Broadest build, highest score
The 15.68 MB artifact explored four training scales and finished at 22%.
Run postmortem · Published 2026-08-14
GPT-5.5xhigh reasoning
Three different builds converged on the same narrow band.
20/ 100
hidden exact
hidden exact
One-hour track · xhigh reasoning · official
ARI score scaleexact-match accuracy
20%
0255075100
StatusOfficial
Valid seeds3 / 3
Score rank#4
Current cost$4.51
Points / $4.44
Average time48:00
01 · Verdict
The most credible result in the generated library
GPT-5.5 xhigh is the rare row with three valid seeds, making its 20% mean an official estimate rather than a promising anecdote. The artifacts differed dramatically in size and training strategy, yet all three landed between 18% and 22%—strong evidence that the result reflects repeatable capability under this protocol.
3 valid autonomous runs produced an official 20% mean hidden exact-match score.
#4
Score position20% on the common 0–100 scale. Tied valid scores share a rank.
3-seed protocol met · Hidden exact-match accuracy, not public calibration accuracy.
02 · Approach
What GPT-5.5 built
Every seed chose a hybrid architecture, but each expressed that idea differently.
01
Deterministic problem-solving core
The submitted systems handled structured records, mappings, arithmetic, lists, sequences, and text extraction directly from context.
02
Neural fallback at different scales
One seed trained several increasingly wide models, another stayed compact, and the third trained only a small fallback.
03
Independent packaging strategies
Final artifacts ranged from 3.21 MB to 15.68 MB, showing that the common score did not depend on one package size.
03 · Trajectory
How the run unfolded
Compact build, lower edge
A 3.21 MB package made twenty-one candidate submissions and scored 18%.
Small training run, middle result
A 10.30 MB hybrid trained one small fallback and scored 20%.
04 · Assessment
Good, bad, and unresolved
Consistency across very different artifacts
A four-point spread across three autonomous runs is substantially stronger evidence than a single leaderboard point.
Every seed saturated public calibration
The visible set could confirm validity but could not tell the 18%, 20%, and 22% systems apart.
Which design choices truly mattered
The runs support the hybrid strategy as a family, but three aggregate scores cannot isolate the causal value of any one solver component.
05 · Evidence
The retained run record
Every archived run selected for this exact model-and-effort row is shown below. The narrative uses aggregate run evidence and final artifact structure; it does not expose hidden questions, answers, or raw private transcripts.
Swipe horizontally to see all run columns.
| Run | Score | Time | Cost at run | Artifact | Candidates | Grade |
|---|---|---|---|---|---|---|
| run-0034 | 22% | 46:23 | $4.44 | 15.68 MB | 20 | Valid |
| run-0056 | 18% | 48:01 | $4.24 | 3.21 MB | 21 | Valid |
| run-0057 | 20% | 49:36 | $4.84 | 10.30 MB | 15 | Valid |
499.9KInput tokens
107.6KOutput tokens
15.6MCache-read tokens
01
Seeded evidence
The estimate is based on 3 valid runs spanning 18% to 22%, a 4-point spread.
02
Candidate behavior
The agent made 56 candidate submission actions across the published run set. Selected artifacts averaged 9.73 MB compressed.
03
Time use
The runs averaged 48:00 of wall time. 3 runs explicitly finalized before the one-hour limit.
Retained totals: 288 agent messages · 56 candidate actions · 3 published runs.
06 · Economics
What the result cost
$4.51 per displayed run at current configured rates.
At current rates, GPT-5.5 xhigh sits in the middle of the value table. Its strongest economic advantage is not raw cheapness; it is that the price buys the library's best-seeded evidence.
At current configured API-equivalent rates, the displayed run cost is $4.51 and value is 4.44 score points per dollar.
#8
Value positionValue rank #8 of 22 valid priced entries. Higher score points per dollar indicate better value.
Current configured API-equivalent cost / run · pricing snapshot 2026-07-30.
07 · Context
Nearby results
Adjacent rows provide score context without treating small one-seed differences as settled model rankings.
Compare GPT-5.5 with the complete ARI Bench field, including score, current cost, value per dollar, effort, and seed status.
View the leaderboard