GPT-5.6 cuts the cost of AI agents

OpenAI has introduced GPT-5.6 as a family of models capable of carrying out long-running tasks and using tools at a much lower cost. The key is not just the model itself. It is also how it preserves previous work, distributes tasks, and processes data outside the context window.

The company explains in a guide based on tests with startups that GPT-5.6 can match or outperform earlier models while using fewer tokens, meaning fewer units of processed text. For anyone building AI products, that could mean cheaper agents to operate without having to rebuild their entire architecture.

Better performance with less reasoning

One improvement is choosing more precisely how much reasoning each task requires. In the Agents’ Last Exam benchmark, GPT-5.6 Sol with a low reasoning level outperformed GPT-5.5 with a high reasoning level, while using the same tool system and rules.

The cost difference is clearer in BrowseComp, a test that measures the ability to find uncommon information through searches. According to OpenAI:

• GPT-5.5 with Extra High reasoning achieved 84.36% and cost $33.27.
• GPT-5.6 Luna with the same level achieved 84.04% and cost $1.33.

The comparison reflects the prices and results reported by OpenAI at launch. The company also notes that it later reduced prices even further.

The family’s smaller models, such as Luna and Terra, are designed for frequent or latency-sensitive tasks. A legal company, for example, could use one of them to extract information from handwritten memos and reserve a more powerful model for the complex analysis that follows.

Three changes for building more efficient agents

OpenAI is also adding new functions to the Responses API, the interface used to create applications that can reason and use tools.

First, the agent can preserve its reasoning. Memory of the work completed between turns and native compaction, which summarizes long conversations, keep the model from having to reconstruct everything from scratch. This helps with tasks that run for a long time.

Second, it can divide the work among multiple agents. A primary agent coordinates others working in parallel and then combines their results. A complex search, for example, can be split by source, time period, or document type instead of being run as a single chain of steps.

Third, repetitive data is handled by code. With programmatic tool calling, the model can write JavaScript to filter, sort, and combine results outside its context. That means that if it needs to review 100 documents and keep only those from a specific date, it does not have to analyze every intermediate result as if it were a new decision.

The model is reserved for tasks that require judgment: interpreting information, detecting relationships, and deciding what to do next.

An example of the impact

On ARC-AGI-3, GPT-5.6 Sol scored 13.3% with the standard system. After persistent reasoning and compaction were enabled, the score rose to 38.3% while using approximately six times fewer output tokens.

The model was not changed. The way the work was organized was. The result shows that an agent’s performance depends as much on its architecture as on the model it uses.

OpenAI also extended the instruction cache duration to a minimum of 30 minutes across the entire family. The cache stores repeated parts of a request so they do not have to be processed again, which can reduce response time and cost in applications that repeat the same context.

For you, the most important change is practical: you do not always need to use the most powerful model at every step. An application can combine smaller models for extraction and repetitive tasks, adjust the reasoning level, and reserve larger models for difficult decisions.

The next question will be how much of these improvements holds up outside benchmarks and in real products. If the startup results are repeated, building AI agents will no longer be about paying for a powerful model on every call. It will depend much more on designing the workflow well.