GPT-5.6 Sol reaches Amazon Bedrock, with regional and retention limits

Amazon’s July 13 product note lists GPT-5.6 Sol, Terra, and Luna as available through Amazon Bedrock’s Responses API. Amazon characterizes Sol as the flagship reasoning tier, Terra as the balanced tier, and Luna as the lower-cost tier. For teams already standardizing on Bedrock, the change is principally a new hosting and procurement path for the GPT-5.6 family rather than a different model family.

OpenAI had described the same three-tier GPT-5.6 family in its July 9 announcement, placing Sol at the high-capability end and Terra and Luna below it on cost and speed. That sequence suggests the Bedrock announcement is a deployment expansion following OpenAI’s broader rollout, not evidence of a separate Sol revision. Engineers should therefore keep model and prompt evaluations tied to the exact platform endpoint they intend to operate, rather than assuming a marketplace listing changes the underlying capability claims.

The OpenAI model documentation defines Sol as a text-and-image-input, text-output model with a 1,050,000-token context window and a 128,000-token output ceiling. It supports the Responses and Chat Completions APIs, structured outputs, function calling, web search, file search, code interpreter, hosted shell, computer use, MCP, and tool search. The same documentation also says fine-tuning, audio, and video are unsupported, which makes Sol a poor fit when those capabilities are hard requirements rather than optional parts of an agent stack.

Region is the most immediate operational constraint on the Bedrock route. Amazon says Sol is available only in US East (N. Virginia) and US East (Ohio), whereas Terra and Luna are also available in US West (Oregon). Amazon also describes its in-region inference design as keeping requests in the region selected by the customer. A workload pinned to Oregon can therefore choose the smaller GPT-5.6 tiers locally, but cannot select Sol there under the announced availability.

The price comparison is also more nuanced than a single per-token number. OpenAI lists Sol at $5 per million input tokens and $30 per million output tokens, with cached input at $0.50 per million tokens; prompts above 272,000 input tokens carry higher input and output multipliers, and cache writes cost 1.25 times the uncached input rate. Amazon says Bedrock exposes explicit cache breakpoints and a 90 percent cached-input discount for reusable context, so repeated agent instructions and tool definitions may benefit, but teams should test their own context shape before treating the published discount as an application-level saving.

Data governance and capacity remain separate questions from model availability. Amazon says Bedrock applies IAM, VPC, and CloudTrail controls and uses a zero-operator-access design, while also stating that classifier-flagged traffic may be retained for automated abuse detection for up to 30 days. Separately, Axios reported that OpenAI chief executive Sam Altman warned of possible near-term service hiccups amid demand for GPT-5.6 Sol. Neither statement proves a problem with Bedrock, but together they are a reason to validate retention requirements, regional failover, and workload behavior before committing a production agent to the new route.