Claude Fable 5.2 and the 41.3% Remote Labor Index Score: What It Actually Measures
The viral PERFORMANCE COMPARISON chart is Scale's Remote Labor Index. Fable 5.2 is a real Anthropic model; here is what 41.3% means, what it doesn't, and how other verified boards compare.
Short answer: Claude Fable 5.2 is a real Anthropic model, released on 9 November 2026, not a spoof or a wrapper alias. The unlabeled "PERFORMANCE COMPARISON" chart that circulated in a FMKorea community post on 10 November 2026 is the Scale Labs Remote Labor Index (RLI). On that board, fetched at 11:23 KST the same day, Fable 5.2 leads at 41.30% automation rate — almost double Claude Opus 5.5 at 21.25% and GPT-6 Astra at 20.83%. That number is the share of private freelance-style projects judged at least as good as a human professional deliverable. It is not a claim that the model beats every rival on every task, and it still means most of those projects would not be accepted as commissioned work.
What the 41.30 figure measures
RLI was introduced by researchers at the Center for AI Safety and Scale AI. The dataset has 240 end-to-end projects sourced from 358 verified Upwork freelancers across 23 of 64 Upwork domains. Each project includes a brief, input files, a human gold-standard deliverable, and the reported human time and cost. 230 projects are a private set used for official scoring; 10 are public. The Scale page lists a total economic value of $143,991, a mean human completion time of 28.9 hours (median 11.5), and an average project value of $632.60 (median $200).
The primary metric, automation rate, is the percentage of projects where a reviewer judges the AI deliverable at least as good as the human reference, in the sense that a "reasonable client" would accept it. Evaluations are manual, with Scale reporting 94.4% inter-annotator agreement by majority vote of three reviewers. The page states there is no confidence interval for automation rate.
The chart below uses only the scores on Scale's live leaderboard as of 10 November 2026, 11:23 KST. Model names are written as the Scale page shows them. The axis starts at zero. The circulating community screenshot showed the same top rows, including Scale's odd display of rank 13 twice; the official board continues below 2.50, down to gemini-2.5-pro-preview-06-05 at 0.83.
| Rank (by score) | Model (Scale label) | Automation rate (%) |
|---|---|---|
| 1 | Fable 5.2 | 41.30 |
| 2 | Opus 5.5 | 21.25 |
| 3 | GPT 6 Astra | 20.83 |
| 4 | Fable 5.1 | 17.92 |
| 5 | Gemini Argon 4 | 17.08 |
| 6 | Fable 5 | 15.83 |
| 7 | Opus 4.8 | 8.33 |
| 8 | Codex GPT 5.5 | 6.25 |
| 9 | Gemini 3.8 Flash | 5.83 |
| 10 | Gemini 3.7 Flash | 5.00 |
| 11 | claude-opus-4-6 (CoWork) | 4.17 |
| 12 | claude-opus-4-5-20251101-thinking | 3.75 |
| 13 | Manus_1.6 (Max) | 2.92 |
| 14 (tie) | gpt-5.2-2025-12-11 (medium) | 2.50 |
| 14 (tie) | Manus 1.5 | 2.50 |
| 14 (tie) | Manus 1.0 | 2.50 |
Sources: Scale Labs RLI leaderboard, RLI project site, RLI paper. The community post that circulated the chart: FMKorea, 10 Nov 2026, 10:04.
Anthropic's own launch table reports the same comparison in rounded form: Fable 5.2 41.3% at High effort, Fable 5.1 17.9%, Opus 5.5 21.3%, GPT-6 Astra 20.8%. That matches Scale once you allow ordinary rounding. Fable 5.2's default API effort is high, so those two statements agree. The Scale row itself does not print a separate "self-reported" or "independently evaluated" tag, so this article treats 41.30 as the figure on the official RLI board and 41.3 as Anthropic's rounded High-effort number, without claiming a second, independent lab run beyond what those pages show.
What RLI does not measure
A 41.30% automation rate is a large jump from the previous leaders near 21%, but it is still a minority of projects. The original RLI paper, written when the best agent sat at 2.5%, is useful for the metric and the dataset. It is not a description of the current leaderboard. Scale's overview text on the same page still says that at launch the highest agent (Manus) reached 2.5%; that sentence is stale. The "Performance Comparison" table is the current one.
The benchmark also does not represent all remote work. Scale excludes projects that need physical labor, long-term evaluation (for example SEO), or direct client interaction such as tutoring or customer service. Failed projects in the original analysis clustered around poor or amateur-quality work (45.6%), incomplete deliverables (35.7%), broken or wrong file formats (17.6%), and inconsistencies across files (14.8%). Those percentages describe the early evaluation, not a fresh audit of Fable 5.2's failures.
Compared with a real job, RLI gives the agent a written brief and files, not a back-and-forth client. Scale standardized tooling and set a $30 generation budget per task. Your repository, your review process, and your budget will not match that harness. A model that wins RLI can still lose a social-deduction game, a cheap coding task, or a board that saturates at 100%.
Our October 2026 leaderboard roundup listed Opus 5.5 first on RLI at 21.25, which was correct for that snapshot. That row is now out of date; use the Scale page above. See LLM Leaderboards Roundup (Oct 2026).
Fable 5.2 is an official model
Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. The product page is Claude Fable. The API model ID is claude-fable-5-2. Anthropic says Fable and Mythos are the same underlying model. Fable ships with cybersecurity and biology safeguards and is generally available. Mythos relaxes those safeguards and is limited to organizations in Anthropic's Life Sciences and Cyber Verification Programs.
Anthropic's own guidance is not "use Fable for everything." For most coding and knowledge work, it recommends starting with Claude Opus 5.5, then moving up to Fable or Mythos only for the hardest, most open-ended technical work, or down to Sonnet 5.5 or Haiku 5.5 when that is enough.
Documented API pricing, from Anthropic's Fable page and the Claude API pricing page:
| Item | Fable 5.2 / Mythos 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| Input / million tokens | $8 | $10 | $4 |
| Output / million tokens | $40 | $50 | $20 |
| Cache read / million tokens | $0.20 | $0.25 | $0.20 |
| Fast mode input / output | $16 / $80 | — | $8 / $40 |
| Default API effort | high | — | medium |
| Context window | 1M tokens | — | 1M tokens |
Fable 5.2's list price is 20% below Fable 5.1. Fast mode is documented as up to 2.5× faster token generation at double the token price. Anthropic also says outputs can be up to 30% faster than Fable 5.1 at the standard price. Fable 5.2 is listed as available to Pro, Max, Ultra, Team, and Enterprise users, and on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. The docs give a June 2026 knowledge cutoff and say Fable requires 30-day data retention for safety monitoring unless a customer is authorized for a different arrangement.
Other charts from the same post, checked against official boards
The FMKorea post is mostly screenshots, with almost no body text and no outbound source links. Several of those images do match official boards. A few do not have a source we could confirm.
ARC-AGI-3 (ARC Prize, verified). The official ARC-AGI leaderboard and the Claude Fable 5.2 results page dated 9 November 2026 match the table in the post:
| Model | Date on ARC board | ARC-AGI-3 Standard | ARC-AGI-3 Provider Adapter |
|---|---|---|---|
| Claude Fable 5.2 | 2026-11-09 | 81.2% (1st) | 97.7% (2nd) |
| GPT-6 Astra | 2026-09-02 | 62.7% (2nd) | 99.9% (1st) |
| GPT-6.1 Sol | 2026-09-29 | 52.7% (3rd) | 96.4% |
| Gemini 4 Argon | 2026-10-13 | 49.8% | 97.5% (3rd) |
Standard and Adapter are different harnesses. The Standard harness lets the model carry notes forward. The Provider Adapter preserves opaque reasoning state and uses compaction, which can reuse prior work. Astra still leads the Adapter column; Fable 5.2 leads Standard. A high Adapter score is not the same result as a high Standard score. ARC also reports Fable 5.2 at 98.5% on ARC-AGI-1 and 93.8% on ARC-AGI-2 at XHigh reasoning — strong, but those older sets are much closer to saturation than ARC-AGI-3 Standard.
Artificial Analysis Intelligence Index. Artificial Analysis's Claude Fable 5.2 page lists Claude Fable 5.2 (Max, Default Fallback) at 65 on Intelligence Index v5.0.1, first among the models on that page, with a 9 November 2026 release date. The same page says the model is expensive and slow on this suite: $26.54 and 46.4 minutes per Index task in Artificial Analysis's figures, versus medians of $2.00 input / $10.00 output pricing and 12.7 minutes in its comparison class. We did not re-transcribe every tiny bar label in the circulating screenshot; use the live Artificial Analysis page for the full ranking.
FrontierMath Tier 4 (v2). An unlabeled 100% bar chart in the post lines up with Epoch AI's FrontierMath Tier 4 (v2), a private set of exceptionally hard math problems. Epoch's Claude Fable 5.2 model page lists Fable 5.2 at 100%, marked as the benchmark's best score, with Claude Opus 5.5 at 95% and GPT-6 Astra at 98% on the same comparison table. Those official figures are rounded; this article does not treat the screenshot's confidence intervals as Epoch's printed values. A 100% score on a 43-problem private set is not a general "solves all mathematics" result.
Blood on the Clocktower. The post included a Korean-translated screenshot attributed to Dwarkesh Patel (@dwarkesh_sp), saying Fable 5.2 still trails GPT-6 Astra on a social-deception game eval, while beating Opus 5.5 and Fable 5.1. We did not independently retrieve the original post, so we are not repeating unofficial Elo numbers from that screenshot. The qualitative point is still useful: a model can lead RLI or ARC-AGI-3 Standard and trail on a very different task.
Company-reported research claims. Anthropic says a team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in 30 hours, while Opus 5.5 reached only 26% in 18 hours under the same setup and barely moved after 30. That is formalization — translating an already-known proof into a language a computer can check — not a new proof of the theorem. Anthropic's earlier 4 September 2026 write-up of an 11-day formalization is here: Formalizing Fermat's Last Theorem. The 30-hour figure comes from the 9 November launch post.
Anthropic also says Fable 5.2 completed two previously unfinished problems in theoretical physics, and links two Anthropic-contributed preprints dated 7 November 2026: The Hawking-Page transition at weak coupling and The spinless modular bootstrap at large central charge. The first paper's own provenance statement says an internal Claude research model obtained the result, a physicist checked it against known special cases, and a separate Claude computation gave the same number. Those papers exist and disclose that origin. This article does not treat them as independently peer-reviewed confirmation of the scientific claims.
"Model 3" and "Model 4." The community post said Fable 5.2 "corresponds to Model 3" and that a smarter "Model 4" also exists, next to an unsourced progression chart. Those labels do not appear as official product names in Anthropic's Fable 5.2 announcement or model docs. Treat that graphic as unverified.
FAQ
Should I switch my daily driver from Opus 5.5 to Fable 5.2? Anthropic says no, not by default. Opus 5.5 is half the API list price ($4 / $20 vs $8 / $40) and is the recommended starting point for most coding and knowledge work. Fable 5.2 is for harder, longer, more ambiguous technical jobs, or as an advisor/orchestrator that other Claude models can consult.
Is 41.3% "almost half of remote work automated"? No. It is 41.3% of this specific private set of freelance projects, under Scale's harness and judgment standard. The set omits client-facing and long-running work, and 41.3% still means most projects in the set fail the bar.
Why do some charts show Fable 5.2 behind GPT-6 Astra? Because the boards measure different things. Astra leads ARC-AGI-3's Provider Adapter harness at 99.9%, and a third-party social-deception eval circulated in the same post also had Astra ahead. Fable 5.2 leads RLI, ARC-AGI-3 Standard, and Artificial Analysis's current Index page for this model. Harness, effort setting, and task type change the ranking.
Is Mythos 5.2 a stronger model? Anthropic presents it as the same model with fewer cyber and biology blocks, not as a higher-intelligence SKU. Access is gated. Some of Anthropic's cyber and biology scores in the launch table are reported for Mythos, or for other models with safeguards off, and are not a clean like-for-like comparison.
Caveats
- We did not run any of these benchmarks. Every score above is copied from the official page cited next to it, as fetched on 10 November 2026 (KST).
- Scale's RLI ranks are messy around ties; we ranked by the printed score.
- Effort settings and harnesses differ across Anthropic's own comparison table (for example, some rival scores are at Medium effort, some Opus cyber/biology scores are without safeguards). Read the system card before treating those rows as a single league table.
- Leaderboards move quickly. The October Terapep roundup is already stale on RLI.
Sources
- Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (9 Nov 2026)
- Claude Fable product page and Fable 5.2 API docs
- Claude API pricing
- Fable 5.2 / Mythos 5.2 system card
- Scale Labs Remote Labor Index, remotelabor.ai, RLI paper
- ARC Prize leaderboard and Fable 5.2 ARC results
- Artificial Analysis: Claude Fable 5.2
- Epoch AI: Claude Fable 5.2 and FrontierMath Tier 4 (v2)
- Anthropic: Formalizing Fermat's Last Theorem (4 Sep 2026)
- Preprints linked from the launch post: hexagon:2611.00175v1, hexagon:2611.00174v1
- Community discovery of the RLI chart: FMKorea post, 10 Nov 2026
- Earlier Terapep snapshot: LLM Leaderboards Roundup (Oct 2026)
← Back to all posts