OpenAI Cuts GPT-5.6 Luna API Prices by 80% to $0.20 per Million Input Tokens

GPT-5.6 Luna is now five times cheaper for developers

OpenAI announced the pricing changes on 30 July 2026, only three weeks after launching the GPT-5.6 family. The models—Sol, Terra and Luna—were introduced on 9 July across ChatGPT, Codex and the OpenAI API. The original launch pricing listed Luna at $1 per million input tokens and $6 per million output tokens, as described in OpenAI’s GPT-5.6 launch announcement.

Under the new standard API rates, Luna costs:

• $0.20 per million input tokens, down from $1;
• $0.02 per million cached input tokens;
• $1.20 per million output tokens, down from $6.

That is a straightforward fivefold reduction in both the uncached input and output rates. A workload using one million input tokens and one million output tokens previously cost $7. Under the new pricing, the same workload costs $1.40—a saving of $5.60, or exactly 80%.

OpenAI bills the API in US dollars. The final amount paid by European customers will depend on the exchange rate, VAT and the billing arrangement used by the customer. In practical terms, however, the percentage reduction is the same for developers in the EU as it is for customers elsewhere.

Terra also becomes cheaper, while Sol gets a speed option

The price reduction is not limited to Luna. OpenAI has reduced GPT-5.6 Terra’s input price from $2.50 to $2.00 per million tokens, a 20% decrease. The update does not indicate an equivalent reduction to Terra’s output price, so developers should check the rate applied to their specific usage before revising a production budget.

At the top of the range, OpenAI is offering a faster processing mode for GPT-5.6 Sol through the API. This is aimed at applications where latency matters more than the lowest possible cost, such as interactive assistants, software development tools and agentic systems that must complete several tool calls in sequence.

The pricing structure now gives developers a clearer choice:

• Luna for high-volume, cost-sensitive workloads;
• Terra for a balance between capability and operating cost;
• Sol for demanding tasks where maximum performance or faster processing is more important than price.

Why the cut matters for AI agents

Token pricing has a direct impact on the economics of AI agents. A simple chatbot may generate only a few thousand tokens per interaction. An autonomous agent that reads documents, calls tools, checks its own work and produces a final report can consume hundreds of thousands of tokens during one task.

That makes the Luna reduction particularly important for software companies building workflow automation. At the old price, a system processing 100 million input tokens and 20 million output tokens in a month would cost approximately $220 before other API charges. At the new rates, the same token volume would cost $44. The monthly saving would be $176, assuming all tokens are billed at the standard rates.

For a larger application processing one billion input tokens and 200 million output tokens, the difference grows from approximately $2,200 to $440 per month. This does not eliminate the cost of engineering, monitoring, storage, retrieval, tool calls or human oversight, but it makes experimentation and scaling less expensive.

OpenAI says Luna offers performance comparable to frontier-class models from 2025 at roughly six cents on the dollar per task and at almost nine times the speed. On the Agents’ Last Exam benchmark, OpenAI also says Luna outperforms Claude Fable 5 at an estimated cost per task nearly 99% lower. These are vendor-reported comparisons, so companies should validate them against their own workloads before switching models.

Usage counting also changes in Codex and ChatGPT Work

The lower prices are not restricted to standalone API bills. OpenAI says the reduced Luna and Terra rates are also reflected in how usage is counted for paid subscriptions across Codex and ChatGPT Work. In effect, the same subscription allocation should cover more model usage than before, although the exact benefit will depend on task length, reasoning effort and the plan’s usage limits.

This could be significant for development teams. Codex users often run repeated coding, testing and debugging cycles, while workplace users may process long documents or ask the model to perform multi-step analysis. Lower internal usage accounting may allow teams to use Luna for routine tasks instead of reserving every request for a more expensive model.

Availability in Europe and compliance considerations

GPT-5.6 Luna, Terra and Sol are available to European developers through the OpenAI API, with access also provided through relevant ChatGPT, Codex and ChatGPT Work plans. Availability can still depend on the customer’s account type, subscription and regional rollout status.

For EU companies, lower prices do not remove compliance responsibilities. From August 2026, legally binding EU AI Act enforcement will require stronger transparency measures, including clear notifications when users interact with certain AI systems and detectable, machine-readable labelling for AI-generated content. Organisations deploying Luna inside customer support, document processing or agentic workflows should therefore treat the price cut as a budgeting change—not as a substitute for governance.

Data handling also remains a separate question from model pricing. European buyers should examine retention settings, contractual terms, access controls and data-residency options before sending personal or confidential information to an external API. A cheaper token is not automatically a compliant processing arrangement under the GDPR.

A sharper price-performance contest

OpenAI’s move arrives during an unusually competitive period. Anthropic’s Claude Opus 5, Google’s Gemini 3.6 Flash, open-weight models from Meta and DeepSeek, and European offerings from companies such as Mistral AI are all competing on different combinations of capability, speed, price and deployment control.

The practical result is that model selection is becoming less about choosing one universal winner. A European business may use a low-cost hosted model for classification, an open-weight system for sensitive workloads and a frontier model for complex reasoning. Luna’s new price makes that kind of tiered architecture more attractive, particularly for applications with large volumes of routine agent activity.

OpenAI’s message is clear: improvements in serving efficiency can now be passed directly to developers. For teams building AI products in Europe, the immediate calculation is simple. GPT-5.6 Luna delivers the same nominal capability tier at one-fifth of its previous token price—but businesses still need to test quality, latency, data handling and regulatory fit before putting the savings into production plans.