Claude Fable 5.2 tops the public Remote Labor Index leaderboard at a 41.30% automation rate as of November 10, 2026, with Opus 5.5 at 21.25% and GPT-6 Astra at 20.83%. The headline is easy to garble: it does not mean AI replaced forty-one percent of the freelance economy. It means expert judges would accept Fable's full project deliverables on roughly four out of ten specific RLI tasks versus a paid human reference.
If you route agents for clients, the number matters because it is one of the few benchmarks built from real briefs, files, and invoices, not synthetic puzzles. Our Fable 5.2 versus Opus routing guide already cited the same row on Anthropic's table; this article unpacks what RLI measures, where it breaks, and how to read it next to frontier model comparisons and OpenAI's new freelance study.
TL;DR: the number in plain language
| Question | Answer | Source |
|---|---|---|
| What is being scored? | Full freelance-style projects, not single prompts | Scale RLI overview |
| Primary metric? | Automation rate — share judged client-acceptable vs human deliverable | Scale methodology page |
| Fable 5.2 score? | 41.30% on the live leaderboard | Scale Performance Comparison, Nov 10, 2026 |
| Dataset size? | 240 projects; 230 private for scoring; 10 public | Scale dataset section |
| Human baseline economics? | About $143,991 total value; median 11.5 hours per project | Scale statistics |
| Does 41% mean most jobs are automated? | No — excluded categories and failed projects still dominate | Scale limitations + failure modes |
Why RLI exists
Research benchmarks such as Agents Last Exam stress isolated reasoning. Customers hire freelancers for finished artifacts: a kitting animation, an architecture plan, a mixed-media ad, a data workbook. RLI, described on Scale's site and in the arXiv paper, tries to close that gap by sourcing 358 verified Upwork freelancers across 23 of 64 Upwork domains, then filtering 550 candidate briefs down to 240 attempts.
Scale explicitly drops work that needs physical labor, long-horizon evaluation (SEO-style), or live client interaction (tutoring, customer service). That boundary is why RLI is useful for digital deliverable agents and misleading if you treat it as economy-wide automation.
Each project ships four bundles:
- A written brief
- Input files the human used
- A human deliverable treated as gold standard
- Reported time and dollar cost from the freelancer
The economic summary on Scale's page — 28.9 mean human hours and $632.60 mean project value — is meant to signal that these are not ten-minute trivia tasks.
What "automation rate" actually counts
Scale defines automation rate as the percentage of projects where a reviewer answers yes to a client-style question: would a reasonable client accept the agent's work at least as much as the human reference? Reviewers use a three-point scale; scores of 2 (meets) or 3 (exceeds) count as success.
That is a holistic judgment, not a linter. Scale reports 94.4% inter-annotator agreement using a majority vote across three independent reviewers. The leaderboard notes no confidence interval for automation rate, so treat small gaps between nearby models as directional, not definitive.
Other RLI metrics — Elo, dollars earned, cost savings — appear on the methodology page. Automation rate is the headline number Anthropic quoted and Scale sorts in the public table.
Remote Labor Index leaderboard ranking Fable 5.2 first among frontier agents on freelance-style automation rate
The November 10, 2026 leaderboard
Scale's Performance Comparison block lists automation rates (not Elo) for frontier agents. The top rows we fetched on November 10:
| Rank | Agent label on Scale | 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% |
Anthropic's Fable 5.2 launch materials list 41.3% at High effort on Remote Labor Index versus 21.3% for Opus 5.5 and 20.8% for GPT-6 Astra — ordinary rounding of the same leaderboard family. The jump from Fable 5.1 at 17.92% to 41.30% is the substantive story: long-horizon agent scaffolding, not a one-point coding tweak.
For contrast, when RLI first publicized results, Scale's overview still says the best agent (Manus) sat near 2.5%. The bottom of today's table still shows agents under 2%. The overview paragraph is stale relative to the live ranking; trust the Performance Comparison table for current scores.
What developers get wrong about the 41% headline
Three confusions show up in launch-week threads (including Hacker News discussion of Fable pricing):
Substitution versus capability ceiling. Forty-one percent on RLI means success on this curated slice. It does not map to "forty-one percent of Upwork GMV." Many high-value gigs fail RLI filters by design.
Environment fairness. Scale tested multiple agent stacks — CLI coding environments and computer-use GUIs — and reports each agent's best run. Your harness, tool budget, and $30 generation cap (Scale's per-task budget on the methodology page) may not match theirs.
Multimodal failure modes. Scale's failure analysis clusters amateur-quality output (45.6% of fails), incomplete files (35.7%), corrupt formats (17.6%), and cross-file inconsistency (14.8%). Fable's lead does not erase those modes; it wins more often despite them.
How RLI relates to OpenAI's freelance study the same day
OpenAI published Work at the Frontier: AI's effect on freelance work on November 10, 2026, partnering with Upwork on marketplace posting trends versus ChatGPT activity shares. That study asks which categories gain or lose Upwork share while ChatGPT use rises. RLI asks whether an agent can ship a whole project to human quality.
The timing is coincidental but complementary. OpenAI's post highlights broader AI-related budgets and more task types bundled per posting; RLI shows technical ceilings when agents attempt those bundles end-to-end. Read both before you change agency pricing — we unpack the OpenAI paper in our companion post.
What pockets of work score well
Scale's published insights note successes concentrated in audio (sound design, separation), image generation (logos, ads), structured report writing, and data retrieval. Those match domains where a single deliverable file is inspectable and briefs are tightly scoped.
Domains with heavy 3D, architecture, or long video chains still appear in the set — Fable's marketing emphasized breakthroughs there — but they are also where failure modes around inconsistency bite. Treat category-level anecdotes as hypothesis generators, not guarantees your stack will pass.
Limits you should bake into planning
Scale lists three constraints worth citing in proposals:
- Manual evaluation cost — no cheap automated grader for CAD, video, and layered design packages.
- Incomplete economic coverage — client-facing and physical work excluded.
- Contamination risk — mitigated by keeping 230 projects private, with 10 public for qualitative replication only.
If a client asks "will agents replace our contractors," the honest answer from RLI is: on these 240 digital project shapes, the best frontier agent wins two in five; on the rest of the market, this benchmark is silent.
Practical checklist for builders
| Step | Action |
|---|---|
| 1 | Read the live RLI leaderboard before pitching automation rates to clients. |
| 2 | Match task shape to RLI successes — scoped deliverables with clear files — before defaulting to Fable 5.2. |
| 3 | Compare cost using Fable versus Opus routing; RLI does not price tokens. |
| 4 | Log your own acceptance tests; RLI's private set is not your client's industry. |
| 5 | Watch Scale's 10 public projects for qualitative failure patterns before scaling spend. |
Freelance-style agent benchmark concept: scored deliverables compared against human professional references
How this fits Anthropic's pacing story
Anthropic executives have argued for pacing, not pausing, on frontier releases while IPO filings circulate — see our Amodei pacing piece. RLI gives regulators and customers a single reproducible scoreboard for "can agents do paid remote work yet?" without conflating it with recursive self-improvement claims from SEC risk factors.
None of that changes API terms today. It does tell platform teams which model to benchmark first when your product looks like outsourced creative or analytics deliverables rather than ticket triage.
What people are asking
Is 41.3% "jagged AGI"? Mollick and others argue expert judgment still matters on partial automation — see our jagged AGI piece. RLI is consistent with jagged progress: strong averages on bundled freelance shapes, not uniform superhuman output.
Should I switch all agents to Fable 5.2? Only if your workload resembles RLI's mix and your budget tolerates Fable pricing. Routine coding loops may still favor Opus on cost per merged PR, per comparison tables.
Will Scale update scores daily? The site is a living leaderboard; cite the date you fetched. Anthropic minor releases could move rows without a launch event.
Can I hire solely from the leaderboard rank? No. Clients still need statements of work, NDAs, and domain experts because RLI excludes client-facing categories and does not score ongoing maintenance.
Related reading
- Fable 5.2 and Mythos launch pricing
- Top demos built with Fable 5.2
- Frontier models compared
- OpenAI and Upwork freelance study
- Agents Last Exam explainer
- Workforce agent adoption patterns
Automation rates, model names, and leaderboard ordering are accurate as of November 10, 2026, per Scale Labs' public RLI page and Anthropic's published benchmark table.