Does Claude Opus 5.5 score higher at high effort than at low, and does it rest at low?
Question
On harness 0.6.10, does Claude Opus 5.5 score higher on the Fitness Index at its own high effort than at its own low effort? Where does any gap come from: more actions, different actions, or the credits the scoring text names? And does Opus 5.5 rest through most of its turns at low effort?
Why
- On 14 October 2026, in this scenario, Claude Sonnet 5.5's own high effort added 23.2 Index points (95% 18.5 to 27.9), on 8 of 8 maps. The roster held two Sonnet 5.5 seats and two Claude Haiku 5.5 seats.
- At low effort Sonnet 5.5 wrote only its tool call and chose to rest on 67 to 68% of its turns. At high effort it acted on 46 to 53%.
- Whether that is one model's habit, or something Claude's larger models do at low effort, is open.
- Harness 0.6.10 shows every seat that asks the whole scoring text, including which actions earn points. A seat that reads the formula and follows it is a different finding from one that plays the world better, so this design splits the score by its parts.
Setup
- Scenario. open-play-v1: the 8 practice maps and five house bots (random, greedy harvester, honest trader, aggressive raider, envoy). 100 rounds, one action a turn, a 180-second turn clock.
- Roster. Two Claude Opus 5.5 seats and two Claude Haiku 5.5 seats, all on the native Anthropic route, with an output ceiling of 8,192 tokens and native tool calls. The Haiku 5.5 seats play at effort low in both arms, as background.
- Arms. Only the Opus 5.5 seats' effort differs.
- Control "Opus low": adaptive thinking at effort low.
- "Opus high": adaptive thinking at effort high.
- What comes with the setting. Thinking summaries are never shown back, so a seat carries into its next turn only the text it wrote. In the pilots, the Opus 5.5 seats carried a note into 49 of 200 play turns at low (median 84 characters) and into none at high. The Haiku 5.5 seats carried one into 20 of 200 in each arm. H1 reads this as part of the setting's effect.
- Versions. Harness 0.6.10 (prompt v6), engine 0.3.3, rubric 2.3.
- Trials. 8 per arm, each practice map once.
- Seeds. Spawn layouts are drawn from the first public drand round after creation. Both arms play the same layout on each trial.
- Pilots. One match per arm was played on 6 November 2026, before this design was locked, and declared as a pilot.
- The hypotheses and predictions below were written before the pilots and were not changed.
- The pilots set only the per-match spend caps.
- The reads under "Where the score comes from" were asked for in review before any pilot result existed.
- The carried-note count under "Mechanism read" was asked for in review while the pilots played, and was first read on them.
- The background read and the ceiling-cut count were asked for in review after the pilots.
Decision rule, fixed before the first trial
- Metric. The Fitness Index, as the mean of the two Opus 5.5 seats in each trial. The Haiku 5.5 seats are background: their mean is printed beside it and decides nothing.
- One look, after trial 8, at level 0.05 two-sided.
- Opus high is ahead only when both the paired t interval and the bootstrap interval clear zero.
- The exact sign-flip test over the pairs is reported beside it. Its floor is 0.0078 at 8 pairs and 0.031 at 6, the fewest the loss tolerance allows.
- Loss tolerance. 2 trials. Past it, there is no verdict, and no trial is replaced.
- Secondary outcomes, read on the same pairs and stopping nothing:
- output per play turn, the check that the setting took;
- survival;
- messages per 100 rounds.
- Reported beside H1: turns cut at the 8,192-token output ceiling, per arm. Each cut turn is a lost action in its arm.
Hypothesis, prediction and surprise
- H1: Opus high against Opus low.
- Prediction: +15, a hit between +5 and +25.
- An interval that reaches zero reads Undecided: an effect inside it is neither shown nor ruled out.
- Surprise. The estimate of Opus high minus low comes out under +5. We put that at about 1 in 4.
- An estimate that small would not show that effort does not matter unless its interval also ruled out larger gaps.
Where the score comes from (declared now, described without a test)
- The five parts. H1's gap split into the five weighted dimension terms of the Index.
- Counts per seat and arm. Trained levels, messages sent, and completed trades.
- Rescoring. H1 is rescored twice in the platform's own scorer: once with trained levels set to zero, which removes both their adaptation points and the wealth they add, and once without the message term. These are accounting splits, not counterfactuals.
- Exposure. For each seat, whether and from which round its observation carried the whole scoring text, by arm.
- Reading rule. Suppose H1 clears zero and most of its weighted gap sits in the training and message terms. Then the reading is "high effort scored higher mostly through credits the scoring text names", not "it played the world better".
Mechanism read (declared, described)
- What is counted. Each Opus 5.5 seat's share of turns spent on chosen rests, beside world actions, lookups and looks.
- Carried notes. Per arm, the share of Opus 5.5 play turns whose observation carried the seat's own text from its previous turn, and that text's median length.
- Prediction. 50% or more at low, against 67 to 68% for Sonnet 5.5 at low on 14 October.
- What would cut against the resting account. Opus 5.5 at low rests on under 40% of its turns while H1's interval lies above +10.
Background read, declared (described, no test)
- The Haiku 5.5 seats' paired difference, Opus high match minus Opus low match, with its t interval, and with its competitiveness term split out.
- Competitiveness scores a seat's wealth rank in the whole field. So part of any gain for Opus 5.5 can be rank taken from the seats beside it. This read describes whether the background seats lost ground when Opus 5.5 gained.
Raider read, declared
- On 14 October every early dormancy fell on the aggressive raider's first target.
- For each trial, the seat the raider attacks first in either arm is named. When that seat is an Opus 5.5 seat, H1 is also described with that slot left out of both arms of its trial.
Read beside the 14 October Sonnet 5.5 row (descriptive, Welch, never pooled)
The rows differ in all of these:
- The model and its effort scale. Each model's effort levels are its own. The default effort is high for Sonnet 5.5 and medium for Opus 5.5, so "low" need not be the same dose on the two.
- The harness. It was 0.6.7 there and 0.6.10 here. 0.6.10 shows the whole lookup result once, and the scoring text with it. No Sonnet 5.5 seat has played 0.6.10, so a difference between the two models cannot be told apart from the harness change.
- Carried notes. On 14 October Sonnet 5.5 at low carried its own text into 71 of 6,245 turns. Opus 5.5 at low carried it into 49 of 200 in the pilot.
- The background seats. Here they are Haiku 5.5 at low only.
- Spawn layouts.
Sizing
- From the 14 October Sonnet 5.5 seats' paired spread (SD about 8.3 on the two-seat mean), 8 trials detect a gap of about 9.6 points at 80% power.
- Power is 0.99 at +15, 0.83 at +10 and 0.31 at +5.
- If Opus 5.5's spread is 12, the gap is about 13.8 and the power at +15 is 0.86.
- A true gap under 10 would likely read Undecided, and this design says so in advance.
- With eight pairs on eight maps, the pair and map readings coincide.
- Until five pairs are in, the platform's card sizes this design from a generic per-seat prior, with no reference study, and prints a detectable gap of about 22 points. The sizing above, from the Sonnet 5.5 spread, is this design's own.
Refusal history
- With a reasoning-recording tool on its menu, Opus 5.5 was declined on every turn in 4 of 4 runs. With the tool hidden it played 1 of 1 (7 October 2026).
- Harness 0.6.6 took the tool off the menu.
- The pilots on 0.6.10 recorded no refusals.
What it cannot show
- Anything past these two models, one scenario, these five house bots, 100 rounds and one harness version.
- Why a seat chose what it chose: thinking summaries are the vendor's summaries, not the reasoning itself.
- Capability in general: the Fitness Index is not a validated measure of capability, and it can be gamed.
- Other routes: an effort label maps to different mechanisms there.
- Anything about other vendors' models. The rows for each model describe this setup and rank nothing.