GPT-5.6’s new economics make model routing the real product decision

OpenAI says that, from July 30, GPT-5.6 Luna costs 80% less and GPT-5.6 Terra 20% less in its API, while Sol keeps its existing price. At the same time, the company replaced Priority Processing with Fast mode for Sol, claiming up to 2.5 times Standard speed at twice the price and no change in intelligence. That is a concrete change to the operating envelope of the GPT-5.6 family: the lower tiers become cheaper to run broadly, while the flagship gains a paid latency option.

The important comparison is not simply Sol versus a prior flagship. OpenAI positions Sol for consequential reasoning, Terra for balanced everyday production tasks, and Luna for high-volume work. Its suggested coding pattern is to use Sol to resolve uncertainty and create a plan, then use Luna for specified implementation, testing, and evaluation. This turns routing into an engineering decision: teams must measure where added reasoning improves task success enough to justify Sol, rather than treating the family as interchangeable.

Amazon’s Bedrock documentation shows that the family is also available through a common OpenAI-compatible Responses API surface. Sol, Terra, and Luna each support text and image input, text output, a 272K-token context window, tool use, and the same reasoning-effort settings. For an existing Bedrock deployment, the migration can therefore be largely a base-URL and model-ID change, although availability is limited to listed US regions and the specific Sol tier is available in fewer regions than Terra and Luna.

The more operationally significant addition on Bedrock is controllable prompt caching. Amazon says cache reads receive a 90% discount relative to uncached input, while cache writes cost 1.25 times the uncached input rate and cached prefixes remain reusable for at least 30 minutes. That favors agent loops with a stable system prompt, tool schema, or reference corpus and a changing suffix. It also means a migration should include instrumentation: cache write tokens without later cache reads are a cost signal, not a performance win.

The performance and cost comparisons in OpenAI’s announcement are provider claims, including its comparison with Claude Fable 5, not independent benchmarks in these sources. The verified practical facts are narrower: revised Luna and Terra prices, a paid faster Sol mode, and a compatible Bedrock route with caching and regional token-throughput quotas. Technical users should validate quality, cache-hit behavior, latency, and 429 handling against their own workload before changing defaults, particularly where a workflow depends on long multi-step runs.