OpenAI has cut the price of Luna, the smallest model in the GPT-5.6 lineup, to one-fifth of its launch price. Starting July 30, 2026, the API price per million tokens will be $0.20 for input and $1.20 for output. Terra also got 20% cheaper, dropping to $2.00 for input and $12.00 for output. However, the flagship model Sol remains unchanged at $5.00 for input and $30.00 for output.
An 80% cut sounds dramatic. But the new Luna's input price is identical to GPT-5.4 nano, and its output price is only 4% lower than the previous $1.25. In other words, OpenAI has significantly revised the pricing it originally attached to Luna during the limited preview on June 26, moving it back to the price band that the old nano tier occupied for high-volume processing.
Nor has this outright leapfrogged China's low-cost APIs. DeepSeek V4 Flash is cheaper than Luna on both input and output. The comparison with V4 Pro, meanwhile, flips depending on the input-to-output ratio, cache usage, and the length of a single prompt. To read this price sheet properly, you need to start by checking what the 80% figure is actually measured against.
The 80% Cut Is Relative to Luna's Launch-Time Price
GPT-5.6 entered limited preview on June 26 and reached general availability on July 9. Luna's initial pricing was $1.00 for input and $6.00 for output, so today's $0.20 and $1.20 both represent an 80% reduction from that baseline. Counting from general availability, that's 21 days; even counting from the limited preview, it's just 34 days before the pricing was restructured.
On the same day, Terra's price dropped 20%, from $2.50 to $2.00 for input and from $15.00 to $12.00 for output. OpenAI is leaving subscription pricing and usage caps for ChatGPT and Codex unchanged, while reducing the credit consumption when using Terra and Luna. The API pricing takes effect from July 30, and OpenAI says it will begin rolling out on AWS the same day.
Luna's role within the product lineup also matters here. OpenAI describes Luna as "roughly corresponding to the nano tier in the previous GPT-5 family." The current GPT-5.4 nano is priced at $0.20 for input, $0.02 for cached input, and $1.25 for output. The new Luna matches the input and cache pricing exactly, with only the output price coming down 4% to $1.20.
So the 80% figure doesn't mean OpenAI's cheapest API got 80% cheaper than before in absolute terms. It's a comparison against Luna's own launch price. It's more accurate to view this change as repositioning a new-generation small model back into the price band previously occupied by the old generation's high-volume tier.
The Comparison with DeepSeek Flips Depending on the Input-to-Output Ratio
Checking each company's official pricing pages as of July 31, 2026, the standard rates per million tokens are as follows. Input prices are for non-cached requests, excluding taxes and tool-usage fees.
| Model | Input | Cached Input | Output | Context |
|---|---|---|---|---|
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | 1.05M |
| DeepSeek V4 Flash | $0.14 | $0.0028 | $0.28 | 1M |
| DeepSeek V4 Pro | $0.435 | $0.003625 | $0.87 | 1M |
| Kimi K3 | $3.00 | $0.30 | $15.00 | 1,048,576 |
Luna is about 93.3% cheaper than Kimi K3 on input and 92% cheaper on output. However, Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-weight flagship model, and it isn't in the same performance class as nano-tier Luna. The gap in unit pricing shouldn't be read directly as a gap in capability or in per-task cost.
The relationship with DeepSeek V4 Pro is more complicated. Suppose each request has input under 272,000 tokens, and the monthly totals come to 1 million input tokens and 100,000 output tokens. In that case, Luna costs $0.32 and V4 Pro costs $0.522 — Luna is cheaper for input-heavy workloads. But if input and output are equal in volume, Luna costs $1.40 while V4 Pro costs $1.305, putting V4 Pro below Luna.
Working from the standard rates, the break-even point falls around where input exceeds output by roughly 1.4 times. For workloads that read long codebases or documents and return short results, Luna's low input pricing works in its favor. But for workloads involving long reasoning or long-form text generation, Luna's $1.20 output price becomes relatively costly. V4 Flash, at $0.14 for input and $0.28 for output, undercuts Luna on both metrics.
The Long-Context Surcharge Kicks in Above 272,000 Tokens
There's a condition attached to Luna that isn't visible from the table alone. If a single input exceeds 272,000 tokens, the entire request is billed at double the input rate and 1.5 times the output rate. Being able to use a 1.05-million-token context window is a different matter from being able to process that entire length at the same unit price.
For example, a single request using 900,000 input tokens and 100,000 output tokens would cost $0.54 with Luna. That's 1 million tokens total, which fits within Luna's context ceiling and its maximum input of 922,000 tokens. DeepSeek's current pricing page contains no comparable long-context surcharge, and for the same token counts, V4 Pro would cost $0.4785 and V4 Flash would cost $0.154. Even if Luna undercuts V4 Pro when sending many short requests, feeding it one enormous context in a single request can flip the ranking.
The difference also shows up for agents that rely heavily on caching. Luna's cached input rate is $0.02, whereas DeepSeek's is $0.0028 for V4 Flash and $0.003625 for V4 Pro. With Luna, writing a new cache entry is billed at 1.25 times the standard input rate. For use cases that repeatedly send conversation history or tool definitions, you also need to factor in cache hit rates and how often the cache gets rewritten.
It's also worth noting that each model uses a different tokenizer and produces different amounts of tokens during inference. The per-million-token price is a useful entry point for procurement, but it doesn't guarantee the total task cost — including retries, tool calls, and reasoning effort needed to reach the same quality bar. You need to run your own evaluation using your actual input distribution and quality standards.
You Can't Extrapolate an 80% Price Cut from a 20% Cost Improvement
OpenAI attributes this pricing change to efficiency gains across the model, inference infrastructure, and agent execution stack. In a technical article published the day before, the company said that GPT-5.6 Sol rewrote its production GPU kernels, and combined with broader kernel improvements, this cut overall serving costs by 20%.
Another lever is speculative decoding. A small draft model proposes several next tokens in advance, and the main model verifies them in parallel. When a proposal is accepted, it reduces the number of sequential steps the main model needs to run. OpenAI says Sol ran hundreds of experiments varying the draft model's architecture and size, improving token-generation efficiency by more than 15%.
However, the 20% figure refers to overall serving cost, while the 15%-plus figure refers to token-generation efficiency — these aren't the same metric. You can't simply add them together to get a 35% cost reduction, nor can you assume Luna's own costs fell by 80%. OpenAI hasn't disclosed Luna-specific cost figures or utilization rates. Neither its profit margin nor a breakdown of the factors behind the 80% price cut has been made public.
Alongside the pricing change, OpenAI also introduced a Fast mode for Sol. It replaces the previous Priority Processing, offering up to 2.5x the speed at twice the standard price. The strategy is to have the cheap Luna handle bulk processing, while charging a premium for Sol in cases where latency matters more than cost. This announcement isn't a blanket price cut so much as a restructuring that widens the price bands available for different use cases.
Cheap APIs and Open Weights Are Different Choices
When thinking about competition with Chinese players, it's important to separate API unit pricing from the availability of open model weights. DeepSeek has released V4 Pro (1.6 trillion total parameters, 49 billion active) and V4 Flash (284 billion total, 13 billion active) with published weights, offering both an API and the model weights simultaneously. Kimi K3 is also a disclosed MoE model with 2.8 trillion total parameters and 104 billion active parameters, with OpenAI having disclosed details such as its architecture selecting 16 out of 896 experts and its MXFP4 weight format.
Luna is a closed API, managed entirely by OpenAI, that gives immediate access to tool use and a 1.05-million-token context window. Open-weight models, by contrast, let users and third-party providers deploy the model weights themselves — within the bounds of the license — and choose their own inference environment. OpenAI's price cut narrows the cost gap against self-hosting for users who'd rather leave operations to someone else, but the distinction of whether you actually hold the weights — and can research, deploy, and modify them — remains.
OpenAI estimates that Luna delivers performance close to a frontier-class model from roughly a year ago, at about 6% of the comparable task cost and roughly 9 times the speed. That's the company's own assessment, not something proven by this pricing announcement itself. What matters next isn't who cuts prices again first, but rather measuring total output, retry counts, and costs after long-context surcharges — using the same real-world data and quality bar across providers. Whether that 80% figure actually shows up on your bill will be decided there.