GPT-5.6 Price Cuts Explained: Luna Drops 80%, Terra Falls 20%, and Sol Gets Fast Mode

Introduction

OpenAI has changed the economics of the GPT-5.6 family less than a month after its general release.

Starting July 30, 2026:

• GPT-5.6 Luna costs 80% less.
• GPT-5.6 Terra costs 20% less.
• GPT-5.6 Sol keeps its existing Standard price but gains a faster API processing option.
• Priority Processing has been renamed Fast mode.

The update is not limited to a temporary discount. OpenAI says the lower prices reflect improvements across the model, inference stack, routing systems, context management, and the agentic software that connects models to tools.

The practical result is a wider range of production choices.

Luna is now priced for high-volume, cost-sensitive agent workflows. Terra remains the balanced option for everyday work that needs more intelligence. Sol continues to serve the most demanding tasks, while Fast mode offers lower latency when waiting time matters more than the token premium.

What Changed on July 30, 2026

When GPT-5.6 launched on July 9, OpenAI listed the following Standard short-context API prices:

The July 30 update changed Terra and Luna:

Luna’s price is now one fifth of its original launch rate.

Terra’s new rate is four fifths of its original price.

Sol remains unchanged under Standard processing, but Fast mode now gives developers an option to receive responses at up to 2.5 times the speed for twice the Standard token price.

The Current Standard API Pricing

The headline input and output prices do not show the complete billing structure.

GPT-5.6 supports cached-input discounts and explicit cache writes. Requests using more than 272,000 input tokens also receive long-context pricing for the entire request.

Short-Context Standard Pricing

The following rates apply per one million tokens:

Cache reads receive a 90% discount relative to ordinary uncached input.

Cache writes are billed at 1.25 times the uncached input rate.

Long-Context Standard Pricing

For prompts containing more than 272,000 input tokens, OpenAI currently charges the following rates for the full request:

The long-context rule means cost estimates should consider both the number of tokens and whether a request crosses the 272,000-token threshold.

A single very large prompt is not billed by splitting the first 272,000 tokens at the short-context rate and the remainder at the higher rate. The long-context multiplier applies to the full request.

Matching Intelligence to the Outcome

OpenAI’s main argument is that efficient AI deployment begins with the result, not the model name.

Different stages of the same workflow may require different levels of intelligence, speed, and reliability.

A task’s ideal model depends on:

• The consequence of an error.
• The difficulty of the work.
• The acceptable response time.
• The volume of requests.
• The cost of human review.
• The ability to verify the result automatically.
• Whether the task is clearly specified or still ambiguous.

When Luna Makes Sense

GPT-5.6 Luna is designed for cost-sensitive, high-volume work.

Its current API model page describes it as corresponding roughly to the nano tier used in earlier GPT-5 families.

Luna is a practical candidate for:

• Classification.
• Structured extraction.
• Background agent steps.
• Routine code changes.
• Test generation.
• Document triage.
• Repetitive tool-calling loops.
• Large-scale processing where each individual task has limited downside.

OpenAI estimates that Luna can deliver performance comparable to models considered frontier-class about a year earlier at roughly six cents per dollar of task cost and nearly nine times the speed.

That comparison is based on OpenAI’s evaluations and cost methodology. Teams should validate it against their own workloads.

When Terra Makes Sense

GPT-5.6 Terra is the balanced member of the family.

OpenAI positions it roughly where mini models sat in earlier GPT-5 generations, but with stronger agentic and professional-work capabilities.

Terra is suited to work such as:

• Workspace question answering.
• Scoped research.
• Everyday coding.
• Document analysis.
• Agent planning with moderate complexity.
• Customer-support workflows that need reasoning.
• Repeated business tasks where Luna is not reliable enough but Sol is unnecessary.

When Sol Makes Sense

GPT-5.6 Sol is the flagship model for complex professional work.

It is the best fit when the model must:

• Resolve ambiguity.
• Build a plan for a difficult project.
• Review high-value decisions.
• Handle complex coding or debugging.
• Coordinate tools across long workflows.
• Perform deep research.
• Produce or verify work where failure is expensive.

The gpt-5.6 API alias routes to gpt-5.6-sol.

One Workflow Can Use More Than One Model

The pricing update makes mixed-model workflows more practical.

A coding agent does not need to use Sol for every token and every tool call.

A possible architecture is:

1. Use Sol to understand the repository and identify uncertainty.
2. Ask Sol to create a plan and define success criteria.
3. Send clearly specified implementation tasks to Luna.
4. Use Luna to write routine tests and perform repetitive checks.
5. Route uncertain or failed steps to Terra or Sol.
6. Use Sol for final review when the consequence of an error is high.

The same pattern can work outside software engineering.

A document-processing system might use Luna for extraction, Terra for synthesis, and Sol only when the material is ambiguous or consequential.

The correct routing strategy should be determined through evaluations rather than assumptions.

Customer Results Shared by OpenAI

OpenAI’s announcement includes early customer feedback from several production users.

These examples are company- and customer-reported rather than independent benchmarks.

These examples show why per-token pricing is not the only useful measure.

A cheaper model can also alter:

• The number of tool calls.
• Cache reuse.
• Context length.
• Output verbosity.
• Retry frequency.
• Human-review requirements.
• Total completion time.

The more meaningful metric is usually cost per successful outcome.

How OpenAI Says It Improved Efficiency

OpenAI attributes the price-performance improvement to three connected layers.

1. The Models

GPT-5.6 models are designed to complete work more directly.

A model that needs fewer output tokens, fewer retries, or fewer reasoning loops can reduce total task cost even before the list price changes.

2. The Inference Systems

The production stack determines how efficiently the model uses hardware.

OpenAI says optimized serving software generates tokens more efficiently and keeps compute resources productive.

3. The Agentic Harness

The harness is the software around the model that provides tools, context, routing, state, and workflow control.

OpenAI says better context management helps agents avoid repeating completed work.

This can reduce:

• Duplicate tool calls.
• Reprocessing of unchanged context.
• Repeated planning.
• Unnecessary model round trips.
• Tokens spent restating prior results.

The three layers interact.

A stronger model may find an optimization in the serving stack. That optimization reduces cost, which makes larger-scale agent usage economical. More usage then creates additional opportunities to improve the system.

GPT-5.6 Sol Helped Optimize Its Own Serving Stack

One of the most notable claims in the announcement is that GPT-5.6 Sol contributed to OpenAI’s internal efficiency work.

Within a human-led engineering process, OpenAI says Sol:

• Rewrote and optimized production kernels.
• Designed and ran hundreds of token-generation experiments.
• Monitored training.
• Intervened when problems appeared.

OpenAI reports that the kernel work reduced the end-to-end cost of serving the model by approximately 20%.

It also says the experiments improved token-generation efficiency by more than 15%.

These are OpenAI’s internal results and have not been independently reproduced through a public benchmark.

The larger point is that the model is becoming part of the process used to improve the infrastructure that runs the model.

That creates a tighter engineering feedback loop:

Stronger model
→ better infrastructure optimization
→ lower serving cost
→ broader use
→ more feedback and investment
→ stronger next model

A Compute Strategy Built for Scale

Lower prices do not mean OpenAI needs less total compute.

The company argues that abundant intelligence requires both:

• More compute.
• More productive compute.

The first expands total capacity.

The second increases the useful work completed by each unit of hardware.

OpenAI says it is building a diversified infrastructure portfolio and matching workloads to the systems best suited to run them.

This supports both ends of the GPT-5.6 family:

• Luna and Terra make high-volume workloads cheaper.
• Sol and Fast mode support difficult, latency-sensitive work.

Examples of workloads that may become easier to run at scale include:

• Large document-analysis pipelines.
• Customer-interaction classification.
• Routine software implementation.
• Background agent automations.
• Repeated data-processing tasks.
• Tool-using workflows with high cache reuse.

Meanwhile, complex Sol requests can use Fast mode when the value of a quicker answer justifies the premium.

Fast Mode Replaces Priority Processing

Fast mode is the new name for OpenAI’s Priority Processing service tier.

Existing API requests that use:

"service_tier": "priority"

remain compatible.

Developers can also use:

"service_tier": "fast"

For GPT-5.6 Sol, OpenAI says Fast mode delivers up to 2.5 times the speed of Standard processing while preserving the same model intelligence.

The trade-off is price.

Fast mode currently costs twice the Standard API token rate for GPT-5.6 Sol.

Cached-input discounts still apply.

Python Example

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-5.6-sol",
    input="Review this migration plan and identify the three highest-risk assumptions.",
    service_tier="fast",
)

print(response.output_text)

Fast mode is available through the Responses API and Chat Completions API for supported models.

For GPT-5.6 and earlier models, the response object may report the service tier as priority even when the request used the newer fast name.

When Fast Mode Is Worth the Premium

Fast mode is not automatically the best setting for every Sol request.

It is most useful when latency directly affects the value of the result.

Examples include:

• A user-facing assistant completing a high-value workflow.
• A coding agent waiting to unblock a developer.
• Interactive research or analysis.
• Incident-response support.
• Real-time review before a decision.
• Production systems with stable, latency-sensitive traffic.

Standard processing may be more economical for:

• Background jobs.
• Overnight research.
• Asynchronous report creation.
• Batch analysis.
• Tasks where a user is not waiting.
• Workflows whose bottleneck is an external tool rather than the model.

A two-times token premium only makes sense when the faster completion creates more value than it costs.

Prompt Caching Matters More in Long Agent Workflows

The new list prices make Luna and Terra cheaper, but prompt caching can have an equally large effect on long-running agents.

Every turn in an agent loop may resend:

• System instructions.
• Tool definitions.
• Shared documents.
• Repository context.
• Conversation history.
• Stable workflow rules.

Without caching, the application can pay repeatedly for the same prefix.

GPT-5.6 supports both automatic caching and explicit cache breakpoints.

OpenAI currently bills:

• Cache reads at 10% of the uncached input rate.
• Cache writes at 125% of the uncached input rate.

A cache write costs more than ordinary input, but later reads are heavily discounted.

Caching is most useful when the same stable prefix is reused enough times to recover the higher write cost.

Applications should track:

• Cache hit rate.
• Prefix stability.
• Time between requests.
• The size of cached context.
• Whether the workflow frequently invalidates the prefix.

Model Specifications Shared Across the Family

The current API model pages list several shared specifications:

All three models are available through the Responses API.

The model pages also list broad tool support, although exact feature availability can vary by endpoint, account, product, and release stage.

Availability in ChatGPT Work, Codex, and the API

OpenAI says GPT-5.6 Terra and Luna remain available in:

• ChatGPT Work.
• Codex.
• The OpenAI API.

Current access described in the price announcement includes:

Sol has its own access rules across ChatGPT, ChatGPT Work, Codex, and the API.

The July 30 update did not reduce ChatGPT or Codex subscription prices.

It also did not increase the stated quota budgets.

Instead, Luna and Terra now consume fewer credits because their underlying usage costs are lower.

OpenAI also said the pricing update would begin rolling out through AWS on the announcement day. Pricing for models offered through third-party platforms can differ from direct OpenAI API pricing.

How to Choose the Right GPT-5.6 Model

A practical selection process can follow five steps.

Step 1: Define the Required Outcome

Write down what counts as success.

Avoid choosing a model based only on a broad label such as “coding” or “research.”

Step 2: Identify the Cost of Failure

A low-risk classification task and a production database migration should not use the same decision rule.

Step 3: Establish a Sol Baseline

For difficult work, test Sol first to understand the highest available quality.

Step 4: Test Terra and Luna on the Same Evaluation Set

Measure whether the lower-cost model preserves the quality that actually matters.

Step 5: Optimize the Whole Workflow

Track:

• Task success.
• Total tokens.
• Cache reuse.
• Tool calls.
• Retries.
• Latency.
• Human-review time.
• Cost per successful result.

Do not optimize only for the cheapest input-token rate.

常见问题

What are the new GPT-5.6 Luna prices?

GPT-5.6 Luna now costs $0.20 per million Standard short-context input tokens, $0.02 per million cached input tokens, and $1.20 per million output tokens. Cache writes cost $0.25 per million tokens.

What are the new GPT-5.6 Terra prices?

GPT-5.6 Terra now costs $2.00 per million Standard short-context input tokens, $0.20 per million cached input tokens, and $12.00 per million output tokens. Cache writes cost $2.50 per million tokens.

Did GPT-5.6 Sol become cheaper?

Its Standard token price did not change. Sol remains $5 per million short-context input tokens and $30 per million output tokens, while Fast mode provides up to 2.5 times faster processing for twice the Standard price.

What happened to Priority Processing?

OpenAI renamed Priority Processing to Fast mode on July 30, 2026. Existing requests using service_tier: "priority" remain compatible, and new requests can use service_tier: "fast".

When does GPT-5.6 long-context pricing apply?

The current model pages state that prompts containing more than 272,000 input tokens use higher input and output rates for the full request. Developers should account for that threshold when estimating the cost of very large prompts.

Which GPT-5.6 model is best for high-volume workloads?

OpenAI positions Luna for cost-sensitive, high-volume work. Terra is the balanced choice when Luna does not provide enough quality, while Sol is intended for the most difficult professional tasks.

Do the price cuts lower ChatGPT and Codex subscription fees?

No. OpenAI says subscription prices and quota budgets remain unchanged. Terra and Luna now consume fewer credits inside supported paid-product workflows.

Are the lower prices available through every cloud provider immediately?

Direct OpenAI API prices changed on July 30. OpenAI said AWS pricing would begin rolling out later that day, but third-party platform pricing and availability can differ, so users should verify the provider’s current rate card.

相关工具

• GPT-5.6 Sol: OpenAI’s flagship GPT-5.6 model for complex professional work.
• GPT-5.6 Terra: The balanced GPT-5.6 model for workloads that need both intelligence and lower cost.
• GPT-5.6 Luna: The fastest and most affordable GPT-5.6 model for high-volume, cost-sensitive tasks.
• Responses API: OpenAI’s primary API for reasoning, tools, multi-turn workflows, and agentic applications.
• Codex: OpenAI’s agentic software-engineering environment using the GPT-5.6 model family.
• OpenAI API Dashboard: The official workspace for API keys, projects, usage, limits, and billing.

Related Links

• Complete OpenAI API Pricing: Current Standard, Batch, Fast, cached-input, cache-write, and long-context token rates.
• Fast Mode Documentation: Official setup instructions, compatibility details, and usage guidance for the renamed service tier.
• GPT-5.6 Model Guidance: Official model-selection, migration, reasoning, caching, and workflow recommendations.
• GPT-5.6 Launch Announcement: The original family launch, model capabilities, initial prices, availability, and benchmarks.
• GPT-5.6 Efficiency Engineering: OpenAI’s explanation of the model and infrastructure work behind the efficiency gains.
• Prompt Caching Guide: Official instructions for automatic and explicit prompt caching.
• OpenAI API Changelog: The dated record of the price update, Fast mode launch, and other API changes.

Summary

OpenAI’s July 30 update reduced GPT-5.6 Luna’s Standard API price by 80% and Terra’s by 20%. Sol’s Standard price remains unchanged, while the new Fast mode can deliver up to 2.5 times faster API responses at twice the token rate.

The announcement is built around a broader price-performance strategy. OpenAI says it is improving the model, inference stack, routing, context management, and agent harness so that each successful task requires less time, fewer tokens, or less compute.

For developers, the right choice is not simply the cheapest model. Luna is designed for high-volume execution, Terra balances cost and intelligence, and Sol handles the most difficult work. Mixed-model routing, caching, evaluations, and cost-per-success measurements can produce larger savings than a model swap alone.

The core change is that GPT-5.6 now offers a wider operating range: much cheaper routine intelligence at the Luna and Terra tiers, with faster frontier intelligence available through Sol when latency justifies the premium.

This Markdown file is an independently written editorial adaptation of the official article. It preserves the topic, factual sequence, and practical meaning without reproducing OpenAI’s wording or customer quotations line by line.