Fable 5 hits 16% on AI freelance benchmark, doubles Opus 4.8

Key Takeaways

• Fable 5 achieved 16.1% automation rate on freelance tasks, more than double Opus 4.8's 8.3%
• AI agent capabilities have quadrupled in under eight months according to CAIS
• Despite the record, 16% success rate shows AI cannot yet replace human freelancers

Anthropic's Fable 5 model scored 16.1% on the Center for AI Safety's Remote Labor Index, doubling the performance of Opus 4.8 and setting a new record for AI systems completing real freelance work. The benchmark measures whether AI agents can finish economically valuable projects at a quality paying clients would accept. Fable 5 hit this mark more than twice as often as its nearest competitor.

The US government re-authorized Fable 5 on June 30 after a brief suspension. Anthropic says the model shares capability similarities with Mythos 5, which remains restricted to select organizations. But before the pullback, CAIS ran Fable 5 through its Remote Labor Index, a benchmark released in October 2025 that tests AI on actual freelance project types.

What the Remote Labor Index actually tests

RLI puts AI models through tasks that real freelancers get paid for: 3D mockups of engagement rings, video advertisements, floor plan mapping. Researchers give each model the same input files a human contractor would receive, like reference documents and project briefs. Human evaluators then compare the AI output against professional-quality deliverables.

The automation rate reflects how often evaluators judged AI work as good as or better than human professional output. At 16.1%, Fable 5 cleared this bar roughly one in six attempts. Opus 4.8 managed 8.3%, and OpenAI's GPT-5.5 came in at 6.3%.

The acceleration is striking. When RLI launched, the best models topped out at 2.5%. The previous published leader, Opus 4.6 running with Claude Cowork scaffolding, reached 4.17%. Fable 5 nearly quadrupled that in less than a year.

Why 16% doesn't mean freelancers should panic

The improvement rate matters more than the absolute number for tracking AI progress. But for anyone worried about near-term job displacement, 16% is still a long way from 100%. Five out of six freelance projects still required human execution to meet client standards.

CAIS noted that the government shutdown of Fable 5 in mid-June cut testing short. Even assuming the model failed every remaining project, its automation rate would still be 14.6%, higher than any other system tested.

The more practical constraint is integration. Security concerns, compliance requirements, and organizational inertia slow AI adoption for most companies. Replacing human freelancers would require networks of agents checking work quality, budget, and timelines. The tradeoff is not one-to-one.

The LLM judge experiment failed

CAIS tried removing humans from the evaluation loop entirely by using an LLM judge instead. It did not work. Evaluating a deliverable requires opening project files in professional applications, operating those applications competently, and forming judgments the way a client would. These computer-use skills remain AI's weakest area.

“Evaluating an RLI deliverable is itself a demanding, agentic task. Doing it properly means opening the project's files in the right professional applications, operating those applications competently, and forming a judgment the way a client would.”

— CAIS study

This creates an interesting ceiling. Until AI can reliably judge its own work, human oversight remains essential. The models might generate acceptable output more often, but someone still needs to verify it.

What this means for teams building AI agents

The quadrupling of agent capabilities in eight months suggests that teams building autonomous workflows should plan for rapid capability changes. A task that fails 94% of the time today might hit acceptable success rates within two or three model generations.

The RLI results also highlight where current agents struggle most: computer use, file manipulation, and multi-application workflows. Teams using tools like Zapier, Make, or n8n for automation should watch for agent integrations that improve these capabilities specifically.

Frequently Asked Questions

What is Anthropic's Fable 5 model?

Fable 5 is an AI model from Anthropic that shares capabilities with Mythos 5. It was briefly suspended by the US government and re-authorized on June 30, 2026.

What does the Remote Labor Index measure?

The RLI benchmark by CAIS measures how often AI agents can complete real freelance projects at a quality that paying clients would accept, including tasks like 3D design, video editing, and data analysis.

Can AI replace human freelancers now?

Not yet. A 16% success rate means AI fails to meet professional standards on five out of six projects. Security concerns and integration challenges further slow adoption.

How fast are AI agent capabilities improving?

According to CAIS, AI agent capabilities on freelance tasks have quadrupled in under eight months, from 4.17% to 16.1% automation rate.

Why did the LLM judge fail in CAIS testing?

Evaluating freelance deliverables requires operating professional applications and forming client-like judgments, skills that current AI agents still lack.

Source: Latest news