GPT-5.6 Sol, Terra, and Luna are now generally available on Amazon Bedrock, with prompt caching and Codex integration. AWS’s announcement comes the same day Anthropic unveils Claude Opus 5 at half the price of its top tier.

Why it matters to you. The question shifts to the cost of a response you can actually use. As we reported on July 22, cost per usable output tells a different story than benchmarks: GPT-5.6 spent $8 where Claude Fable 5 burned $160 on the same task. Now GPT-5.6 on Bedrock has caching active, and Opus 5 positions itself at half price. Competitive pressure shifts entirely to cost per verified result.

For those running agents hundreds of times daily on similar tasks, prompt caching can lower your bill measurably. For those in the Claude ecosystem, Opus 5 narrows the gap. In both cases, provider choice becomes an economic calculation.

If you want to try it. AWS’s post documents how to enable caching via the Responses API on the bedrock-mantle endpoint. Compare costs against your actual workload before migrating.

In detail

GPT-5.6 was on Bedrock since July 14, but under limited access. Today’s news is general availability with two pieces that change the operational math: prompt caching and integration with the Codex coding agent.

Prompt caching works like this: when your agent repeatedly sends the same starting context (a long system prompt, a reference document, conversation history), Bedrock stores it and charges you only the first time. Subsequent calls on that same context cost a fraction of full price. If your workflow repeats the same preamble across hundreds of requests, savings are immediate and measurable.

The Responses API is Bedrock’s interface for OpenAI models, through the bedrock-mantle endpoint. It’s built for agents, with conversation state management and tool calling integrated. For those already familiar with the native OpenAI API, the learning curve is minimal.

Codex integration means those using Bedrock as infrastructure can point the coding agent at GPT-5.6 models hosted there without routing through the direct OpenAI API. This matters for enterprises already holding an AWS contract who want to keep traffic within a single billing and compliance perimeter.

The convergence with Opus 5. The same day, Anthropic announces Claude Opus 5: performance close to Fable 5 at half cost, available as default on Claude Max. The timing isn’t accidental. Both labs are responding to the same pressure: clients no longer choose by benchmark, they choose by cost per usable response. The July 22 piece showed a 20x gap between GPT-5.6 and Fable 5 on the same task. Anthropic closes the gap by lowering price; OpenAI widens the channel by putting GPT-5.6 on Bedrock with caching.

What still needs verification. AWS’s published rates for caching on Bedrock haven’t yet been field-tested by independent sources. Real savings depend on how much context repeats between calls: if every request has different context, caching doesn’t activate and you pay full price. Quotas and scaling limits, documented in the post, vary by region and account.

For those wanting to measure with their own data instead of trusting marketing numbers, the playbook for comparing two models in fifteen minutes is the starting point: same task, same test cases, results table with cost per execution.