GPT-5.6 optimises efficiency. Citation behaviour may follow.

Efficiency, in the context of large language models, is not a neutral word. OpenAI's blog post on GPT-5.6 frames the release as a marriage of "frontier intelligence with frontier efficiency," but the more consequential claim is buried in the subtext: the model is designed to deliver more useful output per unit of inference cost. That shift in design priority has direct consequences for which content gets cited and which gets skipped.

The mechanism matters. When inference is expensive, models have more latitude to retrieve broadly and reason at length. When efficiency is the governing constraint, retrieval becomes more selective. The model optimises for answer quality at lower compute cost, which means it needs to be more confident in what it surfaces. Sources that are already trusted, already cited elsewhere in the training distribution, and already structured in ways that reduce the model's interpretive work become more valuable, not less.

This is not speculation about a future state. It follows from how efficiency-oriented inference works across agentic and non-agentic contexts alike.

What "useful intelligence per dollar" actually selects for

The phrase "more useful intelligence per dollar" is doing heavy lifting in the OpenAI announcement. It describes a model that has been optimised to resolve queries with fewer computational steps. In practice, that rewards content with two properties: high prior citation frequency (the model has seen it cited reliably before) and low ambiguity (the model does not need to spend tokens reconciling conflicting framings).

For B2B brands, particularly in financial services, policy institutions, and multilateral bodies, this creates a specific challenge. Their content is often technical, frequently updated, and written for narrow expert audiences. It does not accumulate citations the way that, say, a widely-shared explanation of a consumer concept does. An institution like the World Bank's CGAP or a standard-setting body like ISO produces authoritative material that is less likely to be cited by a general web population, even if it is definitively correct. Under efficiency-optimised retrieval, the model skews toward content that has already demonstrated citability, creating a compounding advantage for sources that are already prominent.

The implication is direct: brands and institutions whose content lives in specialist databases, behind paywalls, or in PDF formats that are poorly indexed will find their LLM visibility degrading faster under efficiency-oriented models than under brute-force ones. Compute-heavy models can afford to go looking; efficient ones do not.

Agentic workflows raise the stakes

GPT-5.6 is also positioned explicitly for agentic use cases, where the model does not answer a single prompt but executes multi-step tasks with real-world consequences: drafting proposals, running research briefs, populating reports. In these workflows, citation behaviour compounds. A source consulted in step one shapes the framing of steps two through five. An organisation absent from early retrieval does not recover in later steps.

For communications leaders at large industrial groups or financial institutions, this is the sharper concern. A procurement officer using an AI agent to assess a supplier's track record on sustainability, or an analyst using an agent to compile a regulatory briefing, will produce outputs shaped by what the model retrieves first. Efficiency-optimised retrieval means those first-pass sources are drawn from a narrower pool of high-confidence material.

The brands that built credibility in traditional search by publishing volumes of indexable content will retain an advantage. Those that relied on human intermediaries, conference presentations, or proprietary research channels to build authority will find that advantage does not translate.

The practical consequence

GPT-5.6 is not a radical departure from GPT-5, and OpenAI's post makes no specific claim about citation methodology. What it does claim is that the model has been rearchitected around efficiency as a first principle. That architectural choice does not need to be accompanied by an explicit policy change on citations to produce one in practice.

The brands that will hold their LLM visibility through this generation of efficiency-oriented models are those that have already made their content easy for a model to trust quickly: clearly attributed, published on domains with high existing citation rates, structured so that the key claim surfaces in the first sentence rather than the fifth paragraph. In an efficiency-constrained retrieval environment, the cost of burying the lede is no longer just a reader engagement problem. It is a visibility problem.