Lovelace boasts it can equal Gemini Deep Research at less than 1% of the cost

AI startup Lovelace has released a benchmark of 12 financial and business research tasks on which it says its agent achieved parity with Google’s Gemini Deep Research Max at “less than 1% of the cost.”

To carry out the test, the company says it built an investment banking research agent on top its context engine, YottaGraph, and Gemini 3.1 Flash Lite, with no grounding beyond the YottaGraph itself.

The benchmark, a release stated, “evaluated a range of complex financial and business questions, including company comparisons, acquisition scenarios and investment analyses. Reports were judged based on factual accuracy, analytical rigor, use of evidence and citation quality.”

The experiment was designed to answer this question: Can an agent powered by a lightweight LLM hooked up to Lovelace’s YottaGraph, and nothing else, provide deep, research-grade reports significantly faster and more cheaply than a flagship Deep Research model?

Looming crisis due to AI cost

According to a company blog post, it could: “Across 12 investment banking topics judged on a six-dimension, 1–10 rubric, it scored 9.67 mean versus 9.87 for Gemini Deep Research Max (3.1 Pro), at roughly six cents per report instead of seven dollars, and in under five minutes instead of 17 minutes.”

When asked what prompted the decision to embark on the benchmark, Andrew Moore, CEO of Lovelace, and the former head of Google Cloud AI, replied via email, “we know that there is a crisis looming because of the expense of AI.”

The Foundation Labs such as Google, OpenAI and Anthropic are, he said, “racing so fast towards more powerful AI that they, perhaps quite reasonably, are ignoring the expense in order to hit the finish line first. But if society is going to start deploying AI usefully, we feel it must be possible to do so without building thousands of new nuclear reactors and data centers.”

Moore predicted that context, not compute, will define the next generation of AI systems “because LLMs are closer to human cognition than traditional computer science. And they are being asked to do the equivalent of asking a human to hold thousands of things in their working memory while they are reasoning.”

The more things there are in working memory simultaneously, the massively more expensive things become, and that’s the context problem, he said, adding, “solve the context problem and you solve the cost problem.”

Carmi Levy, an independent technology analyst, said that he has long been uncomfortable with the industry’s “arms race-like mentality toward AI dominance,” noting, “Lovelace’s benchmarks suggest the industry may have been focusing on the wrong thing all along.”

Efficiency, he pointed out, has been almost invisible in the industry’s rush toward an AI-enabled future: “Bigger, ever more capable models have grabbed the headlines as vendors fight for bragging rights. And while size certainly matters in terms of any AI model’s ability to effectively crunch massive workloads, we seem to have ignored the costs of all that capability, and whether those costs are even worth it.”

Levy pointed out that someone would not use a sledgehammer to tighten a loose fitting on the front porch. “Instead, we’d choose a smaller, more effective tool to do the job,” he said. “The same logic applies in AI, and while the first few years of the AI era have been almost uniquely focused on the biggest, most powerful tools, we’ve failed to match the costs of all the capability to the underlying business issues that are being addressed.”

Enterprises bringing sledgehammers to every engagement, he said, “are probably significantly overspending relative to organizations that use tools specifically sized to a given business need. As metered AI use becomes a more mainstream reality in the enterprise, efficiency will need to become a priority.”

All of this, Levy added, “is as important for individual enterprises looking to control runaway AI compute costs as it is for vendors building the necessary data center and energy infrastructure to power it all. It’s also critical for governments to ask about efficiency before green-lighting an endless lineup of massively scaled infrastructure.”

Time to redesign AI models

Sanchit Vir Gogia, chief analyst at Greyhound Research, described the benchmark as being “meaningful, but not in the way the market will read it. It is not a clean win for small models over large ones. The defensible reading is narrower: for bounded, evidence-heavy research, the architecture around a model can compress the cost of a good answer by two orders of magnitude without degrading it.”

The industry, he said, “has spent three years acting as though intelligence lived inside the model, reaching for a bigger one whenever quality disappointed. That reflex is now meeting enterprise economics with all the grace of a grand piano pushed down a staircase. A brilliant model fed poor grounding is still a very expensive guesser.”

But financial research, said Gogia, “runs on filings and entities that behave like a graph. It is a graph-shaped problem wearing a research suit. The first question is no longer which model is most capable, but which system answers reliably at a defensible cost. Capability, in the enterprise, is now a property of the system, not the model.”

CIOs and CTOs, he added, “should treat the claim as a prompt to redesign their AI operating model, not as an instruction to buy a product. AI value is now determined by four connected disciplines: the quality of the context supplied, the routing of work to the appropriate model, the governance of the workflow, and the observability of what the system costs and does.”

He pointed out that an agent given poor context does not merely return a weak answer; it takes weak action, which is an operational hazard.

“Buyers should demand the cost of a completed workflow, and ask how portable the architecture is, since context lock-in is still lock-in,” he said. “The CIO’s task is to design the safest path from evidence to outcome.”