GPT-5.6 от OpenAI: ставка на эффективность и больше интеллекта на доллар

OpenAI has introduced GPT-5.6 — an update to its flagship lineup, where the main bet is not on record-breaking benchmark scores but on efficiency: getting more useful intelligence out of every dollar spent. The improvements touched three links at once — the models themselves, inference, and performance in agentic scenarios.

What OpenAI Changed in GPT-5.6

GPT-5.6 is optimized along three directions that OpenAI lists in its blog: the architecture of the models themselves, inference (the process of generating a response), and agentic workflows — multi-step tasks where the model is called dozens of times in a row. The common denominator across all three links is cost: the cheaper each one is, the more tasks can be solved for the same money.

• Update to the flagship lineup — the GPT-5.6 model from OpenAI
• Three axes of optimization: the models themselves, inference, agentic workflows
• The key metric of the release — "useful intelligence per dollar" (intelligence per dollar)
• Source of the statements — OpenAI's official blog

What "Intelligence Per Dollar" Means

"Intelligence per dollar" is the ratio of a model's answer quality to the cost of obtaining it, and it is precisely this metric that OpenAI puts at the center of the GPT-5.6 release. Previously, progress was measured mainly by benchmark scores; now, increasingly, what matters is how much a unit of that quality costs. For business, the metric is direct: the same task can cost noticeably less if the model spends fewer computations to get the same result.

"GPT-5.6 improves AI efficiency at the level of models, inference, and agentic scenarios, helping deliver more useful intelligence for every dollar invested,"

OpenAI's announcement states.

Where the Savings Are Hidden

The main savings in GPT-5.6 are concentrated in inference — the stage at which the model actually generates the response. It is inference, not training, that every user pays for with every request, so making it cheaper directly hits the final bill. OpenAI states that it worked on efficiency both at the level of the models themselves and at the level of how they are executed — meaning the cost reduction comes not only from a "lighter" model but also from inference engineering.

Why Efficiency Matters for Agents

Agentic scenarios are the main beneficiary of GPT-5.6's efficiency gains. Unlike a single question-and-answer exchange, an agent solves a task in several steps: it plans, calls tools, checks the intermediate result, and repeats the cycle. Each such step is a separate call to the model, and over a long chain the cost adds up. When inference gets cheaper, the savings multiply by the number of steps, which is why multi-step agentic workflows are the first to feel the difference — both in cost and in speed.

Why This Is a Move Toward the Mass Market

The bet on efficiency, rather than solely on "smarter at any price," reflects a shift in the phase of AI development. While a model is expensive, only a few use it; when the cost per unit of intelligence falls, scenarios that previously made no sense become viable — from constantly running agents to processing large data streams. GPT-5.6 is presented precisely from this angle: not "the highest score," but "more value for the same money."

What This Means

GPT-5.6 is a signal that the frontier is shifting from the race for scores to the economics of inference: what wins is not only the smartest model, but also the cheapest one per unit of useful output. For developers and companies, this means that agentic products, which just yesterday were too expensive to operate, are becoming more realistic.