32 of 36 at $0.18 per 1,000 answers, at a median of 4.5 s. Haiku on medium got 33 at $0.38.
Experiments / Experiment 03
Which model writes meeting notes you can trust?
The best-value run passed 92% of answers at $0.38 per 1,000. The top run passed 97% at $2.54, about 6.7 times the price.
Their likely ranges overlap, so this test alone does not show a clear lead. What a likely range is ↓
Opens a new Prompt test with this experiment's prompt, checks and models. Paste yours over it. New here? Your email starts a free trial. No credit card needed.
- Best value
- 92% at $0.38 per 1,000 answers Claude Haiku 5.5, medium thinking. 33 of 36 answers passed, likely range 71–98%. One cheaper run scored about the same, but it came from the model that judged the answers, so we do not lead with it. It is in the table below.
- Top score
- 97% at $2.54 per 1,000 answers GPT-6.1 Sol, medium thinking. 35 of 36 answers passed, likely range 78–99%.
- One model, same setting, new prompt
- 31% → 81% 11 of 36 → 29 of 36 passed. Only the prompt changed.
The risk in meeting notes isn't a clumsy summary. It's a decision nobody made, or a task with an owner who never agreed. We tested eight models on invented meetings, where people change their minds.
A test case, not a customer. We built this on invented data and ran it on 4 November 2026. Models and prices change, and we don't rerun these often, so check the date.
What we do with prompts
This page is one run of it.
- StartYour prompt
The one you already use. Here, a plain first draft.
- 1 · ScoreChecks and test cases
6 checks and 24 test cases, written for the task.
- 1 · ScoreEvery answer scored
864 answers, each checked against every rule.
- 2 · CompareOther models
8 models, with cost and speed, to find the same result cheaper or faster.
- 3 · ImproveRewrite, score again
31% → 81% passed on one model, same setting (11 → 29 of 36).
We wrote the checks and test cases, and scored every answer
- 24test cases, 18 kept unseen for scoring
- 6checks on every answer
- 864answers scored, on 4 November 2026
- 69%the best any model passed on the first prompt (25 of 36)
The checks
Must-pass checks one miss fails the answer
- Lists as decisions only things the group actually decided, not ideas raisedmust pass
- Names an owner only if the transcript names one, and otherwise leaves it emptymust pass
- Gives a due date only if the transcript states onemust pass
- Valid JSON with summary, decisions, action_itemsmust pass
- Misses no action item someone clearly agreed to domust pass
- Reflects the final position when people changed their mindsmust pass
An answer passes only if it meets every must-pass check.
Summarise this meeting transcript. Return JSON with the keys summary, decisions and action_items, where each action item has task, owner and due.
Transcript: {{transcript}}The same test on other models: who won
35 of 36 at $2.54. Opus on low also got 35, at $9.14.
33 of 36 in 2.9 s at $0.48.
Hover over a point for the model and setting. Every run is in the table below.
All 16 runs
| Model | Thinking | First prompt | After one rewrite | Likely range | Cost per 1,000 answers | Median time | Thinking tokens |
|---|---|---|---|---|---|---|---|
| GPT-6.1 Sol | Medium | not run | 35 of 36 | 78–99% | $2.54 | 4.6 s | 28 |
| Claude Opus 5.5 | Low | 21 | 35 of 36 | 78–99% | $9.14 | 4.9 s | 59 |
| DeepSeek V4.1 Flash | Medium | not run | 34 of 36 | 74–99% | $0.55 | 6.2 s | 871 |
| Claude Haiku 5.5 | Medium | not run | 33 of 36 | 71–98% | $0.38 | 5.0 s | 368 |
| Mercury 3 | Medium | not run | 33 of 36 | 71–98% | $0.48 | 2.9 s | 1191 |
| GPT-6 Luna | Medium | not run | 32 of 36 | 67–97% | $0.18 | 4.5 s | 155 |
| GPT-6.1 Sol | Low | 25 | 32 of 36 | 67–97% | $2.24 | 4.5 s | 6 |
| Gemini 3.8 Flash | Medium | not run | 32 of 36 | 67–97% | $4.01 | 7.7 s | 862 |
| Claude Opus 5.5 | Medium | not run | 32 of 36 | 67–97% | $14.04 | 6.6 s | 294 |
| GPT-6 Luna | Off | 19 | 31 of 36 | 64–96% | $0.10 | 2.5 s | 0 |
| Gemini 3.8 Flash | Low | 21 | 31 of 36 | 64–96% | $0.83 | 3.0 s | 6 |
| Claude Haiku 5.5 | Off | 18 | 30 of 36 | 61–94% | $0.18 | 2.3 s | 0 |
| Mercury 3 | Off | 11 | 29 of 36 | 58–93% | $0.09 | 1.0 s | 0 |
| DeepSeek V4.1 Flash | Off | 24 | 29 of 36 | 58–93% | $0.11 | 1.3 s | 0 |
| Nex N2.5 Mini | Medium | not run | 22 of 36 | 39–80% | $0.09 | 4.1 s | 796 |
| Nex N2.5 Mini | Off | 15 | 13 of 36 | 18–59% | $0.02 | 1.7 s | 0 |
Likely range. Each run answered 18 test cases twice. With so few cases a score is a rough figure. The range is where the true pass rate probably sits, 95 times in 100. 15 of the 16 runs have a range that overlaps the top run's, so this test does not clearly separate them. The rest are behind by more than chance. Test your own prompt on more cases before you choose.
Sorted by passes, then cost. First prompt was run at the lowest setting only. Thinking tokens are per answer, as each provider reports them. Cost is the model's price for the answer, per 1,000 answers, on the test date. A code fence around JSON counts as valid, but anything that doesn't parse is a fail.
Does more thinking help?
Medium thinking lifted most models, by 3 to 5 cases for Haiku, Mercury, DeepSeek and Sol, and it cost Mercury and DeepSeek about 5 times as much. Nex rose to 22 and stayed the weakest. Opus went the other way, from 35 on low to 32 on medium, at a higher price. Haiku on medium came within 2 of Opus on low at about a twentieth of the cost.
Off or low Medium Cost change
We rewrote the prompt and scored it again
One model, same setting, only the prompt changed
11 of 36 → 29 of 36 passed
The biggest gain of the 8 models, marked “headline” in the list, on 18 cases the rewrite never saw, run twice. Same thinking setting both times, same cases.
The rewrite raised 7 of 8 models, and Nex N2.5 Mini did worse. Each at its lowest thinking setting.
First prompt Rewrite
Summarise this meeting transcript. Return JSON with the keys summary, decisions and action_items, where each action item has task, owner and due.
Transcript: {{transcript}}Summarise this meeting transcript. Return only valid JSON, without markdown or other text, with the keys summary, decisions and action_items. Use a string for summary and arrays for decisions and action_items, where each action item has task, owner and due. Use empty arrays when there are no decisions or action items, even if the transcript is incomplete or contains only small talk.
List as decisions only outcomes explicitly agreed in the transcript. Do not list suggestions, possibilities, conditional offers, unsupported statements of agreement, or the absence of an agreement or assignment as decisions. Put relevant context about unresolved matters in the summary instead.
Include every action item someone clearly agreed to do, but do not turn suggestions, unmet needs or unaccepted assignments into action items. Name an owner only if the transcript explicitly identifies one; otherwise use an empty string, not "Unassigned". Give a due date only if the transcript explicitly states it for that task; otherwise use an empty string. Preserve the stated deadline without adding inferred days or borrowing deadlines from related tasks. Prerequisites and conditions are not due dates.
Reflect the final position when people revise or withdraw earlier statements. Do not carry forward superseded tasks, conditions or commitments, or attach a condition to a task unless the transcript explicitly does so.
Transcript: {{transcript}}What changed in the prompt
- Define a decision
Say a decision is something explicitly agreed. Models listed ideas and clarifications as decisions.
- Say what to do with no owner or date
Use an empty value, not a guess. Models filled these in with "Unassigned" or a likely date.
- Later statements win
Say that the final position replaces earlier ones, and that a prerequisite isn't a due date.
- Ask for raw JSON
Opus wrapped its JSON in Markdown fences and added a note after it, which breaks a parser.
What it fixed, check by check
The share of answers that passed each check, across all the models with thinking at its lowest setting.
First prompt After one rewrite
Real answers, before and after
A recommendation, recorded as a decision
A hiring panel debrief where the chair recommends a practical assessment and Louis prefers another interview for Candidate B.
“Both candidates will be invited to a practical assessment”
The judge found this listed as a decision when it wasn't one.“The chair recommended inviting both to a practical assessment.”
An owner nobody could hear
A noisy call where someone offers to report a flickering light, and nobody can tell whether it was Imran or Callum.
“Unconfirmed (Imran or Callum)”
It wrote a guess into the owner field, where our rule asks for an empty value.“It was unclear whether that speaker was Imran or Callum”
An idea, not a task
Ana: Should we move launch to March? Ben: Maybe. Ana: Let's think about it. Ben: Someone should check with legal.
“Check with legal about moving the launch to March”
It added a task nobody accepted, inside a code block that a parser can't read.“no one agreed to do it and no owner or due date was set”
How we tested it
- We wrote a plain first prompt, the kind you'd write on day one, and three awkward cases by hand. A model drafted the rest, for 24 cases in all.
- The prompt ran on 6 of those cases. GPT-6.1 Sol rewrote it once from what failed, which is what Spring Prompt's assistant does.
- Both prompts then ran on the 18 cases the rewrite never saw, twice, so scores are out of 36. An answer passes only if it meets every must check.
- We set thinking on purpose. Lowest means off where a model allows it, and low where it doesn't. We also ran the rewrite at medium.
- GPT-6 Luna judged each answer with three votes. Word limits and JSON were checked by code, and we read a sample of answers by hand.
What this can't tell you
- The cases are invented, so a real product will behave differently. Your prompt, data and customers matter more than this table.
- 18 cases is small. Gaps of two or three cases can be chance, so read the top few as level.
- A model drafted most cases and a model judged them.
- Costs are model prices for the answers alone, and they change often. Times come from one day's runs.
- We added GPT-6 Luna on 6 November, on the same cases and prompts. Luna is also the judge, so we checked it with GPT-6.1 Sol judging fresh answers: on medium, Luna got 31 of 36 and Haiku 32.
How we know the judge is right
A model judged each answer with three votes. Word limits and JSON were checked by code, and we read a sample of answers by hand. The three votes are there to catch a judge that changes its mind. Code checks what can be counted. Reading by hand is how we look for a judge that is wrong the same way every time.
A judge can still be wrong, and this experiment has no agreement figure of its own. In your own Prompt test you can open any answer and agree with the judge or give your own verdict, and the test shows how often the judge agreed with you. How often our judges agree with people →
Want a hand setting it up?
Tell us what your prompt does. We'll set the test up with you, inside your own free trial account, so your prompt stays in your account. You don't send it to us.
- Score. Checks and test cases written for your task, and every answer scored.
- Compare. The same test on other models, to see if you get the same result cheaper or faster.
- Improve. A rewrite from what failed, scored again on the same cases.