GPT-5.6 Sol gets Fast mode as OpenAI cuts prices below it

OpenAI's July 30 GPT-5.6 update cuts API prices for Luna by 80 percent and Terra by 20 percent, while leaving Sol token pricing unchanged. For Sol, the new item is Fast mode: it replaces Priority Processing, promises up to 2.5 times the speed of Standard processing, costs twice the Standard price, and is described by OpenAI as leaving intelligence unchanged. Existing API requests tagged `priority` continue to work through the new mode.

The split clarifies where user controls lie. Terra costs $2 per million input tokens and $12 per million output tokens, while Luna costs $0.20 and $1.20; the product page says ChatGPT and Codex subscription prices and quota budgets stay unchanged even though Terra and Luna consume fewer credits. Sol users are therefore choosing a latency premium rather than receiving a lower base price, while teams that can decompose work have a stronger incentive to push routine steps down to Terra or Luna.

That routing decision now interacts with cache design. AWS documents explicit cache breakpoints for GPT-5.6 Sol, Terra, and Luna on the Bedrock Responses API, whereas it describes GPT-5.5 and earlier as using automatic prompt caching. A GPT-5.6 breakpoint needs a prefix of at least 1,024 tokens and has a 30-minute minimum lifetime; the feature is meant for stable system instructions, tool definitions, or reference files followed by changing user input.

The economics are conditional, not automatic. AWS bills cache writes at 1.25 times the uncached input rate but discounts cache reads by 90 percent and excludes cached reads from input-token rate limits. Its guidance says net input cost falls once cache reads are roughly 20 percent of tokens moving through the cache. A workload that changes the prefix frequently can therefore pay write overhead without enough reuse to recover it, so cache-read and cache-write telemetry should be part of any migration decision.

For technical users, the practical experiment is small but concrete: keep repeated material before an explicit breakpoint, place variable request data after it, record the cache usage fields, and compare standard Sol against Fast mode under the same latency and quality target. AWS says explicit mode makes the boundary deterministic, while implicit mode can vary the number of writes before reads begin. The evidence supports a new control plane for cost and latency, not a blanket promise that a higher-priced Sol request or a cached prompt will improve every workload.