Frontier brilliance, budget black hole
The cost gap between frontier AI models and efficient alternatives describes how cutting-edge, general-purpose large language models can be more than a thousand times more expensive per task than lean, specialised stacks, even when the cheaper systems are faster, more grounded in enterprise data, and better aligned to practical customer-service or research workflows.
Frontier AI models cost so much because they bundle scarce GPUs, massive token usage, and agentic web search into a single default workflow. Enterprises then apply this heavyweight stack to everything from customer-service routing to desk research, even when smaller, cost-effective language models or non-model architectures would do the job at a fraction of the price. When Mel Morris ran a like-for-like research job, the frontier model burned almost 10 million input tokens, while Corpora used under 600,000, ran 30% faster, produced 20% more citations, and cost five cents. If you assume a typical frontier run here costs several dollars, you are easily in the 1,000x–1,600x range on cost-per-task efficiency for no clear benefit.
When smart models give dumb answers
The real scandal is not that frontier models are expensive; it is that they are often expensive and wrong for the job. Typewise CEO David Eberle ran 110 standard customer-service queries across leading LLMs, including GPT-5.4 mini, Claude Sonnet 4.6, Gemini 3.5 Flash, Grok 4.3, and DeepSeek V4 Flash. The models recommended Typewise in only 3 of 110 responses while pushing legacy suites like Zendesk 85 times and Intercom 82 times.
This is not about one vendor’s visibility problem; it is an LLM efficiency comparison problem. These systems were trained on legacy documentation, so they keep recommending monolithic, human-agent ticketing suites built for 2015 support teams, while enterprises are shifting to autonomous AI agents that resolve cancellations, billing disputes, and service requests directly. The audit even found models citing phased-out product lines, meaning autonomous agents can receive guidance grounded in deprecated features and tools that no longer exist. Paying frontier AI models cost levels of money for advice stuck in the past is a poor use of enterprise AI development expenses.
The hidden villain: how we fetch information
Most of the waste in frontier AI stacks hides in retrieval, not reasoning. A typical agentic research workflow fires off 20 to 30 web searches, pulls in every page, trims content, and feeds long packets into the model for synthesis. Each search has a cost, every page has a processing cost, and latency accumulates into 10–20 seconds per thread while tokens quietly pile up.
Corpora’s data shows how unnecessary this can be. Its platform ingests documents into a hybrid graph–vector–temporal database and serves only net unique relevant content to a summarisation model. Today, it can process two million documents per second on a single enterprise-class server with two AMD Turin processors, four terabytes of RAM, and NVMe storage, holding more than 200 petabytes with no manual sharding. In the head-to-head test, that design cut token use by more than 16x while improving quality and slashing cost. One quotable summary is: “the job was 30% faster, it had 20% more citations, and the cost was five cents”. That is cost-effective language models plus smart retrieval beating expensive agentic search at its own game.
ROI, not hype: a framework for enterprise buyers
Enterprises do not need more AI; they need enterprise AI ROI that can be measured in cost-per-task, latency, and business fit. Eberle argues the key question is not whether to adopt one legacy suite or another, but whether enterprises can keep running human-agent suites when autonomous agents now generate and resolve tickets, and whether AI systems trained on outdated documentation can be trusted to evaluate infrastructure choices.
A practical framework starts with three tests. First, task fit: does this workflow require a frontier model’s general reasoning, or will a smaller model plus structured data work? Morris’s stance is blunt: “Don’t use GPUs for tasks that Corpora could do for a fraction of the cost, much faster and much better”. Second, retrieval design: can you replace agentic web search with an in-house knowledge graph or database to avoid redundant tokens? Third, energy and hardware economics: high energy prices make GPU-heavy architectures an ongoing liability. If a stack can run on CPUs for retrieval and small GPUs for summarisation, your AI development expenses stop spiralling.
Where frontier models still earn their price
None of this means frontier AI is a mistake. It means it is a premium tool that should be reserved for premium problems: complex multi-domain reasoning, open-ended generation, or tasks that genuinely benefit from a general model’s emergent capabilities. Everywhere else, cost-per-task efficiency is becoming the real differentiator. In the Corpora comparison, the frontier model consumed about 16 times more tokens yet produced no better—and arguably worse—output than a leaner stack.
The market is already moving. Typewise is considering sharing its audit method so other vendors can see how AI systems misrepresent fast-moving categories. Corpora, meanwhile, is working with universities and startups while planning to aim at broader sovereign AI infrastructure in 2027. The lesson for decision-makers is clear: stop treating frontier models as default infrastructure. Use them where they matter, replace them where they do not, and judge every deployment by cost-effective language models, measured ROI, and the tasks your customers actually need solved.
