Secured a checkpoint
The first trained candidate was promoted as insurance before the more ambitious inference work began.
Run postmortem · Published 2026-08-14
GPT-5.5low reasoning
It found the right idea, then stopped before it was broad enough.
4/ 100
hidden exact
hidden exact
One-hour track · low reasoning · provisional
ARI score scaleexact-match accuracy
4%
0255075100
StatusProvisional
Valid seeds1 / 3
Score rank#17
Current cost$0.91
Points / $4.39
Average time19:06
01 · Verdict
A useful sketch, not a finished system
GPT-5.5 low quickly recognized that a neural checkpoint alone would not be enough and moved toward a hybrid predictor. The run produced a valid artifact and perfect public calibration, but its 4% hidden score shows that the early solver remained narrow.
One valid autonomous run produced a provisional 4% hidden exact-match score.
#17
Score position4% on the common 0–100 scale. Tied valid scores share a rank.
1 of 3 required seeds · Hidden exact-match accuracy, not public calibration accuracy.
02 · Approach
What GPT-5.5 built
The run combined a trained byte model with task-aware inference logic instead of relying on either one alone.
01
Timeout-safe neural checkpoint
It trained and promoted a valid model almost immediately, protecting the run from ending without a gradeable artifact.
02
Hybrid inference layer
It added contextual copying, extraction, arithmetic, sequence, and record-oriented rules around the neural fallback.
03
Model scaling under a small budget
Six training attempts explored larger configurations, but the final artifact still had to fit the fixed 16 MB cap.
03 · Trajectory
How the run unfolded
Broadened the solver
The agent layered deterministic completions over the trained model and repeatedly resubmitted candidates.
Finalized very early
The run ended with roughly forty minutes unused. Public calibration was saturated, but hidden coverage remained thin.
04 · Assessment
Good, bad, and unresolved
Fast recognition of the real task
The run did not stay trapped in ordinary next-byte training; it pivoted toward contextual problem solving.
Public success hid a narrow solver
A 25/25 public result became 4/100 on the hidden exam, the clearest sign that the visible task family was overfit.
What forty more minutes could buy
More time might have widened the solver, but the public set no longer offered a useful ranking signal.
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-0030 | 4% | 19:06 | $0.91 | 14.03 MB | 11 | Valid |
54.4KInput tokens
10.5KOutput tokens
649.2KCache-read tokens
01
One run, one outcome
Autonomous runs are stochastic. Two more valid seeds are required before this can be treated as an official estimate.
02
Candidate behavior
The agent made 11 candidate submission actions across the published run set. Selected artifacts averaged 14.03 MB compressed.
03
Time use
The runs averaged 19:06 of wall time. 1 run explicitly finalized before the one-hour limit.
Retained totals: 34 agent messages · 11 candidate actions · 1 published run.
06 · Economics
What the result cost
$0.91 per displayed run at current configured rates.
The low-effort run was inexpensive and fast, but the saved spend came with a large accuracy tradeoff. Its value ratio looks ordinary only because both cost and score were small.
At current configured API-equivalent rates, the displayed run cost is $0.91 and value is 4.39 score points per dollar.
#10
Value positionValue rank #10 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