Token Billionaire Economics: Where 1B Output Tokens Per Day Actually Go

Ryan Lopopolo opens his AI Engineer Europe 2026 keynote [8] with a simple statement of identity: “I am a token billionaire” [1]. He means it literally. His team at OpenAI consumes roughly one billion output tokens per day running agents against a one-million-line codebase, a project he also summarized in an OpenAI engineering blog post [4]. When the host translated that into rough economics on the spot, the figure was “over $1,000” at output rates [1]. The Latent Space editors, applying current Codex rates and caching assumptions when they published the companion podcast episode, put it at $2-3k/day [3]. Lopopolo himself has not published an audited dollar figure, and as someone who works at OpenAI, he almost certainly does not pay public API rates. What he has published is where the tokens go.

The answer is the article’s central surprise: two-thirds of a billion daily output tokens have nothing to do with writing code.

Figure 1 - Hero diagram showing 1 billion output tokens split into three equal thirds: planning plus ticket curation, documentation, and implementation plus CI.

Figure 1 - Token Billionaire Allocation: One billion output tokens per day, divided equally among three buckets. Code generation is the smallest single bucket. The compounding advantage sits in the other two.

The Number That Requires a Host Interjection#

During the keynote fireside Q&A, the host noted that Lopopolo was “spending over a billion output tokens a day” and interjected: “That’s like over $1,000” [1]. That math holds precisely if the blended output price across all models used is $1.00 per million tokens [7]. No mainstream frontier model sits exactly at that price. The closest current model is GPT-5.4-nano at $1.25 per million output tokens [7]. GPT-4.1-mini lands at $1.60 [7]. Every other model the serious agent harness work runs on is four to thirty times higher.

At GPT-5.2, the model Lopopolo called his “magic moment” [1] and the one his team built the project on, the Codex-specialized derivative priced at $14.00 per million output tokens [7], one billion output tokens per day equals $14,000 per day, not $1,000. The Latent Space editors’ $2-3k/day estimate [3] implies either a blended model stack that shifts significant volume to cheaper models, some share of the output count being batch-API production (50% discount) [7], or both. None of these scenarios are confirmed by Lopopolo. The article’s value is not in the exact invoice. It is in what the scale implies about where the work actually lives.

Figure 2 - Bar chart of daily cost for 1B output tokens by model: from $1,250 (GPT-5.4-nano) to $30,000 (GPT-5.5), with GPT-5.2-Codex at $14,000 highlighted as the magic moment model.

Figure 2 - Cost of 1B output tokens per day by model: The host’s “$1,000/day” framing is an order-of-magnitude estimate that assumes $1 per million output tokens, a price point no frontier model occupies exactly. At the “magic moment” model GPT-5.2-Codex ($14/1M), the daily total is $14,000. At Claude Opus 4.8 or GPT-5.5, it exceeds $25,000. All figures are from vendor-direct pricing pages [6][7].

A Note on Caching (the Nuance That Changes the Analysis)#

Anyone who has worked with prompt-caching APIs will see a natural objection: heavy cache reuse dramatically reduces the per-call cost, so maybe caching is what makes the “$1K” framing plausible.

That reasoning runs into a hard structural fact. Prompt-caching discounts reduce input token costs, not output token costs [6][7]. The cache hit saves money on the tokens the model reads at the start of each trajectory. The tokens the model writes (the code, the documentation, the PR descriptions, the CI analysis) are charged at full output rates regardless of how many times the same context was loaded before.

For a workload that is fundamentally denominated in output tokens (an agent is doing work by generating output, not by re-reading the same prompt), caching compresses one side of the ledger while the other side remains unchanged. At Anthropic’s Claude Sonnet 4.6, for example, 1 billion output tokens costs $15,000 per day whether you use caching or not [6]. The cache savings on the input side are real, but they are modest relative to the output line.

KEY INSIGHT: Caching saves money on input tokens. For any workload measured in output volume, caching does not change the output line. Engineers optimizing for cost at agent scale need to address the output side directly, through model selection and batch-API routing, not through caching strategy.

Figure 3 - Split diagram showing caching discount applied to input tokens on the left side (reduced) and output tokens on the right side (unchanged at full rate).

Figure 3 - Caching saves input, not output: Prompt-cache discounts (90% off at Anthropic, 75-90% off at OpenAI depending on model) apply exclusively to cached input reads [6][7]. Output tokens are billed at full rate. A billion-token-per-day workload dominated by output is not materially affected by caching strategy.

Where the Billion Tokens Actually Go#

Asked directly where his billion daily output tokens go, Lopopolo gave a precise answer: “I would say probably it’s probably a third, a third, a third between like planning ticket curation, documentation, implementation, and stuff that runs in CI” [1]. The three buckets:

One-third on planning and ticket curation. Producing tickets that an agent can pick up and execute end-to-end without requiring further human direction. The goal is a ticket so well-specified that no human needs to be in the loop during execution.
One-third on documentation. Architecture decision records, persona-oriented “what good looks like” documents, prompting cookbooks, agent skills.
One-third on implementation and CI. Actual code generation plus the reviewer agents and structural tests that run on every push.

The two-thirds that is not implementation is what most teams do not budget for. The common mental model is that agent tokens go toward writing code, and everything else is overhead. Lopopolo’s breakdown inverts that. The planning and documentation spend is not overhead. It is the compounding asset. Every well-specified ticket reduces the agent interaction cost on that task permanently. Every documentation artifact improves the next trajectory that reads it. The implementation third produces the code that ships; the other two-thirds produce the context that makes the code good.

Figure 4 - Pie chart with three equal 33% segments: Planning and Ticket Curation, Documentation, and Implementation plus CI. An amber bracket groups the two non-implementation thirds as 66% of spend.

Figure 4 - The 1/3 + 1/3 + 1/3 breakdown: Lopopolo’s direct answer at AI Engineer Europe 2026 (transcript 00:39:53-00:40:27) [1]. Two-thirds of the daily token budget goes to planning, ticket curation, and documentation. Code generation is the smallest single bucket. This is the economic argument for treating harness quality as primary work.

KEY INSIGHT: Two-thirds of a token-billionaire’s daily spend is non-implementation. If your team is treating planning and documentation tokens as waste to be minimized, you are optimizing the wrong bucket. The planning spend is what makes the implementation spend produce something mergeable.

The Three Things That Are Actually Scarce#

Early in the keynote, before any discussion of costs, Lopopolo names what is actually constrained in an agent-first engineering workflow. Three things [1]:

Human time. The number of hours an engineer can direct, evaluate, and course-correct agent work in a day.
Human and model attention. The cognitive load available for systems thinking versus line-by-line oversight.
Model context window. The amount of information a model can hold in working memory at any one moment.

Tokens are not on the list. At a billion output tokens per day, they are the abundant resource. The constraint has moved upward in the stack.

This inversion is the central economic argument of the whole keynote. Teams that optimize for token cost as their primary constraint are solving the wrong problem. The token spend is elastic: you can always get more tokens by waiting a few hours and paying the bill. Human time is not elastic. Model attention is not elastic. Context window is not elastic. Engineering decisions that conserve the scarce resources are worth far more than decisions that reduce token spend by 15%.

Figure 5 - Diagram with three scarce resources (Human Time, Human and Model Attention, Model Context Window) above a dividing line, and Tokens labeled as the abundant resource below.

Figure 5 - Scarce versus abundant at agent scale: Lopopolo’s enumeration at transcript 00:05:11-00:05:52 [1]. Human time, human and model attention, and model context window are the constrained resources. Token spend is abundant. Engineering decisions should optimize against the scarce side.

The Magic Moment That Made This Scale Viable#

Lopopolo is specific about when the economics became tractable: “For me, the magic moment was GPT 5.2, which when it came out was able to do the full job of a software engineer” [1]. GPT-5.2 was released in December 2025. The Codex-specialized derivative, GPT-5.2-Codex, launched for API access in January 2026.

Before that model, the capability gap between “the agent can write code” and “the agent can do the full job end-to-end” meant that the human was always a synchronous gate on execution. After it, the gate became optional in principle and the harness work became the primary engineering investment. The billion tokens per day, the three-engineer team, the three-to-five PRs per engineer per day, all of these followed from having a model capable of the full job.

The specific model matters for the cost discussion because GPT-5.2-Codex, the vehicle for that capability jump, is priced at $14.00 per million output tokens [7]. The jump in capability came with a price point that makes the “$1K/day” framing obviously an underestimate. The current successor, GPT-5.3-Codex, carries the same $14.00 per million output rate [7]. Lopopolo’s follow-up talk covers the post-GPT-5.2 throughput trajectory in more detail [2].

Figure 6 - Timeline from December 2025 to April 2026 showing the GPT-5.2 release, GPT-5.2-Codex API launch, and AI Engineer Europe 2026 keynote as sequential milestones.

Figure 6 - The capability inflection point: GPT-5.2 (released December 2025) was Lopopolo’s named “magic moment” model [1], the first model his team found capable of the full software-engineering job end-to-end. The Codex-specialized derivative launched for API access in January 2026 at $14.00 per million output tokens [7]. The subsequent three-month project that produced 1M+ LOC and the April 2026 keynote data built on this model.

50 Agents Running 24/7: The Dream State the Economics Enable#

If the token spend is not the constraint, what is the aspirational end state? Lopopolo names it directly: “The dream here is that I actually have 50 agents running 24/7 and I don’t have to interact with them at all. The way to do that is to define the work well, figure out ways for it to automatically be scheduled, and remove myself from having to click the button” [1].

The economics make this tractable. One billion output tokens per day distributed across 50 agents is 20 million output tokens per agent per day. At GPT-4.1-mini rates ($1.60 per million output) [7], that is $32 per agent per day. At GPT-5.2-Codex rates, it is $280 per agent per day. Both figures are within the operating budget of a small engineering team if even a fraction of those agents are producing merged PRs.

What stands between most teams and this dream state is not cost. It is harness quality. Specifically, it is the harness’s ability to define completion, verify output, and route agent work without requiring a human to type “continue.”

KEY INSIGHT: “Every time I have to type ‘continue’ to the agent is like a failure of the harness to provide enough context around what it means to continue to completion” [1]. Lopopolo’s framing makes type-continue a diagnostic, not an inconvenience. Each occurrence identifies a specific gap in the harness: a missing acceptance criterion, an underspecified skill, an unclear test. The gap should be closed permanently, not worked around repeatedly.

Figure 7 - Architecture diagram: 50 agent nodes routed through Symphony orchestrator, with the human on a separate async harness-improvement path, not in the synchronous execution loop.

Figure 7 - The 50-agents-running-247 dream state: Lopopolo’s aspiration at transcript 00:33:17 [1]. Fifty agents running continuously, with the human making asynchronous harness improvements rather than synchronous execution decisions. Symphony [5] is the OpenAI orchestrator that routes agent work. The type-continue failure diagnostic (transcript 00:33:44) is the operational instrument that drives harness improvements.

The Type-Continue Diagnostic in Practice#

The relationship between the token economics and the type-continue principle is tighter than it appears. Consider what a type-continue event costs at agent scale. An engineer stops whatever they are doing, re-reads the agent’s current state, decides what “continue” means in this context, and types it. That intervention takes human time and human attention, both items on the scarce-resources list. The token cost of the continuation is trivial. The opportunity cost of the human’s involvement is not.

In our companion article on Garbage Collection Day, we covered the weekly ritual where recurring type-continue events become the raw material for harness improvements: the bespoke lint, the structural test, the reviewer-agent update. The economics now supply the argument for why that investment is worth making. The token spend on a type-continue continuation is measured in cents. The human attention cost is measured in minutes of context-switching. Over 50 agents and 20 working days, the arithmetic is not close.

The pattern also connects back to the scarce-resources list. Lopopolo explicitly designed his harness so that “the human’s job is to define the work, prioritize it, schedule it, and remove yourself from the loop” [1]. Each type-continue is a diagnostic signal that the removal is incomplete. The harness has not yet encoded enough context to complete the task autonomously. The fix is in the context, not in the human’s willingness to click.

Figure 8 - Flowchart splitting on a type-continue event: left quick path loops back to the same event; right durable path converts it to a harness improvement via GCD so the failure never recurs.

Figure 8 - Two responses to a type-continue event: The quick path resumes work immediately and guarantees the same event will recur. The durable path converts the occurrence into a harness improvement, removing that class of interruption permanently. At agent scale, the compounding cost of the quick path dwarfs the one-time cost of the durable fix.

The Compounding Asset, Not the Recurring Cost#

There is a reframing implicit in all of this that is worth making explicit. In most engineering teams, token spend is treated as a cost line: minimize it, budget for it, report on it. Lopopolo’s framing treats it as evidence of productive work. A team spending a billion output tokens per day is a team that has solved the bottleneck. The tokens represent agent trajectories completing. The spend is not the problem. The spend is the output.

The metric that should concern teams is not the token bill. It is the ratio of agent-to-human in the execution path. A team spending $1,000 per day with humans in the synchronous gate on every PR is less effective than a team spending $14,000 per day with humans building harness improvements asynchronously. The first team’s throughput is bounded by human review capacity. The second team’s throughput is bounded by harness quality, which is improvable.

This is the economic argument for treating harness engineering as first-class work rather than infrastructure overhead [9]. The two-thirds of Lopopolo’s token budget that goes to planning and documentation is not overhead spending. It is the investment that makes the implementation third produce something the harness can verify, merge, and build on.

The number to watch is whether the harness quality is improving each week, because that is what determines whether the cost eventually buys more.

Figure 9 - Two-column comparison: Cost Mindset (token spend as a downward arrow) versus Asset Mindset (token spend reinvested into planning and documentation in a compounding loop).

Figure 9 - Cost mindset versus asset mindset on token spend: The cost mindset minimizes token spend. The asset mindset treats two-thirds of that spend as an investment in context quality that compounds across future agent trajectories. Planning artifacts and documentation produced today reduce the cost per successful trajectory tomorrow.

Conclusion#

The “token billionaire” framing lands as a punchline but its mechanics are the argument. One billion output tokens per day costs somewhere between $1,250 and $30,000 depending on the model mix, whether batch-API routing is in play, and whether Lopopolo’s team pays public API rates at all. The specific invoice is unknowable. What is not unknowable is the 1/3 + 1/3 + 1/3 breakdown: planning and ticket curation, documentation, implementation and CI, each taking an equal share [1].

Two-thirds of a billion-token-per-day operation is making the code-writing work. It is producing the context, the specifications, the review standards, and the architectural decisions that allow the implementation third to produce something mergeable without a human in the synchronous path.

The three genuinely scarce resources at this scale, human time, human and model attention, and model context window [1], are all conserved by this allocation. The team’s human hours go toward harness improvement and work definition, not line-by-line review. The models’ context windows are kept small by structural lints and canonical implementations. The attention budget goes toward systems thinking rather than individual PR oversight.

The dream of 50 agents running 24/7 without a human typing continue [1] is not a cost problem. At the model prices that made GPT-5.2 the “magic moment” [1], it is a harness quality problem. And harness quality improves one Garbage Collection Day at a time.

The Series#

This is Part 2 of the 3-part OpenAI Lopopolo sub-series:

Garbage Collection Day: A Weekly Ritual for Reducing Agent Slop. The Friday slop-elimination ritual, the three durable artifact types, and the throughput flywheel.
Token Billionaire Economics: Where 1B Output Tokens Per Day Actually Go (this article). The cost structure and token allocation of operating at extreme agent throughput.
Code as Disposable Build Artifact: LLMs as Fuzzy Compilers. Treating generated code as ephemeral output and the architectural implications for harness design.

References#

[1] R. Lopopolo, “Harness Engineering: How to Build Software When Humans Steer, Agents Execute,” AI Engineer Europe 2026 keynote, April 8-10, 2026, London. https://www.youtube.com/watch?v=am_oeAoUhew

[2] R. Lopopolo, “Extreme Harness Engineering: 1M LOC, 1B toks/day, 0% human code or review,” AI Engineer follow-up talk. https://www.youtube.com/watch?v=CeOXx-XTYek

[3] R. Lopopolo, “Extreme Harness Engineering for Token Billionaires: 1M LOC, 1B toks/day, 0% human code, 0% human review,” Latent Space podcast, April 7, 2026. https://www.latent.space/p/harness-eng

[4] R. Lopopolo, “Harness engineering: leveraging Codex in an agent-first world,” OpenAI, February 11, 2026. https://openai.com/index/harness-engineering/

[5] R. Lopopolo and A. Kotliarskyi, “An open-source spec for Codex orchestration: Symphony,” OpenAI, 2026. https://openai.com/index/open-source-codex-orchestration-symphony/ (GitHub: https://github.com/openai/symphony)

[6] Anthropic, “Pricing,” Claude Platform Docs, observed 2026-06-08. https://platform.claude.com/docs/en/docs/about-claude/pricing

[7] OpenAI, “Pricing,” OpenAI Developers API Docs, observed 2026-06-08. https://developers.openai.com/api/docs/pricing

[8] AI Engineer Europe 2026, Queen Elizabeth II Centre, London, April 8-10, 2026. https://www.ai.engineer/europe/2026

[9] G. Dotzlaw, K. Dotzlaw, and R. Dotzlaw, “What Is an Agent Harness, Really? Nine Components Most Builders Miss,” dotzlaw.com, 2026. /insights/claude-code-01-agent-harness/