Faster, More Stable, Token‑Efficient: GPT‑5.6 Arrives with Three Fully Available Models

Model lineup and availability – OpenAI released three GPT‑5.6 models (Sol, Terra, Luna) on Amazon Bedrock (US‑East‑1, US‑East‑2, US‑West‑2). Sol is the flagship, Terra targets balanced production workloads, and Luna is optimized for fast, low‑cost inference.

Performance breakthroughs – Sol achieved an 80 point record on the Artificial Analysis Coding Agent Index, beating the runner‑up by 2.8 points while using less than half the output tokens, less than half the latency, and about one‑third the cost. In the ExploitBench security benchmark, Sol scored 73.5 % versus 47.9 % for GPT‑5.5. In the Agents’ Last Exam (55‑domain workflow), Sol posted 53.6 points, 13.1 points above the next best model. Even at moderate reasoning intensity, Sol’s estimated cost lead was roughly 11.4 points, about a quarter of the cost of comparable models, and it offers a “max reasoning effort” option for complex tasks.

Terra and Luna trade‑offs – Terra delivers performance superior to GPT‑5.5 at a lower price, making it suitable for code generation, content workflows, and structured‑data extraction where strong reasoning is needed without flagship pricing. Luna focuses on high‑throughput, latency‑sensitive scenarios such as classification and summarisation, offering the most economical per‑token cost.

Cost‑effective prompt caching – Bedrock’s new cache‑breakpoint feature lets users mark reusable prompt sections; subsequent calls reuse the cached context, incurring only 10 % of the normal token charge and retaining the cache for at least 30 minutes, which can cover an entire multi‑step Agent run.

Enterprise‑grade security – The inference engine runs with zero‑operator access (ZOA), preventing any Amazon staff from seeing prompts or outputs. Each call executes inside the user’s IAM‑controlled VPC, is logged to CloudTrail, and respects data‑boundary policies. A hardware‑level security model, real‑time misuse classifiers, and extensive red‑team testing protect against abuse.

Integration and usage – Developers can invoke the models via the Bedrock Responses API or the console. The article also mentions the concurrent launch of ChatGPT Work (multi‑step Agent) and Codex (developer‑focused coding Agent) on Bedrock, expanding the ecosystem for autonomous agents.