Built in
Unlocking agentic value
EY (Smith, Wanner, Butler, Diasio). Whitepaper 1 of EY's "Total Cost of Agents" series, 1 June 2026. ey.com
What it found
Tokens are one of seven cost categories in running an agent; most business cases count only the first three. A support chat that cost $0.04 in 2023 can cost $1.20 as an orchestrated agent. It calls for "circuit breakers": spend ceilings and automatic shut-offs per agent, workflow and business unit, with a named owner for each cost before it is spent.
What CostGrid does
Spending limits per agent, team or everything, per day, month or run, checked before each call sent through CostGrid. A stop button for any agent or team. An owner and reason on every rule. Non-token costs entered under EY's categories, so the monthly statement shows a full cost.
App: Limits · Results › Full cost
Built in
How Do AI Agents Spend Your Money?
Bai, Huang, Wang, Sun, Mihalcea, Brynjolfsson, Pentland, Pei. University of Michigan, Stanford, MIT and others. April 2026. arXiv:2604.22750
What it found
Agentic tasks use roughly a thousand times the tokens of a single question, mostly as input. Runs of the same task differed by up to 30 times. Spending more did not reliably buy accuracy, and models could not predict their own cost beforehand (correlation of 0.39 at best).
What CostGrid does
It does not try to predict cost. It enforces a ceiling on each run before the next call is sent, shows each agent's typical and bad run (median and 95th percentile) instead of an average, and warns while a run is still going when it passes three times what that agent's runs usually cost.
App: Limits › per-run limits · Results › What a run costs · Alerts
Built in
The Price Reversal Phenomenon
Chen, Zhang, He, Stoica, Zaharia, Zou. Stanford, UC Berkeley, CMU, Microsoft Research. Version 2, May 2026. arXiv:2603.23971
What it found
In 32% of model pairs tested, the model with the lower list price cost more in total, by up to 28 times, because it used more reasoning tokens or more turns. Repeats of the same query varied up to 9.7 times in reasoning tokens. It recommends tracking actual cost per request.
What CostGrid does
Every call is priced from the provider's own token counts, not a list price. When a rule moves traffic to a cheaper model, CostGrid measures cost per call and per whole run before and after on your own traffic, and says plainly if the switch made things more expensive.
App: Limits › Did switching save money? · Results › Reasoning
Built in
The Illusion of Thinking
Shojaee, Mirzadeh, Alizadeh, Horton, Bengio, Farajtabar. Apple. 2025. arXiv:2506.06941
What it found, and how we use it
On simple problems, models without extended reasoning were often as accurate as reasoning models while using fewer tokens. We cite it only for that narrower point: past a certain level, more reasoning adds cost without reliably adding accuracy.
Its stronger claim, that reasoning "collapses" on hard puzzles, was disputed because of output-length limits and puzzle instances that could not be solved (arXiv:2506.09250, arXiv:2507.01231). We don't rely on it. Newer work supports diminishing returns, not collapse (arXiv:2608.16956, arXiv:2604.05164).
What CostGrid does
A reasoning cap: a limit on the effort level or thinking budget a request may ask for. Spend broken down by how much reasoning each call asked for, and a flag on agents that write short answers after long reasoning, the first place to try a lower effort level.
App: Limits › How much the model may reason · Results › Reasoning
What it found
API bills and simple totals miss the lifecycle cost of AI. Like the levelized cost of electricity, the fair comparison is total capital and running cost divided by the number of valid outputs.
What CostGrid does
CostGrid estimates which jobs finished from how each agent behaved (indicative), or your software can report it exactly; CostGrid then shows cost per finished job and the money spent on jobs that didn't finish, and full cost per success once other costs are entered. Partly: it uses monthly entered costs and does not amortise capital spend. LCSO, below, is how we extend it.
App: Results › What the money bought · Full cost
Our method · draft
Levelized Cost per Successful Outcome (LCSO), v0.1 draft
CostGrid. Version 0.1, 3 October 2026. Licensed CC BY 4.0: anyone may use and adapt it, with credit. Builds on LCOAI (Curcio, arXiv:2509.02596).
What does one successful business outcome from an AI workflow actually cost, all-in? LCSO is CostGrid's answer, written so a finance team can budget against it and an auditor can re-perform it. It uses metadata only: identifiers, timestamps, usage counts, charges, outcome states and time records. No prompt or answer content.
A working draft. It has not yet been reviewed by finance, audit or academic readers, and it will change.
What changes from LCOAI
| LCOAI | LCSO |
|---|
| What it divides by | Valid inferences | Successful business outcomes, judged against a written, versioned definition with evidence |
| Failed work | Not treated separately | Charged in full to the successes, and reported as the Failure Cost Share |
| People's time | Not included | Review, rework and hand-off time, as a separate layer |
| Variability | Not measured | Cost spread per task (p50, p90, p99) next to the average |
| Granularity | A deployment scenario | Each workflow and period, from live records |
Three layers, always reported together
LCSO-1 · meteredModel charges ÷ successes. Fully measured, and can be tied to the provider's invoice.
LCSO-2 · operationalAdds other metered tools and the human time the AI caused: reviewing, fixing, and handing off failures.
LCSO-3 · levelizedAdds fixed and platform costs: subscriptions, infrastructure, upkeep, with the allocation key stated.
Measured and estimated figures are kept apart and labelled. Human time from a time study is an estimate and is never blended silently into a measured number.
LCSO does not detect hallucinations; it measures their cost.
Wrong answers show up economically as failures, retries, rework and reversed successes. The Failure Cost Share is the share of model spend that did not directly produce a success: failed, abandoned and unknown tasks, plus the retries inside tasks that eventually worked.
The savings rule
A claim that a change lowered LCSO counts only if all of these hold:
- Tasks are randomly assigned to the old setup or the new one (a holdout), on the customer's own tasks. Before-and-after comparisons are labelled observational.
- Both groups use the same success definition and period.
- The result has a confidence interval, bootstrapped over tasks.
- The success rate stays within a non-inferiority margin declared in advance, such as −1 percentage point.
- The metric, margin, sample size and duration are fixed before the test starts.
Worked example · illustrative numbers
Tier-1 support ticket resolution, September 2026. Success means closed by the agent and not reopened within 7 days. 9,800 tickets resolved, 7,800 successes. Human time is estimated from a time study at $45 an hour. These are made-up numbers to show the arithmetic, not a customer's results.
| Metric | Calculation | Result |
|---|
| LCSO-1 | $3,040 model charges ÷ 7,800 | $0.390 |
| LCSO-2 | ($3,040 + $4,824 human time, estimated) ÷ 7,800 | $1.008 |
| LCSO-3 | ($7,864 + $4,000 fixed) ÷ 7,800 | $1.521 |
| Failure Cost Share | ($620 escalated + $210 abandoned + $330 retries) ÷ $3,040 | 38.2% |
| Success rate | 7,800 ÷ 9,800 | 79.6% |
| Human-Equivalent Ratio | $1.521 ÷ $6.75 (a person: 9 minutes at $45 an hour, time study, estimate) | 0.23 |
In this example, model charges are only about a quarter of the all-in cost of each outcome. Failures and people's time are where much of the rest goes.
Limitations of v0.1
- Success is yes or no. Quality is not weighted yet.
- Success definitions are specific to each organisation, so LCSO compares a workflow with itself over time, not one company with another.
- Human time is often estimated. Estimates stay out of LCSO-1 but drive LCSO-2 and LCSO-3.
- It measures cost, not the value of the outcome.
- Loosening a success definition lowers LCSO. Versioning and sampling successes back to their evidence reduce this risk but don't remove it.
- Every AI request needs task and run identifiers. Where they are missing, spend is reported as unattributed.
Shapes
Tiered Super-Moore's Law
Du. Wuhan University. March 2026. arXiv:2603.28576
What it found
Token prices fell about 600-fold from 2020 to 2026. Cheap models halve in price every 1.1 years and mid-tier every 1.55, but flagship prices barely fall, because reasoning models charge on average 31.5 times more per token.
Caveat: per-token price trends miss two things that drive bills. Newer tokenizers count the same text as more tokens (about 30% more for Claude 4.7 and later, per Anthropic's pricing page), and agentic work uses more tokens per task. Good for list prices, weak as a guide to what you'll pay.
What CostGrid does
Its price half-lives drive the 2023–2030 forecast in the research workbook. In the product, spend is broken down by model tier, and reasoning spend is tracked separately, because the tier a call runs on matters more than the falling headline price.
Workbook · App: dashboard › tiers · Results › Reasoning
Context
Tokenomics
Citadel Securities, Global Macro Strategy. June 2026. Secondary reporting only: the primary note is not public. WebProNews report
What it reportedly found
As reported, buyers are becoming sensitive to the all-in cost of AI, token price times token volume, and are moving work to cheaper models wherever frontier capability is not needed, so "frontier" and "everyday" AI use keep splitting apart.
Secondary reporting. We could not check the details against the original note, so treat them as unverified.
How it is used
Background only. The direction matches Gartner's August 2026 forecast that inference cost per agentic workflow will rise more than fivefold through 2028, which is why CostGrid reports price times volume per job rather than per-token list prices.