🔍 Read the full analysis: Claude Fable 5.2 And AI Progress: A Recalculation Across Three Benchmarks on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 12 analysis of leaderboard data finds Claude Fable 5.2 above estimated progress trends on the Remote Labor Index, Epoch Capabilities Index and RoboDojo. The strongest result is on freelance-style tasks; the Epoch lead falls within the benchmark’s uncertainty range, while the RoboDojo figure comes from Anthropic and varies substantially with the trend-fitting method.
Claude Fable 5.2 scored above estimated progress trends on three AI benchmarks, according to a November 12 analysis that rebuilt the comparisons from leaderboard data and Anthropic’s launch materials. The clearest lead was on a test of paid freelance projects, while the Epoch Capabilities Index result was small enough to fall within its uncertainty range, and the RoboDojo score was reported by Anthropic rather than verified on its leaderboard.
On the Remote Labor Index (RLI), Fable 5.2 passed 41.3% of 240 real freelance projects, or 99 tasks, at a standard intended to represent work a reasonable client would accept. The analysis estimates a trendline value of 27.2% on the model’s November 9 release date: a gap of 14.1 percentage points. At the fitted trend’s pace, reaching 41.3% would take about 90 days. The source analysis says its result remains between 13.0 and 14.4 points above trend under other reasonable fitting choices.
The comparison is less decisive on the Epoch Capabilities Index (ECI). Fable 5.2 scored 173.4, compared with an estimated trend value of 171.1, a difference of 2.4 points, or about 1.7 months at the trend’s pace. Epoch AI’s uncertainty range for Fable 5.2 is 169.9 to 177.5, which extends below the trend value. That means the score is above the point estimate, but the available uncertainty information does not establish a clear separation.
For RoboDojo, Anthropic reported a 64.9% success rate on simulated robot-arm tasks. The analysis places that 8.8 points above its trend fit, but says the score is an internal Anthropic run, not a verified leaderboard entry. The leaderboard includes Q0-Preview at 68.9%. Depending on the fitting choices, the estimated RoboDojo gap ranges from 3.8 to 23.3 points, making this the least stable of the three comparisons.
Claude Fable 5.2 Across Three Benchmarks
A November 12 analysis found scores above estimated trends on three benchmarks, with evidence varying in strength. The clearest lead was on freelance-style projects.
What the scores show
Each score sits above its fitted trend estimate, but the comparisons are not equally decisive.
Clearest measured lead
Fable 5.2 passed 41.3% of 240 projects, versus a 27.2% trend estimate.
Uncertain difference
Its Epoch score was 173.4 against a 171.1 trend estimate, within the stated uncertainty range.
Provisional result
Anthropic reported 64.9% on RoboDojo; the result is not a verified leaderboard entry.
How to read “above trend”
The analysis compared Fable 5.2 with curves fitted to earlier record-setting models.
The analysis used a logistic curve for percentage benchmarks and a straight line for the Epoch index.
The gap measures how far the new score sits above the curve’s estimate for its release date.
Uncertainty, verification and trend-fitting choices affect how firmly each gap can be interpreted.
Evidence strength differs
The three results support different levels of confidence for readers comparing models.
| Benchmark | What supports the comparison | Limit |
|---|---|---|
| Remote Labor Index | Largest reported gap; remains 13.0–14.4 points across fitting choices | 240 projects; no leaderboard confidence interval |
| Epoch Capabilities Index | Score exceeds the trend point estimate | Uncertainty range extends below the trend |
| RoboDojo | Reported score is above the fitted estimate | Company-reported; gap varies from 3.8 to 23.3 points |
Freelance Work Shows the Clearest Lead
The results suggest that Fable 5.2’s most substantial measured improvement is in multi-step, client-style work, not a uniform jump across every kind of AI capability. On the RLI, its 41.3% pass rate is more than twice Claude Fable 5.1’s 17.92% leaderboard score. That is a notable change on a test built around completed projects rather than short, isolated questions.
Still, the headline figure has a practical limit: most projects did not meet the acceptance standard. The RLI’s 41.3% is a share of projects that passed, not a measure of how much work the model can reliably automate across all jobs. The comparison also depends on a sample of 240 projects and on how reviewers judge whether work meets a reasonable-client bar. The source analysis estimates a rough sampling margin of about plus or minus six percentage points; it says that would not erase the 14-point trend gap, though the leaderboard itself does not provide a confidence interval.
For readers comparing models, the distinction between a raw score and a score’s position against an expected trend matters. A model can set a new high without showing that progress accelerated. Here, the RLI result is the strongest evidence of an above-trend performance, while the ECI and RoboDojo results call for more caution. None of the three comparisons alone establishes how a model will perform on a particular organization’s tasks.
Popular right now
Vitamix VX1 Professional-Grade Blender, 64 oz, Black
Roborock RockNeo Q110H Wire-Free Robot Lawn Mower
Dyson V8 Motorhead Origin Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
How the Three Trends Were Rebuilt
The analysis defines “above trend” by comparing each new record-holder with a curve fitted to earlier record-setting models, excluding Fable 5.2 from the fits. It uses a logistic curve for the two percentage-based benchmarks, which cannot exceed 100%, and a straight line for Epoch’s index. The vertical gap shows how far a score sits above the estimated value for its release date; the time estimate represents how long the fitted trend would take to reach that score.
For the RLI, Scale AI and the Center for AI Safety assembled 240 projects from 358 Upwork freelancers across 23 categories, including game development, product design, architecture and data analysis. Scale AI says a human took an average of 28.9 hours per project, with a median of 11.5 hours; the average project was valued at $632.60. Human reviewers compare AI deliverables with professional work. According to Anthropic’s system card, Scale AI ran the Claude models at high effort.
The RLI trend uses 11 record-setting entries with release dates available in Epoch AI’s dataset, from Gemini 2.5 Pro in June 2025 through Claude Opus 5.5 in September 2026. Four agent products without dates in that dataset were excluded. The published chart reportedly gave a 27.1% trend value and a 14.2-point gap, compared with the analysis’s 27.2% and 14.1 points; the source says dating differences likely account for the small discrepancy.
“Absolute ECI values are meaningless by themselves, but meaningful comparisons can be made between models.”
— Epoch AI, describing its Capabilities Index
Limits of the Benchmark Comparisons
The three findings do not carry equal evidentiary weight. The ECI’s reported uncertainty range overlaps the trend value, so its 2.4-point lead should not be read as a firm acceleration signal. For the RLI, the source analysis gives a rough margin of about six percentage points but notes that the leaderboard has no confidence interval; the exact statistical uncertainty remains unspecified.
The RoboDojo comparison is more provisional: Anthropic supplied the 64.9% result, and the source analysis says it is not a verified leaderboard entry. The fitted gap changes considerably under alternative methods. The material provided also does not detail the robot-task run’s sample size or provide an independent replication. More broadly, trendline estimates depend on the selected record-setters, their dates and the chosen curve, and should not be treated as direct measurements of general intelligence or workplace productivity.
The article was researched and drafted with AI assistance, including Anthropic’s Claude, which is the subject of the report. Its author disclosed that Anthropic, Scale AI, Epoch AI and the RoboDojo team did not review or sponsor the piece. The reported scores and the independent trend calculations are distinct: some underlying figures come from benchmark or company sources, while the trend comparisons are the author’s analysis.
Independent Checks and New Scores
The next useful evidence would be independent RoboDojo verification and additional models tested under comparable conditions. Further RLI results could show whether Fable 5.2’s 41.3% score is sustained across a larger or different set of projects, while clearer uncertainty reporting would help readers judge how much the observed gap could reflect sampling variation.
For the ECI, later model results and updates to Epoch AI’s uncertainty estimates may clarify whether the apparent 2.4-point lead marks a lasting departure from trend or falls within ordinary measurement variation. Until then, the defensible conclusion is narrower: Fable 5.2 is above the fitted point estimates on all three comparisons, with strong evidence of a substantial RLI gap, an uncertain ECI difference and a RoboDojo claim that still needs independent confirmation.
Key Questions
What does “above trend” mean in this analysis?
It means a model’s score is higher than the value predicted for its release date by a curve fitted to earlier benchmark record-setters. It is a comparison with an estimated historical pace, not a separate benchmark score.
Which Fable 5.2 result is the clearest?
The Remote Labor Index result is the clearest: 41.3% of 240 projects passed, compared with a fitted trend value of 27.2%. The analysis estimates the gap at 14.1 percentage points and says it stays between 13.0 and 14.4 points under other reasonable fits.
Does the Epoch score prove faster progress?
No. Fable 5.2’s ECI score of 173.4 is 2.4 points above the estimated trend value of 171.1, but Epoch AI’s uncertainty range extends from 169.9 to 177.5 and includes values below the trend.
Has Anthropic’s RoboDojo score been independently verified?
The source analysis says no: Anthropic reported the 64.9% success rate, and it was not a verified leaderboard entry. The same leaderboard lists Q0-Preview at 68.9%.
Source: ThorstenMeyerAI.com
