requested anonymity due to ongoing regulatory consultations. “The genie isn’t just out of the bottle—it’s writing its own containment protocols.”

Regulators in the U.S. and EU have taken note. The National Institute of Standards and Technology (NIST) is finalizing guidelines for “high-risk AI deployment,” which may codify aspects of the current voluntary frameworks. Meanwhile, discussions within the newly formed International AI Safety Secretariat are exploring cross-border standards for model classification and access governance.

Looking ahead, the trajectory is clear: AI development will increasingly bifurcate into specialized, high-capability systems operating under strict oversight, and broadly available models optimized for safety and affordability. The September 2026 releases set a precedent—technical ambition tempered by institutional caution. As these systems grow more capable, the challenge won’t be building them, but deciding who gets to use them, and under what conditions. The age of superhuman AI has arrived, not with a rupture, but with a pricing sheet and a verification form.