AWS says GPT-6.1 Sol supports prompt caching—and also says it does not

AWS’s GPT-6.1 Sol documentation contains a consequential contradiction: the launch notice says the Bedrock model supports explicit prompt caching, while the model card lists explicit prompt caching under Not Supported and repeats that it is unavailable in the pricing section. AWS launch notice AWS model card

Two first-party pages, two answers

AWS’s September 29 announcement calls GPT-6.1 Sol generally available and says it “supports explicit prompt caching on Amazon Bedrock,” presenting caching as useful for agentic workloads that reuse context. The same announcement directs customers to Bedrock documentation for features, APIs, endpoints, regions, inference profiles and pricing. AWS launch notice

The linked OpenAI catalog now includes GPT-6.1 Sol and links a dedicated model card. That card says the model has a one-million-token context window and a 131,072-token maximum output. It supports the Responses and Chat Completions APIs on both bedrock-runtime and bedrock-mantle; Converse and Invoke are supported only on bedrock-runtime. AWS OpenAI catalog AWS model card

But its feature table marks explicit prompt caching unsupported. The pricing section is even more direct: cache-write and cache-read price dimensions are shown, yet AWS says those dimensions do not change the model’s feature support and that explicit prompt caching is not supported. AWS model card

The safe deployment reading

Until AWS reconciles the pages, teams should treat caching as unverified and test the exact endpoint they plan to use. The card identifies openai.gpt-6.1-sol for Mantle in us-east-1 and us.openai.gpt-6.1-sol for Bedrock Runtime’s US geographic inference profile. It says no global inference profile is offered at launch and only the Standard service tier is supported. AWS model card

For analytics agents, this is not a cosmetic discrepancy. Reused schema descriptions, semantic definitions and policy instructions can make caching assumptions part of latency and cost estimates. A migration test should record whether cache controls are accepted, whether usage fields show cache reads or writes, and which endpoint and model identifier produced the result.

Those measurements should stay separate from AWS’s vendor-relayed claim that GPT-6.1 Sol approaches GPT-6 Astra on demanding evaluations at roughly one-fifth of the cost; the launch notice does not publish the evaluation breakdown. AWS launch notice

The practical conclusion is narrow: GPT-6.1 Sol is available, but explicit prompt caching should not enter a production migration plan until AWS makes its launch notice and model card agree.