the-decoder.com web signal

OpenAI Funds 100-Agent Coding Swarm at $1.3M a Month

openai agents coding tools ai-agents coding-tools cost

Key insights

  • A three-person team at OpenAI ran 100 simultaneous AI agents consuming 603 billion tokens and costing $1.3 million in one month.
  • Disabling Fast Mode alone would reduce monthly costs by 70%, making latency preference the dominant cost driver over task volume.
  • All agent outputs are open source and model-agnostic, making the workflow replicable on models other than GPT-5.5.

Why this matters

This is the first detailed public benchmark of what unconstrained agentic engineering actually costs at scale, and the $1.3M/month figure will anchor budget conversations at every company planning agent rollouts in 2025 and 2026. The Fast Mode finding flips the usual assumption: in agentic systems, latency configuration outweighs task volume as a cost multiplier, which has direct implications for how teams should architect workflows before committing to production deployments. The open-source, model-agnostic outputs mean competitors and independent teams can study and replicate the approach without relying on OpenAI's pricing or infrastructure, accelerating the spread of high-volume agentic engineering practices across the industry.

Summary

Peter Steinberger, founder of OpenClaw and now at OpenAI, is running roughly 100 simultaneous AI agents handling code review, PR analysis, bug detection, and security scanning, all on OpenAI's tab as a deliberate research exercise in unconstrained token spend. The three-person team consumed 603 billion tokens and 7.6 million requests in a single month, reaching $1.3 million in API costs. GPT-5.5 is the primary model, and disabling Fast Mode alone would cut total costs by 70%, meaning latency preferences are the dominant cost variable, not task volume. Essentially: (Steinberger, OpenAI) are using a tiny team as a live laboratory for what agent-driven engineering looks like when budget isn't the binding constraint. - 603B tokens and 7.6M requests from three engineers in one month - Fast Mode accounts for roughly 70% of spend, making it the single largest cost lever - All outputs are open source and model-agnostic, replicable outside OpenAI's stack The experiment isn't about whether AI agents can do this work; it's about establishing what a fully agentic engineering team actually costs at full throttle.

Potential risks and opportunities

Risks

  • If OpenAI redefines the project scope or ends the research subsidy, the three-person team loses the economic foundation that makes 100 simultaneous agents viable, with no clear path to self-funding at $1.3M/month
  • Competitors (Anthropic, Google DeepMind) can use the open-source outputs to benchmark their own models against GPT-5.5 on agentic coding tasks before OpenAI has a commercialized product built on the findings
  • Enterprise buyers exposed to the $1.3M/month figure will treat it as a ceiling for what agentic engineering tooling should cost, creating pricing pressure on any vendor attempting to productize similar agent fleets without a research subsidy behind them

Opportunities

  • Observability and cost-management vendors (Langfuse, Helicone, Arize AI) gain a high-profile case study to pitch agentic cost optimization tools to engineering teams planning comparable deployments
  • Teams at AI-forward companies (Cursor, GitHub, JetBrains) can adapt the open-source, model-agnostic workflows to run similar agent fleets on cheaper or self-hosted models, capturing the productivity gain at a fraction of the cost
  • Cloud providers (AWS, Google Cloud, Azure) can use the 603B-token-per-month data point to design reserved-capacity or committed-use pricing tiers specifically targeting high-volume agentic engineering workloads

What we don't know yet

  • Whether OpenAI plans to publish formal research findings from this project, and on what timeline, given it is framed as active research
  • Which specific bug classes or security vulnerabilities the agents have surfaced that human reviewers missed, and at what detection rate
  • How per-agent costs break down across task types (code review vs. security analysis vs. PR review) and whether some are dramatically cheaper than others

Originally reported by the-decoder.com

Read the original article →

Original headline: OpenClaw Founder Peter Steinberger Runs 100 AI Coding Agents Simultaneously at $1.3M/Month — OpenAI Footing the Bill as Research Into Unconstrained Token Spend