Hugging Face's logo

Datasets:
dimhold
/
cutoff-probe

Tasks:
Question Answering
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
1K - 10K
Tags:
llm-evaluation
knowledge-cutoff
claude
npm
pypi
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
cutoff-probe / stats.md
dimhold's picture
dimhold
Mirror of github.com/dimhold/cutoff-probe data (history squashed after redaction)
c3d105b
preview code
|
Raw History Blame Contribute Delete
20 kB
# cutoff-probe (A9): tables
Built on 2026-10-20 from `out/calls.jsonl` (1243 rows, 1075 successful, 168
failed rows excluded). Part 1 is the output of `analyze.py` as is (the CRITERIA
measures plus the arm with a required estimate). Part 2 is the output of
`intervals.py`: confidence intervals, bootstrap and monthly slices; it calls
neither the network nor any model. Shares in part 1 are computed over answers
with no missing items. Column headings were translated from Russian when the
repository was prepared; every number is regenerated by the 2 scripts and is
identical to the original tables.
Regenerate:
python analyze.py > part1.md
python intervals.py > part2.md
# Part 1. analyze.py
## Cutoff: vendor / measured / self report / frontier (median)
| model | vendor | measured | 2024 baseline | false recognition | self report (CLI) | self report with caveat | frontier median | frontier p90 | nonexistent in frontier | batches | unparsed |
|---|---|---|---|---|---|---|---|---|---|---|---|
| claude-opus-5-5 | 2026-06 | 2025-09 | 0.941 | 0/400 | 2026-06 x18 | 2025-01 x5 | 2025-08 | 2025-10 | 2/200 | 76 | 0 |
| claude-sonnet-5-5 | 2026-06 | 2025-12 | 0.955 | 0/400 | 2026-06 x18, 2026-05 x1 | 2025-05 x2, 2025-09 x1, 2025-08 x1 | 2025-10 | 2025-12 | 3/200 | 76 | 0 |
| claude-haiku-5-5 | 2026-06 | 2025-09 | 0.812 | 3/400 | 2026-06 x12, 2026-05 x1 | 2025-03 x2, 2025-01 x2, 2025-05 x1 | 2025-09 | 2025-10 | 5/200 | 76 | 0 |
| claude-fable-5-1 | 2026-06 | 2025-11 | 1.0 | 1/400 | 2026-06 x15, 2026-05 x2, 2026-04 x1 | 2025-01 x3, 2025-03 x1, 2025-05 x1 | 2025-09 | 2025-11 | 4/200 | 76 | 0 |
| claude-opus-5 | 2026-05 | 2025-10 | 0.972 | 4/400 | 2026-05 x21, 2026-03 x1 | 2025-01 x5 | 2025-10 | 2026-02 | 13/200 | 76 | 0 |
| claude-sonnet-5 | 2026-01 | 2025-01 | 0.719 | 2/394 | 2026-01 x18, 2025-12 x3, 2025-02 x1 | 2025-02 x3, 2025-04 x2 | 2025-02 | 2025-09 | 4/200 | 76 | 1 |
| claude-haiku-4-5 | 2025-02 | 2024-12 | 0.043 | 20/392 | 2025-02 x16 | 2024-04 x3, 2025-02 x2 | 2024-02 | 2024-07 | 7/154 | 76 | 2 |
## Share correct by quarter
| quarter | opus-5-5 | sonnet-5-5 | haiku-5-5 | fable-5-1 | opus-5 | sonnet-5 | haiku-4-5 |
|---|---|---|---|---|---|---|---|
| 2023Q1 | 81% | 83% | 65% | 90% | 89% | 66% | 6% |
| 2023Q2 | 85% | 89% | 65% | 97% | 97% | 79% | 6% |
| 2023Q3 | 94% | 96% | 78% | 97% | 97% | 89% | 14% |
| 2023Q4 | 88% | 90% | 79% | 99% | 94% | 79% | 16% |
| 2024Q1 | 92% | 99% | 82% | 100% | 97% | 81% | 9% |
| 2024Q2 | 93% | 94% | 75% | 100% | 96% | 61% | 3% |
| 2024Q3 | 97% | 96% | 79% | 100% | 97% | 76% | 1% |
| 2024Q4 | 94% | 93% | 89% | 100% | 99% | 69% | 4% |
| 2025Q1 | 90% | 96% | 79% | 99% | 100% | 28% | 0% |
| 2025Q2 | 89% | 94% | 76% | 97% | 96% | 15% | 0% |
| 2025Q3 | 61% | 92% | 60% | 88% | 86% | 3% | 0% |
| 2025Q4 | 28% | 86% | 22% | 57% | 26% | 3% | 0% |
| 2026Q1 | 7% | 22% | 3% | 7% | 6% | 0% | 0% |
| 2026Q2 | 0% | 0% | 0% | 0% | 0% | 0% | 0% |
| 2026Q3 | 0% | 0% | 0% | 0% | 0% | 0% | 0% |
| 2026Q4 | 0% | 0% | 0% | 0% | 0% | 4% | 0% |
## Abstention arm: share "known" (correct + wrong) by quarter
| quarter | opus-5-5 | sonnet-5-5 | haiku-5-5 | fable-5-1 | opus-5 | sonnet-5 | haiku-4-5 |
|---|---|---|---|---|---|---|---|
| 2023Q1 | 81% | 83% | 86% | 94% | 92% | 66% | 9% |
| 2023Q2 | 85% | 89% | 97% | 97% | 97% | 93% | 17% |
| 2023Q3 | 94% | 96% | 97% | 97% | 97% | 90% | 24% |
| 2023Q4 | 88% | 90% | 93% | 100% | 96% | 90% | 23% |
| 2024Q1 | 92% | 99% | 97% | 100% | 97% | 84% | 13% |
| 2024Q2 | 93% | 94% | 100% | 100% | 96% | 69% | 6% |
| 2024Q3 | 97% | 96% | 94% | 100% | 97% | 80% | 7% |
| 2024Q4 | 94% | 93% | 99% | 100% | 99% | 70% | 6% |
| 2025Q1 | 90% | 97% | 93% | 99% | 100% | 30% | 0% |
| 2025Q2 | 89% | 94% | 92% | 97% | 96% | 15% | 0% |
| 2025Q3 | 61% | 92% | 76% | 88% | 86% | 3% | 0% |
| 2025Q4 | 28% | 86% | 36% | 57% | 28% | 3% | 1% |
| 2026Q1 | 7% | 22% | 4% | 7% | 6% | 0% | 0% |
| 2026Q2 | 0% | 0% | 0% | 1% | 1% | 0% | 0% |
| 2026Q3 | 0% | 0% | 4% | 0% | 1% | 0% | 0% |
| 2026Q4 | 0% | 0% | 0% | 0% | 0% | 4% | 5% |
## Abstention arm: share confidently wrong by quarter
| quarter | opus-5-5 | sonnet-5-5 | haiku-5-5 | fable-5-1 | opus-5 | sonnet-5 | haiku-4-5 |
|---|---|---|---|---|---|---|---|
| 2023Q1 | 0% | 0% | 21% | 4% | 3% | 0% | 3% |
| 2023Q2 | 0% | 0% | 32% | 0% | 0% | 13% | 12% |
| 2023Q3 | 0% | 0% | 19% | 0% | 0% | 1% | 10% |
| 2023Q4 | 0% | 0% | 14% | 1% | 1% | 11% | 7% |
| 2024Q1 | 0% | 0% | 15% | 0% | 0% | 3% | 4% |
| 2024Q2 | 0% | 0% | 25% | 0% | 0% | 8% | 3% |
| 2024Q3 | 0% | 0% | 15% | 0% | 0% | 4% | 6% |
| 2024Q4 | 0% | 0% | 10% | 0% | 0% | 1% | 1% |
| 2025Q1 | 0% | 1% | 14% | 0% | 0% | 1% | 0% |
| 2025Q2 | 0% | 0% | 15% | 0% | 0% | 0% | 0% |
| 2025Q3 | 0% | 0% | 17% | 0% | 0% | 0% | 0% |
| 2025Q4 | 0% | 0% | 14% | 0% | 1% | 0% | 1% |
| 2026Q1 | 0% | 0% | 1% | 0% | 0% | 0% | 0% |
| 2026Q2 | 0% | 0% | 0% | 1% | 1% | 0% | 0% |
| 2026Q3 | 0% | 0% | 4% | 0% | 1% | 0% | 0% |
| 2026Q4 | 0% | 0% | 0% | 0% | 0% | 0% | 5% |
## Estimate arm: month within 1 by quarter
| quarter | opus-5-5 | sonnet-5-5 | haiku-5-5 | fable-5-1 | opus-5 | sonnet-5 | haiku-4-5 |
|---|---|---|---|---|---|---|---|
| 2023Q1 | 94% | 92% | 67% | 97% | 92% | 86% | 25% |
| 2023Q2 | 97% | 86% | 69% | 97% | 97% | 81% | 11% |
| 2023Q3 | 100% | 97% | 92% | 100% | 100% | 100% | 47% |
| 2023Q4 | 100% | 97% | 89% | 100% | 97% | 89% | 36% |
| 2024Q1 | 100% | 97% | 83% | 100% | 100% | 92% | 31% |
| 2024Q2 | 100% | 100% | 81% | 100% | 100% | 86% | 36% |
| 2024Q3 | 100% | 100% | 86% | 100% | 97% | 94% | 36% |
| 2024Q4 | 100% | 100% | 94% | 100% | 100% | 97% | 36% |
| 2025Q1 | 100% | 94% | 94% | 100% | 97% | 92% | 28% |
| 2025Q2 | 100% | 100% | 89% | 100% | 100% | 78% | 0% |
| 2025Q3 | 100% | 100% | 94% | 100% | 97% | 75% | 3% |
| 2025Q4 | 100% | 97% | 94% | 94% | 86% | 56% | 3% |
| 2026Q1 | 97% | 94% | 89% | 86% | 69% | 19% | 3% |
| 2026Q2 | 64% | 83% | 53% | 58% | 42% | 19% | 3% |
| 2026Q3 | 58% | 53% | 44% | 44% | 50% | 17% | 3% |
| 2026Q4 | 17% | 25% | 42% | 42% | 25% | 17% | 8% |
## Estimate arm: share "remember" by quarter
| quarter | opus-5-5 | sonnet-5-5 | haiku-5-5 | fable-5-1 | opus-5 | sonnet-5 | haiku-4-5 |
|---|---|---|---|---|---|---|---|
| 2023Q1 | 78% | 86% | 39% | 81% | 53% | 14% | 0% |
| 2023Q2 | 92% | 100% | 53% | 92% | 75% | 19% | 3% |
| 2023Q3 | 78% | 97% | 50% | 89% | 81% | 33% | 8% |
| 2023Q4 | 83% | 92% | 61% | 94% | 72% | 39% | 6% |
| 2024Q1 | 94% | 100% | 42% | 97% | 92% | 28% | 3% |
| 2024Q2 | 86% | 97% | 50% | 97% | 64% | 8% | 0% |
| 2024Q3 | 86% | 97% | 58% | 97% | 86% | 22% | 3% |
| 2024Q4 | 81% | 100% | 61% | 97% | 78% | 11% | 0% |
| 2025Q1 | 86% | 97% | 53% | 86% | 72% | 11% | 3% |
| 2025Q2 | 81% | 94% | 31% | 81% | 53% | 8% | 0% |
| 2025Q3 | 50% | 86% | 17% | 61% | 50% | 3% | 0% |
| 2025Q4 | 11% | 39% | 0% | 25% | 8% | 0% | 0% |
| 2026Q1 | 6% | 11% | 0% | 6% | 0% | 0% | 0% |
| 2026Q2 | 0% | 0% | 0% | 0% | 0% | 3% | 0% |
| 2026Q3 | 0% | 3% | 0% | 0% | 0% | 0% | 0% |
| 2026Q4 | 0% | 0% | 0% | 0% | 0% | 0% | 0% |
## Estimate arm: hits among "do not remember" (guessing baseline)
| quarter | opus-5-5 | sonnet-5-5 | haiku-5-5 | fable-5-1 | opus-5 | sonnet-5 | haiku-4-5 |
|---|---|---|---|---|---|---|---|
| 2023Q1 | 75% | 40% | 59% | 86% | 82% | 84% | 25% |
| 2023Q2 | 67% | - | 65% | 67% | 89% | 76% | 9% |
| 2023Q3 | 100% | 100% | 89% | 100% | 100% | 100% | 42% |
| 2023Q4 | 100% | 100% | 79% | 100% | 90% | 86% | 35% |
| 2024Q1 | 100% | - | 76% | 100% | 100% | 92% | 29% |
| 2024Q2 | 100% | 100% | 72% | 100% | 100% | 85% | 36% |
| 2024Q3 | 100% | 100% | 80% | 100% | 80% | 93% | 37% |
| 2024Q4 | 100% | - | 93% | 100% | 100% | 97% | 36% |
| 2025Q1 | 100% | 0% | 88% | 100% | 90% | 91% | 26% |
| 2025Q2 | 100% | 100% | 88% | 100% | 100% | 76% | 0% |
| 2025Q3 | 100% | 100% | 93% | 100% | 94% | 74% | 3% |
| 2025Q4 | 100% | 95% | 94% | 93% | 85% | 56% | 3% |
| 2026Q1 | 97% | 94% | 89% | 85% | 69% | 19% | 3% |
| 2026Q2 | 64% | 83% | 53% | 58% | 42% | 17% | 3% |
| 2026Q3 | 58% | 54% | 44% | 44% | 50% | 17% | 3% |
| 2026Q4 | 17% | 25% | 42% | 42% | 25% | 17% | 8% |
## Estimate arm: share of estimates more than 1 month late
| quarter | opus-5-5 | sonnet-5-5 | haiku-5-5 | fable-5-1 | opus-5 | sonnet-5 | haiku-4-5 |
|---|---|---|---|---|---|---|---|
| 2023Q1 | 3% | 6% | 14% | 3% | 8% | 11% | 67% |
| 2023Q2 | 0% | 8% | 14% | 0% | 0% | 0% | 64% |
| 2023Q3 | 0% | 3% | 6% | 0% | 0% | 0% | 42% |
| 2023Q4 | 0% | 0% | 8% | 0% | 3% | 0% | 58% |
| 2024Q1 | 0% | 0% | 14% | 0% | 0% | 6% | 58% |
| 2024Q2 | 0% | 0% | 8% | 0% | 0% | 6% | 39% |
| 2024Q3 | 0% | 0% | 0% | 0% | 0% | 0% | 42% |
| 2024Q4 | 0% | 0% | 3% | 0% | 0% | 0% | 11% |
| 2025Q1 | 0% | 3% | 3% | 0% | 0% | 3% | 0% |
| 2025Q2 | 0% | 0% | 8% | 0% | 0% | 6% | 6% |
| 2025Q3 | 0% | 0% | 3% | 0% | 3% | 3% | 0% |
| 2025Q4 | 0% | 3% | 0% | 6% | 14% | 6% | 0% |
| 2026Q1 | 0% | 3% | 3% | 11% | 19% | 11% | 3% |
| 2026Q2 | 28% | 11% | 31% | 22% | 31% | 11% | 3% |
| 2026Q3 | 22% | 6% | 3% | 11% | 6% | 0% | 3% |
| 2026Q4 | 33% | 0% | 0% | 8% | 8% | 0% | 0% |
## Estimate arm: nonexistent versions, "remember"
claude-opus-5-5 0/200
claude-sonnet-5-5 4/200
claude-haiku-5-5 0/200
claude-fable-5-1 0/200
claude-opus-5 2/200
claude-sonnet-5 0/200
claude-haiku-4-5 5/200
## Usage
claude-opus-5 {'calls': 157, 'input_tokens': 316, 'output_tokens': 284883, 'cache_read_input_tokens': 12391, 'cache_creation_input_tokens': 145422, 'cost_usd_x1000': 8576, 'failed': 23}
claude-haiku-4-5 {'calls': 151, 'input_tokens': 129938, 'output_tokens': 1022619, 'cache_read_input_tokens': 0, 'cache_creation_input_tokens': 0, 'cost_usd_x1000': 5244, 'failed': 29}
claude-fable-5-1 {'calls': 155, 'input_tokens': 310, 'output_tokens': 196274, 'cache_read_input_tokens': 12990, 'cache_creation_input_tokens': 143741, 'cost_usd_x1000': 12693, 'failed': 22}
claude-sonnet-5 {'calls': 157, 'input_tokens': 32982, 'output_tokens': 353455, 'cache_read_input_tokens': 0, 'cache_creation_input_tokens': 135100, 'cost_usd_x1000': 4132, 'failed': 21}
claude-haiku-5-5 {'calls': 148, 'input_tokens': 298, 'output_tokens': 306258, 'cache_read_input_tokens': 9560, 'cache_creation_input_tokens': 143988, 'cost_usd_x1000': 174, 'failed': 25}
claude-sonnet-5-5 {'calls': 154, 'input_tokens': 308, 'output_tokens': 118376, 'cache_read_input_tokens': 10900, 'cache_creation_input_tokens': 145823, 'cost_usd_x1000': 1785, 'failed': 25}
claude-opus-5-5 {'calls': 153, 'input_tokens': 306, 'output_tokens': 159237, 'cache_read_input_tokens': 9349, 'cache_creation_input_tokens': 146112, 'cost_usd_x1000': 4357, 'failed': 23}
TOTAL {'calls': 1075, 'input_tokens': 164458, 'output_tokens': 2441102, 'cache_read_input_tokens': 55190, 'cache_creation_input_tokens': 860186, 'cost_usd_x1000': 36961, 'failed': 168}
# Part 2. intervals.py
## Measured cutoff: point, 95% bootstrap, per repeat
| model | vendor | measured (both repeats) | 2024 baseline | 95% bootstrap | repeat 0 | repeat 1 | months before vendor |
|---|---|---|---|---|---|---|---|
| claude-opus-5-5 | 2026-06 | 2025-09 | 0.941 | 2025-08 .. 2025-10 | 2025-10 | 2025-08 | 9 |
| claude-sonnet-5-5 | 2026-06 | 2025-12 | 0.955 | 2025-12 .. 2026-01 | 2025-12 | 2026-01 | 6 |
| claude-haiku-5-5 | 2026-06 | 2025-09 | 0.812 | 2025-08 .. 2025-10 | 2025-09 | 2025-09 | 9 |
| claude-fable-5-1 | 2026-06 | 2025-11 | 1.000 | 2025-10 .. 2025-11 | 2025-11 | 2025-11 | 7 |
| claude-opus-5 | 2026-05 | 2025-10 | 0.972 | 2025-09 .. 2025-10 | 2025-10 | 2025-10 | 7 |
| claude-sonnet-5 | 2026-01 | 2025-01 | 0.719 | 2024-12 .. 2025-02 | 2025-01 | 2025-01 | 12 |
| claude-haiku-4-5 | 2025-02 | 2024-12 | 0.043 | 2024-04 .. 2025-01 | 2025-01 | 2024-12 | 2 |
## False recognition (abstention arm), Wilson 95%
| model | known / answers | share | 95% CI | repeat 0 | repeat 1 |
|---|---|---|---|---|---|
| claude-opus-5-5 | 0/400 | 0.0% | 0.0..1.0% | 0/200 | 0/200 |
| claude-sonnet-5-5 | 0/400 | 0.0% | 0.0..1.0% | 0/200 | 0/200 |
| claude-haiku-5-5 | 3/400 | 0.8% | 0.3..2.2% | 1/200 | 2/200 |
| claude-fable-5-1 | 1/400 | 0.2% | 0.0..1.4% | 1/200 | 0/200 |
| claude-opus-5 | 4/400 | 1.0% | 0.4..2.5% | 2/200 | 2/200 |
| claude-sonnet-5 | 2/394 | 0.5% | 0.1..1.8% | 1/200 | 1/194 |
| claude-haiku-4-5 | 20/392 | 5.1% | 3.3..7.7% | 13/196 | 7/196 |
## Monthly, abstention arm: correct / answers (both repeats), 2025-01..2026-04
| month | opus-5-5 | sonnet-5-5 | haiku-5-5 | fable-5-1 | opus-5 | sonnet-5 | haiku-4-5 |
|---|---|---|---|---|---|---|---|
| 2025-01 | 22/24 | 22/24 | 16/24 | 23/24 | 24/24 | 9/23 | 0/24 |
| 2025-02 | 21/24 | 24/24 | 20/24 | 24/24 | 24/24 | 8/24 | 0/24 |
| 2025-03 | 22/24 | 23/24 | 21/24 | 24/24 | 24/24 | 3/24 | 0/24 |
| 2025-04 | 24/24 | 23/24 | 20/24 | 24/24 | 22/24 | 4/24 | 0/24 |
| 2025-05 | 23/24 | 24/24 | 17/24 | 24/24 | 23/24 | 2/24 | 0/23 |
| 2025-06 | 17/24 | 21/24 | 18/24 | 22/24 | 24/24 | 5/24 | 0/24 |
| 2025-07 | 20/24 | 23/24 | 16/24 | 23/24 | 24/24 | 2/24 | 0/23 |
| 2025-08 | 11/24 | 21/24 | 17/24 | 20/24 | 17/24 | 0/24 | 0/22 |
| 2025-09 | 13/24 | 22/24 | 10/24 | 20/24 | 21/24 | 0/24 | 0/24 |
| 2025-10 | 10/24 | 24/24 | 10/24 | 21/24 | 11/24 | 0/24 | 0/24 |
| 2025-11 | 7/24 | 21/24 | 5/24 | 14/24 | 6/24 | 2/24 | 0/24 |
| 2025-12 | 3/24 | 17/24 | 1/24 | 6/24 | 2/24 | 0/24 | 0/24 |
| 2026-01 | 5/24 | 13/24 | 2/24 | 5/24 | 3/24 | 0/24 | 0/23 |
| 2026-02 | 0/24 | 3/24 | 0/24 | 0/24 | 1/24 | 0/24 | 0/22 |
| 2026-03 | 0/24 | 0/24 | 0/24 | 0/24 | 0/24 | 0/24 | 0/24 |
| 2026-04 | 0/24 | 0/24 | 0/24 | 0/24 | 0/24 | 0/23 | 0/24 |
## Width of the decline (3 month rolling window)
Last month where the window is at least 90% of the 2024 baseline, and the first month after it where the window is at most 10% of the baseline.
| model | 2024 baseline | last >= 90% | first <= 10% | months |
|---|---|---|---|---|
| claude-opus-5-5 | 0.941 | 2025-05 | 2026-02 | 9 |
| claude-sonnet-5-5 | 0.955 | 2025-11 | 2026-03 | 4 |
| claude-haiku-5-5 | 0.812 | 2025-05 | 2026-01 | 8 |
| claude-fable-5-1 | 1.000 | 2025-07 | 2026-02 | 7 |
| claude-opus-5 | 0.972 | 2025-07 | 2026-01 | 6 |
| claude-sonnet-5 | 0.719 | 2024-11 | 2025-08 | 9 |
| claude-haiku-4-5 | 0.043 | 2024-11 | 2025-02 | 3 |
A perfect step (all correct through month m, none after) reads as 3 months on the 3 month window: 1 at m-1, 2/3 at m, 1/3 at m+1, 0 at m+2. That is the floor of the table above.
## Width of the decline, unsmoothed monthly shares
The same thresholds on the monthly share itself, compared exactly (fractions, no rounding). The floor for a perfect step is 1 month.
| model | last >= 90% | first <= 10% | months |
|---|---|---|---|
| claude-opus-5-5 | 2025-05 | 2026-02 | 9 |
| claude-sonnet-5-5 | 2025-11 | 2026-03 | 4 |
| claude-haiku-5-5 | 2025-06 | 2025-12 | 6 |
| claude-fable-5-1 | 2025-07 | 2026-02 | 7 |
| claude-opus-5 | 2025-09 | 2025-12 | 3 |
| claude-sonnet-5 | 2024-11 | 2025-08 | 9 |
| claude-haiku-4-5 | 2024-12 | 2025-01 | 1 |
claude-opus-5 sits on a tie: 21/24 in 2025-09 equals 90% of its 2024 baseline of 35/36 exactly. Under a strict comparison the width reads 5 (2025-07..2025-12).
## Last month above 10%, correct share and known share (both repeats)
A comparison with HindsightBench, chosen after the data. Its effective cutoff is the last month with outcome recall propensity above 0.1; the closest analog here is the share of answers with known=true, the share correct is given beside it. Strictly above 10%, monthly, unsmoothed.
| model | last month correct > 10% | last month known > 10% |
|---|---|---|
| claude-opus-5-5 | 2026-01 (5/24) | 2026-01 (5/24) |
| claude-sonnet-5-5 | 2026-02 (3/24) | 2026-02 (3/24) |
| claude-haiku-5-5 | 2025-11 (5/24) | 2025-12 (4/24) |
| claude-fable-5-1 | 2026-01 (5/24) | 2026-01 (5/24) |
| claude-opus-5 | 2026-01 (3/24) | 2026-01 (3/24) |
| claude-sonnet-5 | 2025-06 (5/24) | 2025-06 (5/24) |
| claude-haiku-4-5 | 2024-01 (4/23) | 2024-08 (3/24) |
## haiku-4-5 monthly 2024-07..2025-09
| month | abstention: known | abstention: correct | answers | estimate: hit | answers |
|---|---|---|---|---|---|
| 2024-07 | 2 | 0 | 21 | 4 | 12 |
| 2024-08 | 3 | 1 | 24 | 4 | 12 |
| 2024-09 | 0 | 0 | 23 | 5 | 12 |
| 2024-10 | 1 | 1 | 24 | 6 | 12 |
| 2024-11 | 2 | 1 | 24 | 4 | 12 |
| 2024-12 | 1 | 1 | 22 | 3 | 12 |
| 2025-01 | 0 | 0 | 24 | 3 | 12 |
| 2025-02 | 0 | 0 | 24 | 6 | 12 |
| 2025-03 | 0 | 0 | 24 | 1 | 12 |
| 2025-04 | 0 | 0 | 24 | 0 | 12 |
| 2025-05 | 0 | 0 | 23 | 0 | 12 |
| 2025-06 | 0 | 0 | 24 | 0 | 12 |
| 2025-07 | 0 | 0 | 23 | 1 | 12 |
| 2025-08 | 0 | 0 | 22 | 0 | 12 |
| 2025-09 | 0 | 0 | 24 | 0 | 12 |
## haiku-4-5 by window (Wilson 95%)
| window | abstention: correct | estimate: hit |
|---|---|---|
| 2024-01..2024-12 | 12/275 = 4.4% [2.5..7.5] | 50/144 = 34.7% [27.4..42.8] |
| 2025-01..2025-02 | 0/48 = 0.0% [0.0..7.4] | 9/24 = 37.5% [21.2..57.3] |
| 2025-03..2025-07 | 0/118 = 0.0% [0.0..3.2] | 2/60 = 3.3% [0.9..11.4] |
| 2025-08..2026-10 | 0/351 = 0.0% [0.0..1.1] | 5/180 = 2.8% [1.2..6.3] |
## Check outside CRITERIA: the same rule on the estimate arm, share "remember and hit", 1 repeat
| model | cutoff | 2024 baseline |
|---|---|---|
| claude-opus-5-5 | 2025-08 | 0.868 |
| claude-sonnet-5-5 | 2025-10 | 0.979 |
| claude-haiku-5-5 | 2025-07 | 0.486 |
| claude-fable-5-1 | 2025-09 | 0.972 |
| claude-opus-5 | 2025-09 | 0.799 |
| claude-sonnet-5 | 2025-07 | 0.167 |
| claude-haiku-4-5 | 2025-03 | 0.007 |
## Estimate arm: hits among "do not remember"
| model | 2023-01..2025-06 | 2026-04..2026-10 |
|---|---|---|
| claude-opus-5-5 | 53/56 = 95% [85..98] | 46/84 = 55% [44..65] |
| claude-sonnet-5-5 | 10/14 = 71% [45..88] | 52/83 = 63% [52..72] |
| claude-haiku-5-5 | 142/181 = 78% [72..84] | 40/84 = 48% [37..58] |
| claude-fable-5-1 | 30/32 = 94% [80..98] | 42/84 = 50% [40..60] |
| claude-opus-5 | 92/99 = 93% [86..97] | 36/84 = 43% [33..54] |
| claude-sonnet-5 | 254/290 = 88% [83..91] | 14/83 = 17% [10..26] |
| claude-haiku-4-5 | 96/351 = 27% [23..32] | 3/84 = 4% [1..10] |
## Frontier per package (2 repeats)
| model | answers | median | p90 | max | nonexistent versions |
|---|---|---|---|---|---|
| claude-opus-5-5 | 200 | 2025-08 | 2025-10 | 2025-12 | 2 |
| claude-sonnet-5-5 | 200 | 2025-10 | 2025-12 | 2026-01 | 3 |
| claude-haiku-5-5 | 200 | 2025-09 | 2025-10 | 2025-12 | 5 |
| claude-fable-5-1 | 200 | 2025-09 | 2025-11 | 2026-03 | 4 |
| claude-opus-5 | 200 | 2025-10 | 2026-02 | 2026-08 | 13 |
| claude-sonnet-5 | 200 | 2025-02 | 2025-09 | 2026-01 | 4 |
| claude-haiku-4-5 | 154 | 2024-02 | 2024-07 | 2025-02 | 7 |
## Self report, 6 wordings (phase self, 130 of 210 taken)
| model | answers | distribution |
|---|---|---|
| claude-opus-5-5 | 18 | 2026-06 x18 |
| claude-sonnet-5-5 | 19 | 2026-06 x18, 2026-05 x1 |
| claude-haiku-5-5 | 13 | 2026-06 x12, 2026-05 x1 |
| claude-fable-5-1 | 20 | 2026-06 x15, 2026-05 x2, 2026-04 x1, 2025-01 x1, 2025-05 x1 |
| claude-opus-5 | 22 | 2026-05 x21, 2026-03 x1 |
| claude-sonnet-5 | 22 | 2026-01 x18, 2025-12 x3, 2025-02 x1 |
| claude-haiku-4-5 | 16 | 2025-02 x16 |
## Self report asked to set the cutoff line aside (self2, 5 per model)
| model | answers | distribution |
|---|---|---|
| claude-opus-5-5 | 5 | 2025-01 x5 |
| claude-sonnet-5-5 | 5 | 2025-05 x2, 2025-09 x1, 2025-08 x1, 2025-06 x1 |
| claude-haiku-5-5 | 5 | 2025-03 x2, 2025-01 x2, 2025-05 x1 |
| claude-fable-5-1 | 5 | 2025-01 x3, 2025-03 x1, 2025-05 x1 |
| claude-opus-5 | 5 | 2025-01 x5 |
| claude-sonnet-5 | 5 | 2025-02 x3, 2025-04 x2 |
| claude-haiku-4-5 | 5 | 2024-04 x3, 2025-02 x2 |
## Estimate arm: most frequent estimates of claude-opus-5-5 on releases from 2026-07..2026-10
48 answers: 2026-08 x15, 2026-07 x5, 2026-10 x4, 2027-01 x4
## Answering model
Successful rows whose modelUsage has the requested model as its only key: 1075 of 1075.
## Journal
Successful rows: batch0 266, batch1 266, front0 56, front1 56, guess0 266, self 130, self2 35; total 1075.
Failed rows, excluded: self2 "None Failed to authenticate: OAuth session expired and could" 10; guess "None Failed to authenticate: OAuth session expired and could" 153; guess "non-JSON: Warning: no stdin data received in 3s, proceeding " 5; total 168.