AI Frontier Institute

Assurance

How well do AI models do ordinary business work, what does it cost, and how often do they get it wrong? We are building a public test program to answer that on real tasks, with a published method and cross-vendor grading.

Status: stage one, demonstration run

Stage one ran the five public tasks three times on each of four models. It is a demonstration of the method, not a finding: five tasks are too few to rank models, and the grading has not been checked against people. The held-back tasks run in stage two.

Why this exists

California enacted two laws in September 2026 that create independent verification organizations and a registry of AI auditors. Colorado's own automated decision law takes effect on January 1, 2027. Businesses are being asked to vouch for AI tools they have little independent evidence about. Useful, public test data can exist before any certificate does, and that is where we start.

Details on both states' laws are in the California and Colorado policy tracker entries.

What we test

Business tasks only. No trivia, puzzles or general knowledge. Ten tasks across five kinds of work, two each. One task per kind is public, and its answer key is kept on file and available on request. The other will stay private until the first report, and a fingerprint of it will be published before stage two runs, so it cannot be changed afterward.

What we measure

Accuracy
The share of runs that pass every required check, and the share of checks passed, so a nearly right answer is visible.
Cost
Actual billed tokens times the vendor's published price on the run date, per passing run. Stage two will add cost per task and will report human review time separately.
Speed
Wall-clock time from request to full response. Stage one shows the median; stage two will add the range.
Consistency
Every task runs three times per model with identical settings. Stage one shows how many of the three passed; stage two will publish all three scores and their spread.
Failure modes
Each failed run is labelled by what went wrong. Stage two will use the full set of labels: missed field, fabricated fact, wrong number, stale clause, format failure, failed test, refusal, timeout or truncation.

How cross-vendor grading works

What we will publish

In stage two we will publish the test harness, scoring scripts, public tasks and answer keys, prompts and settings, the exact model identifiers the vendor returned, run dates, raw scores, token counts, cost, speed and failure labels. Held-back tasks will be released after the first report if that does not compromise a repeat test. Until then, these files are kept on file and available on request.

Models in stage one:

The model name each vendor reports back is recorded with every run. Codex does not report one, so OpenAI runs record the model that was requested.

Stage one demonstration results

Five public tasks, three runs each, run on 2026-09-24. A run succeeds only if every required check passes. Read these as a test of the method: with five tasks, one task can move a model's total by a fifth.

ModelRuns that passedChecks passedMedian timeCost per passing runGraded items passed
claude-opus-5-515 of 15100%4.2 s$0.015721 of 21
claude-sonnet-515 of 15100%3.6 s$0.006318 of 21
gpt-6-astra15 of 15100%13.5 snot comparable18 of 21
gpt-6-sol15 of 15100%10.7 snot comparable18 of 21

OpenAI times include Codex start-up and OpenAI costs are not comparable; see the disclosures below. Graders agreed on 21 of 21 items for claude-opus-5-5, 19 of 21 items for claude-sonnet-5, 20 of 21 items for gpt-6-astra, 20 of 21 items for gpt-6-sol.

Every model passed every required check on every run. These five public tasks are too easy to separate the models, and stage two's held-back tasks need to be harder. Graded items and times cannot be compared across vendors in this run, because each vendor was graded by a different pair of models, the graders changed some verdicts when re-grading unchanged outputs, and OpenAI times include Codex start-up. Within a vendor they can be compared, because both models shared a grader pair: claude-sonnet-5 against claude-opus-5-5, and gpt-6-sol against gpt-6-astra.

Grading corrections

The first scoring failed 6 of 60 runs. On inspection those failures were faults in our grading, not in the models: a check that flagged the correct phrase “not a guaranteed delivery date” as a promise, and a meeting task whose instructions and format asked for different wording. Separately, one Claude-side grader was weaker than the rest and failed correct work its partner passed, so it was replaced. All of this was fixed, every model was rescored against the same corrected keys, and the first version's scores are kept on file and available on request.

The grader swap was made after the results were seen. The inspection that found these faults was done by the builder, a Claude model, not by a person.

Graded items passed and grader agreement, first scoring against the rescoring after the grader swap.
ModelGraded items passed, firstGraded items passed, rescoredAgreement, firstAgreement, rescored
claude-opus-5-521 of 2121 of 2121 of 21 items21 of 21 items
claude-sonnet-518 of 2118 of 2119 of 21 items19 of 21 items
gpt-6-astra16 of 2118 of 2116 of 21 items20 of 21 items
gpt-6-sol15 of 2118 of 2117 of 21 items20 of 21 items

Runs that passed, by task

Taskclaude-opus-5-5claude-sonnet-5gpt-6-astragpt-6-sol
contract extraction3 of 33 of 33 of 33 of 3
email drafting3 of 33 of 33 of 33 of 3
meeting summary3 of 33 of 33 of 33 of 3
spreadsheet analysis3 of 33 of 33 of 33 of 3
business coding3 of 33 of 33 of 33 of 3

How runs failed

Where this fits

Others already test AI on professional work. Vals.ai publishes scores, cost and latency on legal and finance tasks. Scale's READY measures enterprise workflows, including the cost of human review. OpenAI's GDPval uses blind experts to grade professional work. Artificial Analysis sells custom task testing. We learn from all of them.

What we add is a public-interest program, independent of any model vendor, that tests everyday small-business work and reports how models fail as well as how often they succeed.

Disclosures