OutYet reporting
GPT-5.6 Sol’s new Fast mode changes the routing decision, not the model price
OpenAI has held Sol’s token price steady while cutting the cost of its lower tiers and selling a faster Sol path. The practical question for engineering teams is when latency justifies that premium.
OpenAI’s July 30 GPT-5.6 update changes the operating economics around the family without changing Sol’s standard token price. The company cut Luna’s API price by 80 percent to $0.20 per million input tokens and $1.20 per million output tokens, and cut Terra by 20 percent to $2 and $12 respectively. For Sol, the change is Fast mode: OpenAI says it can deliver up to 2.5 times the speed of Standard processing at twice the price, with no stated change in intelligence.
The timing clarifies the intended product split. OpenAI’s earlier GPT-5.6 product page describes Sol as the flagship tier, Terra as a lower-cost tier positioned against GPT-5.5, and Luna as the fastest and most affordable tier. That page listed Sol at $5 per million input tokens and $30 per million output tokens, compared with $2.50 and $15 for Terra and $1 and $6 for Luna. The July 30 update therefore makes routine and high-volume routing cheaper while keeping the expensive tier’s standard price intact.
OpenAI’s published comparisons suggest why teams may still reserve Sol for difficult work: its GPT-5.6 table reports a 64.6 percent SWE-Bench Pro score for Sol versus 59.4 percent for GPT-5.5, and 88.8 percent versus 85.6 percent on Terminal-Bench 2.1. Those are provider-published evaluation results, not independent validation, but they frame the commercial choice OpenAI is making. The new discounts make it easier to use Terra or Luna as defaults, then escalate to Sol when an application’s own evaluation shows that the additional capability is material.
Amazon’s Bedrock documentation adds a concrete implementation angle. It describes GPT-5.6 Sol as the tier for complex reasoning and agentic coding, and says its explicit prompt-caching feature can reuse selected prompt sections for 30 minutes. Amazon says cache reads receive a 90 percent discount while writes cost 1.25 times the uncached input rate, making the feature best suited to stable, repeated prefixes such as system instructions, tool definitions, and reference material in an agent loop.
There are meaningful limits to this optimization. On Bedrock, GPT-5.6 uses the OpenAI-compatible Responses API rather than the older Chat Completions shape, so a migration includes request and response handling changes. Amazon also lists Sol only in US East regions while Terra and Luna have an additional US West region. Teams should therefore test routing, cache-hit behavior, regional deployment, and end-to-end task quality together; lower list prices or faster responses alone do not prove a lower cost per successful task.