FlowerBench Ranks Enterprise AI Agents on Real Workflows; OpenAI's Codex Leads
Key Takeaways
- ▸FlowerBench provides the first comprehensive ranking of AI agents evaluated on authentic enterprise workflows across seven business domains in real organizational environments
- ▸OpenAI's Codex with GPT-5 achieves the highest performance score (5.50), followed by two Anthropic Claude Opus variants in the top 3, establishing the current leadership in enterprise agent capability
- ▸Cost and efficiency metrics vary dramatically—top-ranked models are not necessarily the most efficient, with costs ranging from $1.19 to $28.58 per task and execution times from 38 to 170+ minutes
Summary
A comprehensive new benchmark called FlowerBench has been released to evaluate AI agents' performance on real, long-horizon enterprise workflows. The benchmark tests agents across seven critical business domains—Finance, Healthcare, Insurance, Operations, MLOps, Legal, and Marketing—with tasks running in private organizational environments that include proprietary data, context, and tools. OpenAI's Codex with GPT-5 ranks first overall with a score of 5.50, executing benchmark tasks in 87 minutes with 14 million tokens at a cost of $17.12. Anthropic's Claude Opus 4.8 follows in second place (4.80 score) and Claude Opus 4.7 in third (4.70 score), with significant differences in execution time, token efficiency, and cost.
Beyond raw performance rankings, FlowerBench provides crucial operational metrics that reveal the trade-offs between different solutions. Execution times range from 38 to 170 minutes, token usage spans from 1.6 million to 40 million per task, and cost-per-task completion varies dramatically from $1.19 to $28.58—indicating that the highest-scoring model is not necessarily the most cost-efficient. The benchmark's focus on authentic enterprise workflows, rather than synthetic tasks, offers organizations genuine insight into how various AI agents would perform in production settings.
- The benchmark's use of private organizational environments with proprietary tools and data provides realistic performance signals unavailable from synthetic benchmarks
Editorial Opinion
FlowerBench represents a critical advancement in evaluating AI agents by testing them on genuine enterprise work rather than theoretical tasks. Running agents inside real organizational contexts with proprietary systems and data reveals production-ready performance in ways standard benchmarks cannot. However, the wide variance in cost and efficiency suggests organizations should adopt a portfolio approach—matching specific models to use cases rather than defaulting to the highest-ranked solution, as cost-per-task and execution speed often diverge significantly from raw performance scores.
