Workflows written by Claude Opus 5.5 raised four weaker AI models’ scores by 6.41 points on average, while workflows written by GPT-6 Astra lowered them by 2.53 points, according to WorkflowArena, a new benchmark from Einsia AI. That is an 8.94-point gap in who makes the better manager, by our arithmetic.
On the main leaderboard, Claude Opus 5.5 in Claude Code at max effort resolved 36.95% of tasks, the highest of 25 configurations. The best GPT configuration, GPT-6 Astra in Codex at xhigh effort, resolved 30.80%, a 6.15-point gap by our arithmetic. The paper was posted to arXiv on 9 November.
Business Pill · AI Agents Explained
An evergreen Business Engineer explainer on AI agents. It teaches the general idea only and does not discuss WorkflowArena.
The key insight: Workflows written by Claude Opus 5.5 raised four weaker executors by 6.41 points on average, while GPT-6 Astra’s static workflows cut them by 2.53, per WorkflowArena. The 8.94-point gap between the two is our arithmetic.
The benchmark has 145 long-horizon tasks across eight domains and 17 task forms: software (17 tasks), enterprise operations (18), science (16), formal mathematics (24), finance (16), law (22), content design (14) and 3D production (18), per the project page.
The model under test decides “whether and how to delegate”. Up to 16 concurrent workers share a four-hour deadline, and every delegation is recorded. Hidden verifiers grade the integrated deliverable: “Correct components without reconciliation cannot earn full acceptance”, the page says.
The main metric is Overall Resolved, the mean of the eight domain resolved rates, where a task must meet every acceptance criterion. A second metric, Overall Score, credits partial progress on a 0–100 scale.
WorkflowArena overall resolved rate, best configuration per model, as published by Einsia AI and Navers Lab (arXiv 2611.12856, 9 November 2026). A task counts only if every acceptance criterion passes. No Gemini or Fable model was tested.
What Was Posted on X
Agent harnesses now let a model spawn its own workers and write the workflow that coordinates them. But can it actually orchestrate them well?
Introducing Einsia AI’s WorkflowArena, the first benchmark for autonomous multi-agent workflow orchestration on real, long-horizon work.
– 145 tasks across 8 professional domains, from SWE and Lean proofs to law and 3D – 25 configurations of 9 frontier models in Claude Code & Codex – Every delegation recorded: we score the deliverable and see the workflow behind it
Einsia AI announced WorkflowArena at 14:30 UTC on 10 November: 145 tasks across 8 professional domains and 25 configurations of 9 frontier models in Claude Code and Codex, with every delegation recorded. The thread below it gives the leaderboard and transfer results this piece checks against the paper and GitHub results.
WorkflowArena makes the model the manager and grades how it hands out the work. really want to see who’s on top here, because the model that’s best at doing a task itself could easily be bad at handing it off. same as people
Zoaina quoted the announcement, writing that WorkflowArena makes the model the manager and grades how it hands out the work, and that the model best at a task itself could easily be bad at handing it off. It is a viewpoint post, not a result.
The Leaderboard
Nine models were tested in 25 configurations, Claude models in Claude Code and GPT models in Codex. At their best settings, Opus 5.5 resolved 36.95%, Sonnet 5.5 33.21%, GPT-6 Astra 30.80% and GPT-6.1 Sol 29.50%, per the GitHub results. Kimi K3.1 Pro reached 16.48% and GPT-6 Luna 13.32%.
On partial credit, Opus at high effort posts the best Overall Score, 80.21. Einsia’s thread says “No configuration wins every domain”: formal math is near saturation at 95.83% for Opus, while law tops out at 9.09%. No Gemini or Fable model is among the nine, so the results say nothing about Gemini 4 Carbon.
Solo Skill Versus Management Skill
To isolate orchestration, the authors held the executor fixed on 32 hard tasks and swapped only the model writing the workflow, which cannot touch the workspace. With Opus-written static workflows, Kimi K3.1 Pro rose from 35.75 to 45.04; with Astra-written ones it fell to 29.69.
Einsia’s thread puts it as “Being good at the task โ being good at orchestrating others”. It calls Astra “a much stronger solo system than these executors”. The two families also delegate differently: at high effort, Opus launches 0.86 workflows per task with 5.55 workers each, while Astra launches 2.09 with 1.72 workers.
As we read the transfer table, the benchmark separates two skills: doing the work and organising it. GPT-6 Astra resolves 30.80% on the main leaderboard, close to Opus’s range, yet its static workflows made all but one executor worse, per the paper’s Table 2.
Einsia AI thread on X, 10 November 2026
“Being good at the task โ being good at orchestrating others.”
Three Implications
FOR TEAMS BUILDING AGENTS The paper reports that the model writing the workflow changed a fixed executor’s score by up to 10.86 points, for Kimi K3.1 Pro with Opus dynamic workflows.
FOR MODEL BUYERS Einsia’s page puts Sol at low effort at 27.59% resolved for an estimated $3.61 per task, Opus at high effort at 34.11% for $6.66, and Astra at xhigh at 30.80% for $29.26.
FOR READING BENCHMARKS Opus at high effort scored 96.98 in software on partial credit but fully resolved 17.65% of those tasks, per the project page.
The framework reads: “The next competitive battle is no longer about building the smartest model.” It adds: “It is about controlling the operating layer through which autonomous work is delegated, orchestrated, and governed.”
As we read it, WorkflowArena is an attempt to measure that layer directly, by grading how a model delegates rather than only what it produces. The framework predates the benchmark and does not discuss it.
What Is Not Established
WorkflowArena is a new benchmark from one lab, and its results have not been independently reproduced. Each model ran in only one harness, Claude Code or Codex, so model and harness effects are not separated. The transfer test covers 32 tasks and four executors.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Claude Opus 5.5 leads WorkflowArena at 36.95% resolved, and its workflows lifted four weaker executors by 6.41 points on average where GPT-6 Astra’s cut them by 2.53, an 8.94-point gap by our arithmetic.
A note on sourcing. Scores, task counts and delegation figures are from Einsia AI and Navers Lab’s WorkflowArena project page, paper and GitHub results, and from the lab’s own X thread. Gaps between scores are our arithmetic. The benchmark has not been independently reproduced.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
