Foreword¶

On July 30, 2026, OpenAI announced a price cut for two tiers of models in the GPT-5.6 family via its API. The fastest and most affordable GPT-5.6 Luna saw an 80% reduction in input token pricing, dropping to $0.2 per million tokens. The mid-tier GPT-5.6 Terra received a 20% price cut, while the flagship GPT-5.6 Sol maintained its original pricing.

This comes less than three weeks after the three GPT-5.6 models (Sol / Terra / Luna) officially reached General Availability (GA) on July 9. For developers running Agents, Codex pipelines, or large-scale API calls, this price adjustment is not merely a “promotion”: it represents OpenAI directly passing on internal inference efficiency dividends to customers. It also indirectly signals that the AI inference cost war in the second half of 2026 has officially entered the open stage.

This article梳理 new pricing tiers, the engineering logic behind the price cuts, and practical billing-related details such as Prompt Caching and Codex quotas, based on official OpenAI announcements and authoritative reports from CNBC and other sources.

GPT-5.6 Three-Tier Model Lineup and Price Adjustment Overview¶

The GPT-5.6 family is tiered by capability, with numbers representing generations, and Sol / Terra / Luna representing independently and sustainably evolving capability tiers:

All three models share a 1.05 million token context window and 128K maximum output length. Their API model IDs are gpt-5.6-luna, gpt-5.6-terra, and gpt-5.6-sol respectively (the gpt-5.6 alias routes to Sol).

OpenAI stated in its announcement that its strategy remains “every generation of AI accomplishes more work at a lower cost”. The decision to not cut Sol’s price while significantly reducing costs for Luna and Terra essentially guides developers to select models based on task complexity, rather than defaulting to the flagship tier for all use cases.

Efficiency Behind the Price Cut: Models Save Costs for Themselves¶

OpenAI attributes this price cut to a ~20% reduction in end-to-end costs for serving GPT-5.6 itself, as well as a over 15% improvement in token generation efficiency. On July 29, the company disclosed that GPT-5.6 participated in the rewriting and optimization of production code during its development, with the model “reducing its own operational overhead” — this kind of self-bootstrapping efficiency improvement is becoming the hard foundation for large model vendors to cut prices.

For enterprise customers, the background is more practical: after the “use tokens freely” era sparked by ChatGPT in 2022, many companies’ AI bills have reached hundreds of millions of dollars, and they have begun to actively control costs. At the same time, low-cost or open-weight models such as Moonshot’s Kimi K3 and Google Gemini 3.6 Flash are closing in on closed-source flagship models on some benchmarks, and Anthropic also launched the cheaper Claude Opus 5 in July. OpenAI’s second price adjustment within three weeks can be seen as a direct response to the cost-sensitive workload market.

Prompt Caching: Luna’s Input Cost Can Be As Low As $0.02 per Million Tokens¶

GPT-5.6 has supported more controllable Prompt Caching (explicit cache breakpoints, minimum 30-minute cache lifecycle) since its launch. Combined with the new pricing announced on July 30, actual bills can be reduced even further:
- Cache Read: Charged at 10% of the standard input price (equivalent to a 90% discount). After a cache hit for Luna, input costs drop to approximately \(0.02 per million tokens**. - **Cache Write**: Charged at **1.25 times** the uncached input price. - **Batch / Flex Processing**: Approximately half of the standard price; Luna's price in this mode is approximately **\)0.1 / $0.6 per million tokens (input/output).

For Agent-style applications, system prompts, tool schemas, and RAG knowledge base prefixes are often highly repetitive. Placing stable prefixes before the prompt_cache_breakpoint and user dynamic content after it is the key engineering approach to making Luna usable while remaining “fast and affordable” — this is also why OpenAI positions Luna as a “callable tool for completing multi-step workflows” rather than just a small classification model.

Codex, ChatGPT Work, and Fast Mode¶

The price cut covers OpenAI API, Codex, and ChatGPT Work:
- Subscription fees and quota budgets remain unchanged, but usage of Terra and Luna will consume fewer credits, equivalent to an indirect capacity expansion.
- In Codex: Free / Go users can use Terra; Plus, Pro, Business, and Enterprise users can choose between Terra and Luna.
- GPT-5.6 Sol has added an API Fast Mode, replacing the original Priority Processing. It claims to be up to 2.5 times faster than standard processing, with a price twice the standard rate; the corresponding endpoint in Codex is /fast. Coding Agents requiring low latency can use this to balance speed and costs on Sol.

Pricing via channels such as AWS Bedrock is billed separately by cloud vendors, and OpenAI noted that prices may differ from the official rate card, with rollouts to be rolled out in batches.

How Developers Can Select the Right Model¶

Based on verified pricing, we can roughly categorize workloads and corresponding model choices:
1. High-throughput, structured extraction, routing/classification, simple toolchains → Luna. With $0.2 per million input tokens + Prompt Caching, it is suitable for moving Agents from “demo” to “full-scale launch”.
2. Daily knowledge work, medium-complexity Agents, default Codex tier → Terra. Its performance matches the GPT-5.5 tier, and its cost-performance ratio is more prominent after the price cut.
3. Long reasoning, complex coding, Multi-agent / Programmatic Tool Calling → Sol (use Fast Mode when necessary).

If your pipeline was previously stuck at POC due to Luna’s $1 per million input tokens, this 80% price cut is worth re-running a cost model for; if you are already using Terra, the 20% price cut will directly show up on your monthly bill.

Summary¶

The July 30, 2026 price cut for GPT-5.6 Luna / Terra marks OpenAI’s second price signal less than three weeks after GA: inference efficiency improvements can be quickly translated into API unit prices, and the long context and multi-round tool calls in the Agent era will turn “how much per million tokens” into a core indicator of whether a product can scale. Luna’s $0.2 per million input tokens, which can drop to $0.02 per million with caching, has lowered the threshold for cost-sensitive Agents even further — the next competition will not only be about model benchmark scores, but also about who can sell the same quality inference at a lower price.