The promise of an always-on agent is simple. You hand it your coverage list, your open questions and your recurring work, and it keeps going while you sleep. It follows the evidence, updates your thesis, and tells you what changed before you think to ask.
That future is arriving fast. Agents now run 99% of the searches on AlphaSense. In October, we were processing tokens at an annualized rate of roughly 75 trillion, and that number keeps climbing as users hand more of their work to SuperAnalyst. Across the industry, the move from chatbots to multi-step agents has raised the tokens consumed per task by up to 30x.
An agent that never stops working never stops spending. How can an always-on system work at all hours without growing its bill exponentially?
There are two ways to make an agent cheaper. You can give it a smaller model, or you can make it more efficient. A smaller model cuts the bill by cutting the intelligence, and users pay for that in the quality of every answer. We take the second route, and it's the system around the model that enables SuperAnalyst to stay on the frontier.
Why Price Per Token Is Misleading
Token prices are the number everyone quotes, but they hide most of what you pay. An agent works in hops: plan, search, read, reason, search again. On every hop, the model is sent the entire conversation so far, so you pay for that context again each time.
Take an illustrative 10-step research task that starts with 20K tokens of instructions and adds 5K tokens of new evidence per step:
The agent reads 65K tokens of unique material and is billed for 425K. About 85% of the bill is the same context, re-sent. Errors compound the same way. If each step is right 95% of the time, a 10-step chain is fully right only about 60% of the time.
Rather than cost per token, the metric that matters is cost per correct answer: every token in the research loop, weighted by whether the answer holds up in front of an investment committee. A model's list price, in other words, says little about what an answer costs, because of how much the system around a model affects the quality of answers.
Most AI research setups today are assembled from parts: a general-purpose model and a few data feeds connected through MCP. Frontier models are extraordinary reasoners but mediocre searchers. They don’t know which sources are authoritative, and embedding distance can't tell an earnings transcript from a news article. The models don't know when to stop, so they compensate with volume, which drives up costs and leaves blind spots.
AlphaSense is built the other way up. We create, license, index, tag, and host the content. Then we build the search indexing, the context graph, the tools, and the agent harness on top of it, and tune them all for the same goal: the best answer at the lowest cost per correct answer.
We found that a frontier model doing its own retrieval over our content with raw vector search, a good proxy for a typical market-data MCP, costs about 3x more per question than the same model running in our harness with AlphaSense Search. The AlphaSense answers are also preferred more than 2:1, even when we give both setups our proprietary content. Relying primarily on the open web is worse still: it costs 2.2x more on average and is 3.6x less citable.
Why We Keep Users on the Frontier
Shrinking the model is a tempting route. A smaller model (ideally one fine-tuned for finance) is cheaper to run and may look competitive on the day it ships. But it is a snapshot of the frontier at the moment it was trained. Frontier models keep getting better at reasoning, writing code, analyzing data and building deliverables, and a fine-tuned model has to be retrained to keep up.
Our efficiency lives in the system around the model, so it carries over to every new frontier model we adopt. Users' skills, memory and workflows stay in place when the model underneath changes.
The Five Levers We Pull
In order to drive efficiency and keep SuperAnalyst on frontier models, we pull five levers. Most of these only work because of our full stack solution, integration from the content to the agent.
- Retrieval: Finding, checking and assembling evidence drives more than 80% of the cost of a SuperAnalyst task and up to 70% of its quality. Our search index, tuned on tens of millions of searches, returns fewer, better results with relevance, freshness, and authority scores, so the model knows when to stop.
- Pre-compute once, for everyone. Supply chain relationships, key debates, channel check trends and financial performance live in our context graph, built once and kept current. SuperAnalyst spends inference only on the delta: the new question, the new document. Updates are pushed when they land, so agents stop polling for changes.
- Tools first, documents when needed. One tool call returns a company's pre-computed supplier and customer map, each link sourced. One call can replace thousands of tokens of reasoning.
- Write a workflow once, run it many times. A skill captures the method, sources, and template, then follows the most efficient path on every scheduled run. Because the data in AlphaSense is kept up to date in real time, the agent only needs to act when something changes.
- A frontier lead agent with a frontier research agent beneath it. The lead agent stays on the frontier model, and our research agent assembles the evidence for it. We specialize where our domain makes the biggest difference: finding and assembling evidence from our content with our tools. Our research agent is itself frontier intelligence for context retrieval. It finds better evidence in fewer turns and with lower latency than a general model searching on its own. That leaves the frontier model with more of its budget for reasoning, coding, data analysis, and building rich deliverables.
What SuperAnalyst Buys You
These design choices show up in the economics of real work. SuperAnalyst’s cost varies with the assignment and the work required to complete it. In the sample shown here, the median cost per run for each of our 10 most-used skills was below the Deep Research benchmark (an AlphaSense feature our users have leveraged at scale for more than a year). Individual runs vary, as the ranges show, but the pattern is consistent across workflows from company research to competitive mapping and earnings analysis. SuperAnalyst carries complex assignments through research, reasoning, code execution, and a finished deliverable, at a cost low enough to scale.
And the work holds up. Sasha Miagkyi, a former banking analyst on our team, used SuperAnalyst to build a model and a 20-page take-private discussion deck in 51 minutes, with nine editable charts, native tables on 14 slides and 47 cited sources. He judged it about 95% of the way to presentable to a VP.
The Real ROI of SuperAnalyst
Sasha’s deck is one example of a broader pattern. Recurring work that takes hours or days by hand comes back as analyst-level deliverables, and users describe the output by comparing it to the people they would trust with it.
“[The first draft SuperAnalyst sent me] is what I would have expected of a McKinsey business analyst that I’ve had for two years.”
“Some things, like the clinical trial predicting tool, I would say are as good as a managing director. It’s really good. It can predict a clinical trial, and I think the way I’ve built it, it does so as well as most humans. I’m pretty confident in saying that.”
That is where the return builds. When the first draft is that strong, analysts spend their hours on judgment, follow-up questions and client conversations instead of assembly. Teams can cover more names, follow more questions to the end, and turn work around when a deal or an earnings print can’t wait.
“Some of the work we did with SuperAnalyst last week, we did it in a day, when it would have taken a week before.”
What This Means for You
An agent that keeps working while you sleep only pays for itself if each task costs less as you hand over more of them. If your AI bill is growing faster than the value you're getting, take a look at the system around it and how efficient it is. The cheapest token is the one your agent never has to spend.
That is why we have concentrated on the system: We tune retrieval so the model gets fewer, better results and knows when to stop. We pre-compute what’s shared, so no one pays to build the same context twice. The agent reaches for tools first and documents only when it needs them. You write a workflow once and it runs many times, acting only when something changes. And the lead agent stays on the frontier model, with our research agent assembling the evidence beneath it, so we never trade intelligence for margin.
The more research you entrust to the system, the more you benefit. That is the economics of delegation, and it is how SuperAnalyst can work at all hours without its bill growing at the same pace.
