Anthropic's Fable 5.2 Completes 41 Percent of Real Freelance Jobs on Industry Benchmark

Anthropic’s Fable 5.2 model can now complete 41.3 percent of real freelance assignments at a professional standard, according to benchmark results reported by The Decoder. That marks a massive leap from just a year ago, when top-performing AI agents completed only 2.5 percent of real-world freelance projects. The benchmark evaluates models against actual Upwork contracts requiring multi-step workflows across design, audio, and 3D modeling software.

What happened

The Decoder reports that Scale Labs and the Center for AI Safety (CAIS) published the latest results for their Remote Labor Index (RLI), which tests AI models on 240 real Upwork projects completed by 358 verified professionals. These jobs—worth a total of about $144,000, averaging $632.60 per project and requiring roughly 29 hours of human labor—span tasks like CAD design, video editing, game development, and data analysis. Agents operate inside a virtual environment preloaded with professional software like Blender, GIMP, and Audacity alongside specialized AI generation tools.

Anthropic’s Fable 5.2 achieved a 41.3 percent completion rate, leaving competing models far behind. According to The Decoder, Anthropic’s Opus 5.5 took second place with 21.25 percent, closely followed by OpenAI’s GPT 6 Astra at 20.83 percent and Google’s Gemini Argon 4 at 17.08 percent. Older models like GPT-5.2 and Gemini 3 Pro completed less than three percent of assignments. Scale Labs noted on X that Fable 5.2 made its strongest gains in audio and music editing, architectural and interior design, and 3D modeling, successfully completing nearly 50 project types that no previous AI agent had passed.

However, the benchmark highlighted significant limits. Scale Labs noted that Fable 5.2 still struggles with animations and industrial product design. When AI agents fail in the test, CAIS researchers found they deliver amateur-quality work, incomplete results, broken files, or contradictory designs. Furthermore, attempts to automate the evaluation process using AI reviewers failed because evaluator models were easily fooled by superficial visual polish despite underlying technical flaws.

Why it matters for small business operations

If you manage operations or hire creative freelancers for a small company, this score jump is a signal that AI agents are transitioning from simple text completion to handling functional, multi-step asset creation. A 41 percent success rate on complex tasks averaging 29 human hours means agents can realistically assist with preliminary draft work, basic 3D modeling, or routine audio cleanup.

However, you shouldn't fire your freelance contractors just yet. The 58.7 percent failure rate proves that agents still regularly generate unusable files, incomplete deliverables, and amateur designs. Furthermore, because CAIS demonstrated that automated AI graders fail to catch technical flaws hidden behind pretty surfaces, a skilled human expert still needs to inspect every single deliverable inside the actual software before anything goes to production. For small teams, the practical strategy is leveraging agents for rough drafting and modular task prep while keeping experienced professionals responsible for review and assembly.

What to do this week

1. Audit your company's creative and technical outsourcing backlogs to flag recurring tasks under 30 hours that involve 3D modeling, audio editing, or basic spatial design.
2. Establish a mandatory manual review protocol requiring human technical leads to inspect all AI-assisted deliverables inside native software before signing off.
3. Test an agent tool powered by high-performing models like Fable 5.2 on a archived freelance assignment to benchmark its preliminary output against your original freelancer's work.

Sources

• Anthropic's Fable 5.2 completes 41 percent of real freelance jobs, up from 2.5 percent a year ago — The Decoder