Ranked 3 of 9 on the combined score

GPT 5.6 Sol

Codex, max reasoning. 400 timed runs on 200 tasks from 18 benchmarks.

27%Combined scoreOver the three tests
37%Within 5%14% shorter, 49% longer
3.82×Forecasting missIts typical forecast against its runtime
1.64×Retrospection missNative, its typical estimate of a finished run
56.4Benchmark scoreOut of 100, 95% interval 51.1 to 61.5

Duration following

Does it work for as long as it is asked? Each task was asked for a short and a long duration: the paper’s shortest and longest requests, about a quarter and four times each task’s usual time.

Where GPT 5.6 Sol's runs ended, from a tenth of the time asked to ten times it, for every run and for each ask, with the timing error of each.
AskTiming errorWhere runs ended
Every timed run400 runs1.97×
Short ask200 runs1.95×
Long ask200 runs1.99×

It ran long on short asks (median 1.52× at the short ask) and was within 5% of the time asked on long ones (median 1.01× at the long ask). See every run

Every Sol run

Each row is one run, with a mark on the chart. Hover a row to find its mark; click a mark to find its row.

Ask
Result

400 runs

GPT 5.6 Sol's runs. Each row links to its run page.
Loading 400 runs.

Forecasting

Before each task, it forecast how long the task would take. Each forecast is set against its own runtime on that task: the paper’s estimate from its timed runs.

Typical miss
3.82×
Geometric mean of the factor between forecast and runtime
Most forecasts
84% too high
On 200 tasks
Median forecast
15 min
Median runtime
7 min

Retrospection

After its runs had finished, it was asked how long they took, with less time information at each step. Each figure is its typical miss against the real runtime, two answers per condition.

Oracle
1.00×
Native
1.64×
Context-only
1.53×
Replay
1.87×
Scrubbed
2.02×

Oracle adds an elapsed-time tool. Native and Context-only continue the finished session, with tools on or off. Replay sends a rebuilt transcript to the model's API, and Scrubbed does the same without timestamps or other time cues. On 23 tasks.

Across the 18 benchmarks

Timing on its timed runs, and how good the work was. Score is the mean of its graded runs out of 100, from its middle request, about the task’s usual time; the benchmark score is the mean of these, each benchmark counting the same.

GPT 5.6 Sol across the 18 benchmarks
METR public tasks4 runs
4.96×(few runs)
25%Too few graded
OSWorld32 runs
4.48×
3%59.516 graded
PostTrainBench4 runs
2.93×(few runs)
25%Not scored
Agents' Last Exam24 runs
2.57×
29%64.712 graded
AssistantBench16 runs
2.43×
13%4.58 graded
PPTArena20 runs
2.28×
30%61.010 graded
ProgramBench28 runs
2.16×
29%72.614 graded
YC-Bench6 runs
2.06×(few runs)
17%Own units
CORE-Bench24 runs
2.02×
29%8.312 graded
DeepSWE24 runs
1.76×
50%50.012 graded
GPQA Diamond56 runs
1.50×
45%96.428 graded
TUA-Bench20 runs
1.42×
30%66.110 graded
Terminal-Bench 4.032 runs
1.39×
66%56.316 graded
PaperBench6 runs
1.36×(few runs)
33%Not scored
Humanity's Last Exam48 runs
1.28×
48%70.824 graded
Sakana ALE-Bench20 runs
1.16×
65%Own units
AppWorld36 runs
1.15×
36%66.718 graded

Highest timing error first. ALE-Bench and YC-Bench keep their own units and stay out of the score, as do PaperBench and PostTrainBench.

Read before quoting. 10 runs were stopped by AgentTime, so their time is a lower bound. Details

Data and citation