This is a report by Florian Brand, a PhD student at Trier University working on language-model evaluations, and Greg Burnham, head of benchmarks at Epoch AI, published on Epoch AI's site on 13 February 2026. It reviews three benchmarks that aim to measure whether AI systems can perform real-world, digital, non-coding work of economic value: the Remote Labor Index (RLI), APEX-Agents and GDPval. Its summary position is that progress on them will correlate with real utility, but that because the tasks are well-defined and self-contained, high scores "would not imply end-to-end automation of digital professions" so much as a shift toward delegating work to AI, as coding agents have done in software engineering.
Summary
The three benchmarks compared
The authors tabulate the benchmarks on nine facets. RLI, from CAIS and Scale AI, has 240 tasks (10 public) drawn from real freelance projects with visual outputs such as web design, product art and video editing, an average reported human time of 29 hours, web access during evaluation, and expert head-to-head judgment of AI against human output. APEX-Agents, from Mercor, has 480 public tasks in investment banking, management consulting and corporate law, devised by experts and "optimized to be hard for models," with much more information than necessary in the environment, an average human time of 2 hours, no web access, and rubric grading by an AI judge. GDPval, from OpenAI, has 1,320 tasks (220 public) from 44 occupations, mostly research and writing, devised by experts, an average human time of 7 hours, web access, and expert head-to-head judgment. Top scores as of February 2026 were 4%, 30% and 74% respectively.
The per-benchmark takeaways are: RLI's early evaluations "likely under-elicited models" but this has improved and top scores remain under 5%; APEX-Agents' environments are messier than the other two, which "probably helps with realism"; and GDPval's realism is "doubly limited" by self-contained tasks and the absence of an environment, while it is the closest to saturation.
Task sourcing and realism
RLI tasks are real Upwork projects, selected so that output can be judged visually or aurally, packaged as a prompt, input files and the freelancer's deliverable, with freelancers paid an average of $41 to provide a work sample. APEX-Agents tasks were created from scratch by professionals, reviewed in multiple rounds including "LLM-based refinements to make the tasks harder," and the authors say it is unclear how far this made the distribution unrepresentative of real work. GDPval started from the nine industries contributing most to US GDP, identified 44 occupations with substantial digital work using O*NET data and an LLM classifier, and had expert teams (five contributors per task on average) write 30 tasks per occupation. How APEX-Agents and GDPval task creators were paid is not known to the authors.
All three are one-shot: a single static prompt, then a single set of outputs. GDPval and RLI supply only the necessary information, which the authors call cleaner than a typical office context, though more realistic for freelancing; APEX-Agents provides an intentionally messy file system with duplicates and unrelated half-finished projects, whose realism the authors cannot judge. None situates tasks within a broader business context, and data is cleaned of proprietary and personal information.
Time, web access and evaluation
Human time estimates come from experts completing the tasks for GDPval and a subsample for APEX-Agents, and from freelancers' self-reports for RLI, which correlate with the wages paid. Among RLI's public tasks, the earbuds animation task was reported at 60 hours and the Singing in the Rain arrangement at about 30. The authors suggest economic value may be more similar across benchmarks than time horizons: RLI projects were billed at an average of $600, which two hours of an experienced corporate lawyer's time could exceed.
RLI and GDPval allow web browsing; RLI keeps a blocklist for outputs available online. Over half of GDPval's tasks depend on live web access, two public samples already had inaccessible links, and about 5% of tasks assume a now-past point of view that models may find easier in hindsight.
GDPval has experts compare the human reference and AI output with a tie option and reports win and tie rates separately; the authors treat win-plus-tie as most informative and most comparable to RLI's "at least as good as" judgment. GDPval used at least two judges and RLI three in the original papers. The authors note they do not know how judges are induced to adopt an economically motivated customer's perspective. APEX-Agents' rubric-and-LLM-judge approach is described as reasonable given the objective nature of its rubrics.
Progress and elicitation
As of writing, GPT-5.2 Pro held the GDPval top score at 74% win-plus-tie (60% win, 14% tie), with GPT-5.2 (xhigh) at 71%; Artificial Analysis's third-party GDPval-AA implementation placed Claude Opus 4.6 well ahead of GPT-5.2 (xhigh), suggesting it might exceed 74% on an official run. The RLI top score was 4% (Opus 4.6) and the APEX-Agents top score 30% (Opus 4.6).
The authors argue all three may under-elicit capability through their scaffolds. RLI's Ubuntu-based computer-use scaffold cannot run Photoshop or Fusion360, forcing fallbacks to GIMP or FreeCAD, and its CLI scaffold routes video generation through Veo 3 Preview, whose weaknesses the evaluated model inherits; the original runs had a one-hour limit and little pre-installed software, which CAIS and Scale AI told the authors they had since upgraded before the Opus 4.5 evaluation. APEX-Agents uses a ReAct loop with 79 tools, many of them wrappers around python-docx and python-pptx, which the authors suspect models would use more capably directly and which add management overhead; they suspect a coding scaffold such as Claude Code would do better. GDPval evaluated OpenAI models with web search and a code interpreter and Claude models through the claude.ai web interface, with no details for Gemini or Grok.
Conclusion
The authors compare the benchmarks to SWE-bench Verified: useful instruments that capture a narrow slice of the relevant jobs, not "remote labor" or GDP as a whole. Saturation would change how work is done, with execution accelerating and planning, delegation and verification taking a larger share of human time, and further progress toward measuring replacement would require benchmarks of end-to-end tasks in the messy context of real work.
Key claims
- Facet table: RLI 240 tasks / 29 hours / web access / expert judgment / top 4%; APEX-Agents 480 / 2 hours / no web / AI-rubric / 30%; GDPval 1,320 / 7 hours / web / expert judgment / 74%.
- GDPval top score 74% win-plus-tie (GPT-5.2 Pro); GDPval described as closest to saturation.
- Position: high scores would indicate a shift toward delegation, not end-to-end automation.
- Scaffolds likely under-elicit all three; task-creator pay unknown for APEX-Agents and GDPval.
Provenance
Retrieved 22 September 2026 from epoch.ai. The page carries a 21-minute audio version and a citation block dated 2026.
Relationships
- supports: Remote Labor Index, APEX-Agents, GDPval — third-party comparison of the three benchmarks' design and limits.
- related: Remote Labor Index: Measuring AI Automation of Remote Work (Mazeika et al., 2025), APEX-Agents (Vidgen et al., Mercor, 2026), GDPval Paper (OpenAI, Oct 2025) — the primary texts reviewed.
- related: SWE-Lancer: Can Frontier LLMs Earn $1 Million from Real-World Freelance Software Engineering? (Miserendino et al., 2025) — a market-priced software benchmark the review's SWE-bench comparison bears on.
- related: Epoch AI, AI Benchmarks and Evaluation, Agentic AI