GPT-5.6 API Price Drop and Fast Mode Guide: New Prices for Luna, Terra, Sol

On July 30, 2026, OpenAI adjusted the pricing and low-latency processing of the GPT-5.6 API. This change includes three things:

• GPT-5.6 Luna price reduced by 80%;

• GPT-5.6 Terra price reduced by 20%;

• Priority Processing renamed to Fast mode. GPT-5.6 Sol can reach up to 2.5 times the processing speed of Standard in Fast mode, and cost twice as much as Standard. Existing code does not need to be changed immediately. Continuing to use service_tier: "priority" in requests will result in the same processing behavior as service_tier: "fast". Official source:

• OpenAI Product Update Notes

• Latest OpenAI API Price

• Fast mode user documentation If you want to understand the original positioning and release background of Sol, Terra, and Luna, you can first read: GPT-5.6 Sol Limited Preview and Model Layering.

Let’s look at the conclusion first: who benefits the most

Luna is the model with the largest price drop this time. Its Standard short-context input price dropped from $1 per million tokens to $0.20 per million, and output dropped from $6 to $1.20. Terra’s Standard short-context input price dropped from $2.50 to $2, and output dropped from $15 to $12. Sol’s Standard short-context price remained $5 input and $30 output. Its main change is not the Standard price cut, but the acceleration of Fast Mode. You can understand the three-tier model like this:

GPT-5.6 Standard Latest Price

The official price is charged per million tokens. Here is the short-context Standard price:

• Uncached token input;
• Cache input token;
• Cache writes tokens;
• Output token;
• The request ratio between Standard and Fast mode;
• Short context to long context request ratio.

What does it mean after Luna’s 80% price cut?

Luna’s input and output prices are now 20% of their original value. The old short-context Standard price is:

• Input: $1.00 / 1M token;

• Output: $6.00 / 1M token. The new prices are:

• Input: $0.20 / 1M token;

• Output: $1.20 / 1M token. Suppose a batch task uses 100 million input tokens and 20 million output tokens per month: Old price estimate:

1
2
3

输入：100 × $1.00 = $100
输出：20 × $6.00 = $120
合计：$220

New Price Estimate:

1
2
3

输入：100 × $0.20 = $20
输出：20 × $1.20 = $24
合计：$44

With the token usage unchanged, the fee in this example drops from $220 to $44. But price drops don’t mean all tasks should migrate to Luna. First, check command adherence, structured output, tool calls, and error rates with real requests. If the model becomes cheaper but leads to more retries or manual corrections, the total cost may not actually decrease.

How to calculate after Terra’s 20% price cut

Terra’s old short context Standard price is:

• Input: $2.50 / 1M token;

• Output: $15.00 / 1M token. The new prices are:

• Input: $2.00 / 1M token;

• Output: $12.00 / 1M token. Assuming 100 million input tokens and 20 million output tokens are used monthly: Old price estimate:

1
2
3

输入：100 × $2.50 = $250
输出：20 × $15.00 = $300
合计：$550

New Price Estimate:

1
2
3

输入：100 × $2.00 = $200
输出：20 × $12.00 = $240
合计：$440

For example, you can reduce $110 per month, which is exactly a 20% reduction. Terra is better suited for teams that already use it as the default production model. There’s no need to change the model ID; bills are calculated at the new price.

What is Fast mode?

Fast mode is the new name for the original Priority Processing. It remains a pay-as-you-go, low-latency processing method, but the official product positioning and parameters have become more intuitive. The goals of Fast mode are:

• Reduced response wait times;
• Provides more stable latency;
• Retain pay-as-you-go methods;
• Services have high-value user requests with stable traffic. For gpt-5.6-sol, Fast mode can reach speeds up to 2.5 times faster than Standard. “Up to 2.5x” does not mean every request is fixed at 2.5x increase. Actual latency still depends on input length, output length, tool calls, network, and system load.

Latest Fast mode prices

The current short-context Fast mode prices are as follows:

1
2
3

输入：10 × $5.00 = $50
输出：2 × $30.00 = $60
合计：$110

Fast mode：

1
2
3

输入：10 × $10.00 = $100
输出：2 × $60.00 = $120
合计：$220

Whether it’s worth paying an extra $110 depends on whether latency reduction improves conversion rates, task completion rates, or labor efficiency.

Enable Fast mode for a single request

Responses API usage service_tier: "fast":

1
2
3
4
5
6
7
8

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "input": "分析这段错误日志并给出修复顺序",
    "service_tier": "fast"
  }'

Python SDK example:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-5.6-sol",
    input="分析这段错误日志并给出修复顺序",
    service_tier="fast",
)

print(response.output_text)

Fast mode can be used for both the Responses API and the Chat Completions API.

Do you need to modify the old priority parameters?

No immediate modification is needed. The following old requests are still valid:

1
2
3
4
5

response = client.responses.create(
    model="gpt-5.6-sol",
    input="检查这个发布方案",
    service_tier="priority",
)

For supported models, priority and fast will have the same handling behavior. It is recommended to use fast for new code, as the name matches the current documentation. Old code can be migrated gradually according to the normal release cycle without urgently replacing everything. Special attention should be paid to the response object. For GPT-5.6 and earlier models, even if the request is written as fast, the service_tier in the response may still return priority. Therefore, monitoring systems should not directly judge return priority as if the configuration is invalid.

Fast mode is enabled by default at the project level

If most user requests in the project require low latency, you can enable the following in the project settings:

1. Open the project settings in OpenAI API Platform;
2. Enter General;
3. Find the Project Service Tier;
4. Choose Fast. After setting, requests without explicit service_tier will gradually switch to Fast mode. The official explanation states that project requests will gradually switch over time. Do not treat the first minute of saved data as the final result. A safer approach is to enable it as requested, use function switches to control a small amount of traffic, and then decide whether to change the project’s default values.

Rate limits do not double

Fast mode and Standard handle the same model rate limit. Enabling Fast mode does not automatically increase TPM or RPM quotas. Existing retry, backoff, and concurrency controls still need to be retained. If the system immediately sends more follow-up requests due to faster returns, it may trigger rate limits more quickly. During verification, the following should be observed:

• Request delay;
• 429 Response;
• Tokens per minute;
• Number of concurrent tasks;
• Number of reattempts;
• Actual completion time.

Sudden traffic may be downgraded to Standard

Fast mode has a ramp rate limit. When traffic reaches at least 1 million tokens per minute and TPM increases by more than 50% within 15 minutes, some Fast mode requests may be downgraded. After downgrading:

• Requests processed at Standard speed;

• Fees are calculated at Standard rates;

• The service_tier in the response is default. Methods to reduce trigger risk include:

• Gradually scaling up when switching models or snapshots;

• Use functional switches to migrate traffic within hours;

• Avoid putting large ETL or batch tasks into Fast mode;

• Keep the Standard downgrade path;

• Statistics on delays and fees by service level.

Which scenarios are not suitable for Fast mode?

Fast mode is suitable for requests where latency directly affects user experience, such as:

• Real-time coding assistant;

• Interactive Agent;

• The complex analysis users are waiting for;

• High-value customer service or sales processes;

• Products sensitive to initial word latency and total time consumption. Unsuitable scenarios include:

• Offline batch processing;

• Nighttime data cleaning;

• Queues insensitive to completion times;

• Tasks that can use Batch or Flex;

• Large-scale ETL;

• Full traffic enabled solely for benchmarking. If users don’t notice a difference of just a few dozen seconds, paying twice the token price is usually unnecessary.

Current limitations of Fast mode

The restrictions listed in the official documentation include:

• Does not support long contexts;
• No support for fine-tuning models;
• Embeddings are not supported;
• Regional availability depends on local laws and regulations;
• No guarantee that every GPT model will be supported in the future;
• Scale Tier and Fast mode are billed separately. Fast mode supports the multimodal capabilities already available in Standard mode, including image input. Cache input can still be discounted. Fast mode is compatible with data residency, Zero Data Retention, and BAA, but original endpoints, tools, eligibility, and contract requirements still apply.

How to verify if Fast mode is really faster

Don’t jump to conclusions after just one request. Prepare a set of real business samples, run them separately in Standard and Fast modes. At the very least:

A set of secure migration steps

Production environments can be migrated in the following order:

1. Retain the current Standard baseline data;
2. Choose the interface that relies most on low latency;
3. Set service_tier: "fast" for 1% of traffic;
4. Compare P50, P95, success rates, and unit task cost;
5. Check the priority and default in the response;
6. Gradually expand to 5%, 10%, and 25%;
7. Observe whether the ramp rate limit is triggered;
8. Confirm improvement in business metrics before continuing to increase volume;
9. Tasks that do not require low latency can continue using Standard, Batch, or Flex;
10. Finally, consider enabling the project-level default setting. Don’t switch across the entire organization uniformly by model name. Interfaces within the same project may have completely different sensitivities to latency and cost.

How to re-split between Luna, Terra, and Sol

After the price cut, you can redesign the task route:

• Luna: Lightweight tasks with high throughput, clear rules, and automatic acceptance;
• Terra: default production traffic and medium complexity tasks;
• SOL: complex reasoning, difficult coding, and high-value failed retrys;
• Sol Fast mode: Complex requests where users are waiting and delays can affect business outcomes. First, Luna or Terra handles it, then upgrades failed or low-confidence requests to Sol, which is usually more cost-effective than all requests going directly into Sol. Fast mode should be delayed routing, not capability routing. It won’t turn Luna into Sol, nor will it automatically improve the quality of model answers.

Frequently Asked Questions

Have the model IDs of Luna and Terra changed?

No. gpt-5.6-luna and gpt-5.6-terra are still in use, with prices adjusted starting July 30, 2026.

Has the Standard price of Sol dropped?

The official focus of this update is on price cuts for Luna and Terra, as well as the speed up of Sol’s Fast mode. Sol Standard short-context prices remain $5 for input and $30 for output.

Will priority be abandoned?

The current official documentation clearly states that both priority and fast can be used. It is recommended to use fast for new code, while old code can be migrated gradually.

Why does the request write fast, but the response returns priority?

This is the current compatibility behavior of GPT-5.6 and earlier models. Don’t judge that a request hasn’t entered Fast mode just by the response name.

Is Fast mode always 2.5 times faster?

No. The official statement is up to 2.5 times, and specific requests are affected by length, load, utility calls, and network.

Does Fast mode add a rate limit?

No. Fast mode shares the model rate limit with Standard.

How to charge after downgrading sudden traffic issues?

Downgrade requests are processed at Standard speed and Standard price, and the response service_tier will display default.

Can you still enjoy discounts when entering with cache?

Yes. Fast mode still applies eligible cache input discounts.

Summary

This adjustment in GPT-5.6 simultaneously changes cost and latency strategies. Luna has dropped by 80%, significantly lowering the cost of high-throughput, lightweight tasks. Terra has dropped by 20%, directly lowering the cost of default production traffic. Sol Standard remains price-unchanged, but Fast Mode offers up to 2.5 times the speed, at the cost of double the token price. Old service_tier: "priority" remain compatible and do not urgently modify all code. New access is recommended to use service_tier: "fast" and gradually scale up by interface. What really needs to be verified is not “faster speed,” but whether faster response improves business outcomes. If latency gains cannot offset the price premium, continuing with Standard, Batch, or Flex is more reasonable.