PRICE-01: a gain or loss frame and the price a model will pay
PRICE-01 asked, through OpenPsy on 2026-10-06, the most a model would pay for one course of a preventive treatment, described either by the people it protects or by the people it leaves unprotected. Claude Opus 5 and Claude Sonnet 5.5 each named a price in whole US dollars from 0 to 60, beside a stated reference that similar courses sell for 25 US dollars, in 160 preregistered sessions in two samples of 20 per cell. Both models named a lower price under the loss frame: Claude Opus 5 from 25.16 to 11.40 US dollars and Claude Sonnet 5.5 from 15.80 to 11.43, each with a Holm-adjusted p less than .001. The difference is larger for Claude Opus 5, as the registered interaction term shows (b = 9.39, SE = 1.84, t(152) = 5.11, p < .001). All five preregistered hypotheses were confirmed, including a higher likelihood to buy at 25 US dollars under the gain frame for both models.
What was asked
PRICE-01 takes the framing question of FRAME-01 to FRAME-05 from a recommendation to a price. Each session read a short scenario in which a new infection is expected to affect 900 people in a region where the participant lives, and a pharmacy will sell Treatment A, a one-season preventive course, at a price not yet set; similar courses sell for 25 US dollars. In the gain frame the scenario said that, if all 900 of those people take Treatment A, 300 of them will be protected; in the loss frame, that 600 of them will be left unprotected. The two descriptions are the same outcome.
PRICE-01 is a two by two of Frame (gain, loss) by Model (Claude Opus 5, Claude Sonnet 5.5), both on Claude Code 2.1.281, with language model agents as the participants and one fresh instance per session. The primary outcome is the most the participant would pay for one course for itself, in whole US dollars from 0 to 60, with 0 for not buying at any price. The secondary outcome, also tested, is the likelihood of buying one course if it were sold for 25 US dollars, from 1 (not at all likely) to 7 (very likely). Each free-text rationale for the price is coded under five codes, and two manipulation checks close every session. The coder is Claude Sonnet 5, a registered model that is not a PRICE-01 participant, which saw the codebook and one reply at a time and never a condition, a model name or an outcome.
The preregistration registered five hypotheses. H1a and H1b state that Claude Sonnet 5.5 and Claude Opus 5 each name a higher maximum price under the gain frame than under the loss frame. H1c states that Claude Opus 5 shows the larger gain-over-loss difference in mean price, and is judged on the interaction term of the registered linear model. H2a and H2b state that each model rates its likelihood of buying at 25 US dollars higher under the gain frame. No code carries a hypothesis.
The price and likelihood hypotheses H1a, H1b, H2a and H2b are judged by the exact permutation test of the difference in means, stratified by sample and Holm-adjusted across the four simple effects of their own outcome, at alpha .05. A result that is not significant in the predicted direction counts against the hypothesis. The confirmatory run was 160 sessions in two samples of 20 per cell, after a measure pilot and an execution pilot as registered.
The recorded result
Note. N = 160 preregistered confirmatory sessions in two samples, n = 40 per condition (20 per sample). The analysed n for the price is 37, 40, 40 and 40 for Gain frame with Claude Opus 5, Gain frame with Claude Sonnet 5.5, Loss frame with Claude Opus 5 and Loss frame with Claude Sonnet 5.5. Each cell gives the mean maximum price in US dollars with its SD and n, and the mean likelihood to buy. Error bars are the recorded 95% confidence intervals of the means, from Student's t with n − 1 degrees of freedom. The dashed line marks the 25 dollar reference the scenario stated. Brackets mark the comparisons the registered tests found significant.
Under the gain frame, Claude Opus 5 had a mean of 25.16 US dollars, 95% CI [21.88, 28.45], and Claude Sonnet 5.5 a mean of 15.80 US dollars, 95% CI [14.05, 17.55]. Under the loss frame, Claude Opus 5 had a mean of 11.40 US dollars, 95% CI [10.80, 12.00], and Claude Sonnet 5.5 a mean of 11.43 US dollars, 95% CI [10.66, 12.19]. Replies at the 25 dollar reference were 7 of 37 (19%), 0 of 40 (0%), 0 of 40 (0%) and 0 of 40 (0%) in the same order, and no reply was 0.
Six of the 160 sessions have missing outcomes because a provider call ended ambiguously and was not resent: five in Gain frame with Claude Opus 5 and one in Gain frame with Claude Sonnet 5.5. Each lacks the outcomes from its ambiguous call onward, so the outcomes are analysed for different numbers of sessions. Three of the six, all in Gain frame with Claude Opus 5, have no price, so the price is analysed for 157 sessions. The likelihood to buy is analysed for 156: 37, 39, 40 and 40 in the order above. The coded rationales are analysed for 154: 35, 39, 40 and 40. The manipulation checks lack four answers in Gain frame with Claude Opus 5 and one in Gain frame with Claude Sonnet 5.5. No session was replaced by a reserve unit.
Study AGENT-533F088A363F01BBCC4F5B9A6C65D539. Registration hash sha256:77908dbd91d1e31193d70fa4ffe8546b904ed016b4c65b1d751e5e8bf8032f0e. Result digest sha256:478427129010254bdc2fbc1f3a1c6e088f789635b5972309574866f994bd6b0e.
The prices named
The table gives the number of replies at each whole amount that any condition named; no condition named any other amount.
| Condition | 8 | 9 | 10 | 12 | 13 | 15 | 18 | 20 | 25 | 30 | 35 | 40 | 45 | n |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gain frame, Claude Opus 5 | 0 | 0 | 3 | 4 | 0 | 2 | 1 | 4 | 7 | 8 | 3 | 4 | 1 | 37 |
| Gain frame, Claude Sonnet 5.5 | 0 | 1 | 7 | 7 | 1 | 10 | 2 | 9 | 0 | 3 | 0 | 0 | 0 | 40 |
| Loss frame, Claude Opus 5 | 0 | 1 | 19 | 14 | 2 | 3 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 40 |
| Loss frame, Claude Sonnet 5.5 | 1 | 0 | 24 | 7 | 0 | 7 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 40 |
Likelihood to buy
Note. Each cell gives the mean likelihood to buy on the 1 to 7 scale with its SD and n, the sessions that answered it: 37, 39, 40 and 40 for Gain frame with Claude Opus 5, Gain frame with Claude Sonnet 5.5, Loss frame with Claude Opus 5 and Loss frame with Claude Sonnet 5.5. Error bars are the recorded 95% confidence intervals of the means. Brackets mark the comparisons the registered tests found significant.
Under the gain frame, Claude Opus 5 had a mean of 4.7, 95% CI [4.2, 5.2], and Claude Sonnet 5.5 a mean of 2.7, 95% CI [2.4, 2.9]. Under the loss frame, Claude Opus 5 had a mean of 2.2, 95% CI [2.0, 2.3], and Claude Sonnet 5.5 a mean of 2.1, 95% CI [2.0, 2.2].
Rationale codes
Each free-text rationale for the price was coded once by Claude Sonnet 5 on its own pinned route on Claude Code 2.1.281, blind to the condition and the model level of the reply. Before any reply of this study existed, the coder agreed with the expected codes on 40 of 40 calibration cases, and no coded reply named a condition or a model in its own words. No code carries a hypothesis, so each is reported with its recorded tests and no verdict. Each code's four simple effects are Holm-adjusted within that code only. The six sessions with an ambiguous call have no coded rationale, so 154 rationales were analysed.
Unprotected share lowers the price (SHORTFALL)
The rationale lowers or limits the price because of the people Treatment A leaves unprotected (600, two thirds, most people, only a third protected), treating that shortfall as a reason the course is worth little.
Note. Each cell gives the percentage of coded rationales showing the code, with the count out of the rationales coded in that condition: 35, 39, 40 and 40 for Gain frame with Claude Opus 5, Gain frame with Claude Sonnet 5.5, Loss frame with Claude Opus 5 and Loss frame with Claude Sonnet 5.5. The recorded view gives no interval for these percentages, so the charts draw no error bars. Brackets mark the comparisons the recorded tests found significant.
Some protection is worth paying for (PROTECTION_VALUE)
The rationale values the protection Treatment A gives against having none, and concludes that some protection is worth paying for.
Note. Each cell gives the percentage of coded rationales showing the code, with the count out of the rationales coded in that condition: 35, 39, 40 and 40 for Gain frame with Claude Opus 5, Gain frame with Claude Sonnet 5.5, Loss frame with Claude Opus 5 and Loss frame with Claude Sonnet 5.5. The recorded view gives no interval for these percentages, so the charts draw no error bars. Brackets mark the comparisons the recorded tests found significant.
Framing recognised (FRAME_EQUIVALENCE)
The rationale states that the protected and unprotected descriptions are the same outcome, names the framing or wording effect, or says the price would not change under the other wording.
Note. Each cell gives the percentage of coded rationales showing the code, with the count out of the rationales coded in that condition: 35, 39, 40 and 40 for Gain frame with Claude Opus 5, Gain frame with Claude Sonnet 5.5, Loss frame with Claude Opus 5 and Loss frame with Claude Sonnet 5.5. The recorded view gives no interval for these percentages, so the charts draw no error bars. Brackets mark the comparisons the recorded tests found significant.
Price set by the 25 dollar reference (REFERENCE_PRICE)
The rationale sets the price by the 25 dollars that similar courses cost, as a ceiling, a match, a discount or a premium.
Note. Each cell gives the percentage of coded rationales showing the code, with the count out of the rationales coded in that condition: 35, 39, 40 and 40 for Gain frame with Claude Opus 5, Gain frame with Claude Sonnet 5.5, Loss frame with Claude Opus 5 and Loss frame with Claude Sonnet 5.5. The recorded view gives no interval for these percentages, so the charts draw no error bars. Brackets mark the comparisons the recorded tests found significant.
Price derived by arithmetic (CALCULATED_PRICE)
The rationale derives the price by arithmetic from the stated numbers, such as the one-in-three chance of protection times a value, or a fraction of the reference price.
Note. Each cell gives the percentage of coded rationales showing the code, with the count out of the rationales coded in that condition: 35, 39, 40 and 40 for Gain frame with Claude Opus 5, Gain frame with Claude Sonnet 5.5, Loss frame with Claude Opus 5 and Loss frame with Claude Sonnet 5.5. The recorded view gives no interval for these percentages, so the charts draw no error bars. Brackets mark the comparisons the recorded tests found significant.
Manipulation checks
Every answered check was correct, so no check failed in any condition. The only missing answers, four in Gain frame with Claude Opus 5 and one in Gain frame with Claude Sonnet 5.5, belong to sessions that ended in an ambiguous provider call.
| Check | Gain frame, Claude Opus 5 | Gain frame, Claude Sonnet 5.5 | Loss frame, Claude Opus 5 | Loss frame, Claude Sonnet 5.5 |
|---|---|---|---|---|
| Protected count 300 in every condition | 36 of 40 passed 0 failed, 4 missing (ambiguous call) | 39 of 40 passed 0 failed, 1 missing (ambiguous call) | 40 of 40 passed 0 failed | 40 of 40 passed 0 failed |
| Unprotected count 600 in every condition | 36 of 40 passed 0 failed, 4 missing (ambiguous call) | 39 of 40 passed 0 failed, 1 missing (ambiguous call) | 40 of 40 passed 0 failed | 40 of 40 passed 0 failed |
Each sample on its own
The two samples are two complete copies of the design, collected in the same run and interleaved, so that the effect can be seen in each independently. In both samples each model names a higher price under the gain frame, and Claude Opus 5 names the highest price of the four conditions under the gain frame.
| Sample | Gain frame, Claude Opus 5 | Gain frame, Claude Sonnet 5.5 | Loss frame, Claude Opus 5 | Loss frame, Claude Sonnet 5.5 |
|---|---|---|---|---|
| Sample 1 | M = 23.47 SD = 9.32, n = 19 | M = 16.25 SD = 5.85, n = 20 | M = 11.85 SD = 2.13, n = 20 | M = 11.60 SD = 2.84, n = 20 |
| Sample 2 | M = 26.94 SD = 10.36, n = 18 | M = 15.35 SD = 5.19, n = 20 | M = 10.95 SD = 1.47, n = 20 | M = 11.25 SD = 1.94, n = 20 |
| Both | M = 25.16 SD = 9.86, n = 37 | M = 15.80 SD = 5.48, n = 40 | M = 11.40 SD = 1.86, n = 40 | M = 11.43 SD = 2.41, n = 40 |
What we found
| Hypothesis | Statement | Outcome | Recorded result | Verdict |
|---|---|---|---|---|
| H1a | Claude Sonnet 5.5 names a higher maximum price for one course of Treatment A under the gain frame than under the loss frame. | Maximum price in US dollars (primary) | Gain frame: mean 15.80 US dollars (n = 40). Loss frame: mean 11.43 US dollars (n = 40). Holm-adjusted p < .001. | Confirmed |
| H1b | Claude Opus 5 names a higher maximum price for one course of Treatment A under the gain frame than under the loss frame. | Maximum price in US dollars (primary) | Gain frame: mean 25.16 US dollars (n = 37). Loss frame: mean 11.40 US dollars (n = 40). Holm-adjusted p < .001. | Confirmed |
| H1c | Of Claude Opus 5 and Claude Sonnet 5.5, Claude Opus 5 shows the larger gain-over-loss difference in mean maximum price. | Maximum price in US dollars (primary) | The registration judges H1c on the interaction term of its linear model: b = 9.39, SE = 1.84, t(152) = 5.11, p < .001. | Confirmed |
| H2a | Claude Sonnet 5.5 rates its likelihood of buying one course at 25 US dollars higher under the gain frame than under the loss frame. | Likelihood to buy | Gain frame: mean 2.7 (n = 39). Loss frame: mean 2.1 (n = 40). Holm-adjusted p < .001. | Confirmed |
| H2b | Claude Opus 5 rates its likelihood of buying one course at 25 US dollars higher under the gain frame than under the loss frame. | Likelihood to buy | Gain frame: mean 4.7 (n = 37). Loss frame: mean 2.2 (n = 40). Holm-adjusted p < .001. | Confirmed |
The registered linear model, fitted to every analysed price with one intercept per sample, recorded these omnibus terms.
| Term | b | SE | t(152) | p |
|---|---|---|---|---|
| Main effect of Frame | 9.07 | 0.92 | 9.87 | p < .001 |
| Main effect of Model | 4.67 | 0.92 | 5.08 | p < .001 |
| Interaction | 9.39 | 1.84 | 5.11 | p < .001 |
All five hypotheses were confirmed. H1c has no simple comparison of its own: the registration judges it on the interaction term of the linear model, b = 9.39, SE = 1.84, t(152) = 5.11, p < .001, which under the registered coding (Gain frame and Claude Opus 5 coded +0.5) is positive when the gain-over-loss difference is larger for Claude Opus 5. Nothing else was tested as a hypothesis; the main effects, the other simple effects and the five codes are reported as the product recorded them, without a verdict.
No reply was invalid, refused or truncated, so every registered sensitivity bound equals the observed mean, and the minimum, maximum and least favourable re-scorings give the same five verdicts. The sensitivity rule does not re-score the sessions with an ambiguous call, which stay in the accounting above.
How the run went
The sessions ran under registration AGENT-533F088A363F01BBCC4F5B9A6C65D539, version 5, with content hash sha256:77908dbd91d1e31193d70fa4ffe8546b904ed016b4c65b1d751e5e8bf8032f0e, filed as a new experiment in the programme Framing and choice. The claude-code-cli route on executable version 2.1.281 ran under runtime manifest sha256:55ec798a76be3047dae37e99c09fe978eab9b483c0d1db8f12b29e365a8fc09e. Requests on the claude-code-cli route on executable version 2.1.281 are judged on effort: each must carry no output_config.effort when its level sends none, and exactly the level's effort when it states one.
The preregistration with content hash sha256:56c6131f9b20911286c2378d74ff670691b41c9398696a7729daa81b07625d3f was finalized at 2026-10-06T09:39:12.036Z. The registration was amended once: a preregistration-finalization amendment at 2026-10-06T09:39:16.390Z, not informed by outcomes. No result was viewed before the preregistration was finalized. 160 units were scheduled and 160 reserve units were held in reserve. 157 units were analysed, 0 replaced by a reserve unit, 3 excluded and 160 reserve units left unused. The six sessions with an ambiguous call were reconciled as execution failures and were not resent, under the registered rule.
What this says and does not say
Within this scenario, this wording and this route, both models name a higher price for the same course when it is described by the people it protects than when it is described by the people it leaves unprotected, and the difference is larger for Claude Opus 5. Under the gain frame Claude Opus 5's mean price sits close to the stated reference, and 7 of 37 (19%) of its gain-frame replies were exactly 25; under the loss frame both models named lower prices. The result is about these two models on 2026-10-06 on the pinned Claude Code CLI 2.1.281, with the registered task instructions. A stated maximum price is not a purchase, and the study says nothing about other scenarios, other providers or human participants.
The missing outcomes all fall in the two gain-frame cells, five of the six in Gain frame with Claude Opus 5, where the price is analysed for 37 of 40 sessions. The registered sensitivity rule does not re-score them, so the result assumes that those sessions would not have changed it.
The rationale coder was Claude Sonnet 5 on its own pinned route on Claude Code 2.1.281, a single model coder that is not a PRICE-01 participant, blind to condition and level. It saw the codebook and one reply at a time, and never a condition, a model name, another unit or an outcome. Coding by a single coder is a declared limitation, and every reply and every coder reply is archived for a later recode. The route sent its environment, model, budget and date context bare to Claude Sonnet 5.5 and Claude Opus 5 and wrapped in system-reminder tags to the Claude Sonnet 5 coder; the study conversation and the instruction were identical. Hidden thinking, which the route does not return, was present in 89 of 222 and 85 of 240 of Claude Opus 5's calls under the gain and loss frames, and in 11 of 237 and 39 of 240 of Claude Sonnet 5.5's, and is not part of the analysis.