Tech • IA • Crypto
GPT 5.6 Just Blew Up the AI World
9.4/10TL;DR
OpenAI’s GPT 5.6 launches under restricted access amid U.S. government oversight, signaling a shift toward treating advanced AI as strategic infrastructure while the company simultaneously pushes into custom chip development to cut costs.
KEY POINTS
Restricted launch under government pressure
OpenAI released GPT 5.6 to a limited group of roughly 20 partners, with broader access delayed. The U.S. government requested restrictions, and partner identities were reportedly shared with authorities. This marks a departure from typical AI rollouts and suggests increasing state involvement in frontier model deployment.
AI increasingly treated as strategic technology
The limited release reflects growing concern that advanced AI systems have national security implications, particularly in cybersecurity and biological research. Policymakers are moving toward pre-release review frameworks, raising concerns about an emerging de facto licensing system for powerful models.
Three-model lineup with advanced capabilities
The new family includes GPT 5.6 Soul (flagship), Terra, and Luna. Soul introduces multi-agent “ultra mode” and deeper reasoning capabilities, enabling coordinated task execution across sub-agents. This architecture improves performance in coding, automation, and technical workflows, but increases compute demands.
Benchmark gains and efficiency improvements
OpenAI reports state-of-the-art performance on Terminal Bench 2.1 and improved results over GPT 5.5 with fewer tokens. External comparisons indicate competitiveness with top Anthropic models while using roughly one-third fewer output tokens, lowering operational costs for large-scale deployments.
High-risk classification and safety controls
All three models are classified as high capability in cybersecurity and bio/chemical domains. OpenAI states the models can identify vulnerabilities but do not reliably execute full exploit chains. Safeguards include built-in refusal behaviors, real-time misuse detection, and account-level monitoring.
Tensions over regulatory precedent
A recent executive order requires companies to submit advanced models for government review up to 30 days before release. Critics argue this creates an “involuntary licensing regime”, potentially slowing innovation and introducing opaque approval processes that could disadvantage U.S. firms globally.
Massive safety testing effort
OpenAI conducted over 700,000 A100 GPU-hours of automated red teaming, alongside human and third-party evaluations. Additional testing will continue during the preview phase, with updated disclosures expected before general availability.
Pricing and infrastructure considerations
Pricing starts at $5 per million input tokens and $30 per million output tokens for Soul, with cheaper tiers for Terra and Luna. New caching mechanisms reduce repeated prompt costs by up to 90%, improving efficiency for agent-based systems.
Custom AI chip “Jalapeno” unveiled
OpenAI introduced its first custom inference chip, Jalapeno, developed with Broadcom. Early tests show around 50% cost savings compared to GPUs and improved performance per watt. The chip targets inference workloads, which now dominate operational expenses.
Shift toward full-stack AI control
Jalapeno is part of a broader strategy to reduce reliance on NVIDIA and optimize cost at scale. OpenAI joins Google, Amazon, Microsoft, and Meta in building custom silicon, as ASIC-based systems are projected to capture 27.8% of the AI server market.
AI designing its own hardware
OpenAI used its own models to accelerate chip design, shortening development to nine months. This creates a feedback loop where AI improves hardware, which in turn enhances AI performance and reduces costs.
Long-term expansion and geopolitical competition
Large-scale deployment of Jalapeno is planned between 2026 and 2028, with ambitions for gigawatt-scale infrastructure. Meanwhile, global competition is intensifying, with China also investing heavily in custom AI chips amid export restrictions.
CONCLUSION
The GPT 5.6 rollout highlights a pivotal shift where cutting-edge AI is shaped as much by government oversight as by technological progress, while OpenAI’s move into custom hardware signals an escalating race for control over the full AI stack.