OpenAI GPT-5.6 Sol, Terra, and Luna are now generally available on Amazon Bedrock
AWS Tanvi Girinath ● Covered by 5 sources
OpenAI's GPT-5.6 family — Sol, Terra, and Luna — just went live on Amazon Bedrock. It's tuned for AI agents that run hundreds of steps deep, with AWS security baked in.
Based on reporting by AWS, Tanvi Girinath — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI's newest model generation has landed on Amazon Bedrock, and it comes with a naming scheme that ditches the usual mini/pro suffixes for something more permanent. Sol, Terra, and Luna aren't just size variants — OpenAI is calling them durable capability tiers, meaning each can get upgraded on its own schedule going forward. Sol is the flagship reasoner, Terra the everyday workhorse, and Luna the cheap-and-fast option for high-volume tasks like classification or routing.
The numbers OpenAI is touting for Sol are aggressive. It scores 80 on the Artificial Analysis Coding Agent Index, 2.8 points ahead of the next-best model, while burning less than half the output tokens and roughly a third less cost. On ExploitBench, a cybersecurity benchmark, Sol hits 73.5% versus 47.9% for GPT-5.5 at similar token budgets — a jump that matters given how much attention vulnerability research and red-teaming are getting right now. On Agents' Last Exam, a test spanning 55 professional fields, Sol posts 53.6, beating its nearest competitor by 13.1 points, and it still leads by double digits even when dialed down to medium reasoning effort at a quarter of the cost.
What's arguably more interesting than the benchmarks is the infrastructure AWS built underneath. Agent workloads are bursty by nature — one user prompt can spawn hundreds of model calls in a short window — so Bedrock's new inference engine pools capacity across customers while keeping each one's throughput isolated. Paired with that is prompt caching using explicit cache breakpoints: mark the reusable chunk of a prompt (system instructions, tool definitions, reference files) and Bedrock reuses that processed context for 30-plus minutes, billing cached tokens at a 90% discount. For agents that repeat the same scaffolding across dozens of calls, that's real money saved, not just a marketing bullet.
Security gets equal billing here, which makes sense given who's buying this — genomics labs, coding-agent shops, security researchers, the kind of customers who can't have their prompts sitting somewhere an engineer could peek at them. Bedrock's zero-operator-access model is enforced at the chip level, every call runs under a customer's own IAM policies and VPC, and everything gets logged to CloudTrail. OpenAI layers its own safety stack on top — refusals, real-time misuse classifiers, account-level enforcement — with flagged traffic retained for 30 days for abuse review.
For now, availability is narrow: Sol only runs in US East (N. Virginia and Ohio), while Terra and Luna add US West (Oregon). Alongside the model launch, OpenAI also shipped a new ChatGPT desktop agent called Work, aimed at multi-step tasks like building docs and slides, sitting next to Codex for developers — both of which can be pointed at GPT-5.6 through Bedrock's Responses API.
My take — AI-written commentary, not fact-checked reporting
The tiered naming trick is smart marketing dressed up as architecture — it lets OpenAI ship incremental upgrades without the awkward 'is 5.6 better than 5.5-turbo' confusion that plagued earlier releases. But the real story is AWS quietly becoming the compliance-friendly backdoor into frontier models for regulated industries that would never touch OpenAI's own API directly. Watch this pattern spread to Anthropic and Google too — the cloud vendors, not the model labs, may end up owning enterprise AI distribution.
Read more about this at: AWS