OpenAI announced on July 30, 2026 that the GPT-5.6 API will cost dramatically less than its predecessor, rolling out three tiered models—Sol, Terra and Luna—plus steep discounts for cached reads.
Why the new pricing matters
Since the launch of GPT-5, API costs have been a major line item for anyone building conversational agents, retrieval-augmented generation (RAG) pipelines or large-scale data-extraction tools. By slashing input fees to $5, $2 and $0.20 per million tokens for Sol, Terra and Luna respectively, OpenAI gives developers a cheaper way to choose the level of reasoning they need without redesigning their entire stack. The savings hit hardest for Terra, which now runs at 40 % of the old flagship price, and Luna, which sits at just 4 % of that benchmark.
The three models, broken down
| Model | Input price (per 1 M tokens) | Output price (per 1 M tokens) | Typical use |
|---|---|---|---|
| Sol | $5 | $30 | Deep reasoning, chain-of-thought tasks |
| Terra | $2 | $12 | Production workloads, balanced performance |
| Luna | $0.20 | $1.20 | High-volume, simple extraction or classification |
Sol remains the most expensive but offers the richest reasoning depth. Terra is positioned as the default for most applications, while Luna is a “high-scale” option where depth can be sacrificed for cost.
Caching cuts the bill even further
OpenAI introduced a caching layer that discounts read operations by 90 %. The rates for cached input are:
- Sol: $0.50 / M tokens
- Terra: $0.20 / M tokens
- Luna: $0.02 / M tokens
Writes to the cache cost 1.25 × the corresponding input rate, and a cache entry must live at least 30 minutes before it can be evicted.
Context-length surcharges keep you honest
The API now imposes a surcharge when a request exceeds 272 K tokens of input. The base price doubles for the input and rises by 1.5 × for the output. Sol’s overage rates are listed at $10 input and $45 output per million tokens, making it easy to calculate the penalty for unusually long contexts.
What developers lose
The price cuts come with two notable omissions. First, there is no permanent free tier, meaning every request incurs a charge no matter how small. Second, OpenAI has not yet introduced batch-endpoint discounts for GPT-5.6, a feature that helped large-scale users lower per-call costs in earlier versions.
How to keep costs down
- Pick the right model: Use Sol only for tasks that truly need deep reasoning; default to Terra for most production work; reserve Luna for high-throughput, low-complexity jobs.
- Use caching: Store prompts and intermediate results that are reused across calls; the 90 % read discount can dwarf the base price reduction.
- Mind context length: Split very long inputs into smaller chunks or truncate non-essential parts to avoid the 2× input surcharge.
- Watch cache lifetimes: The minimum cache duration is 30 minutes.
What’s next
Developers should keep an eye on the official pricing page for any tweaks, especially around cache-write fees or context-length thresholds.
Takeaway: GPT-5.6’s tiered pricing and aggressive caching discounts make it an affordable OpenAI model, but the lack of a free tier and batch discounts means savvy cost-management will still be essential for any sizable deployment.