OpenClaw creator and OpenAI technical staff member Peter Steinberger estimated that he spent about $10 million in AI tokens during the previous month, according to The Register’s October 21, 2026 report. He made the comment during a conversation with Cloudflare CTO Dane Knecht at Cloudflare Connect 2026 in San Francisco.
Steinberger also acknowledged that the spending did not translate entirely into useful work.
“What people don't always understand is not all of that is useful,” he said, according to the report.
That admission is the consequential part of the story. A large token budget tells us something about the scale of experimentation. Without a breakdown of completed tasks, failed attempts and reusable results, it tells us very little about productivity.
The same conversation covered two practical questions facing developers who use coding agents: where the agents should run, and how much authority they should have. Steinberger argued for moving work into cloud infrastructure and suggested that one agent could negotiate with another agent responsible for granting permissions. Together, those comments describe a system in which agents can work more independently, but also create costs and risks that need explicit controls.
What the $10 Million Estimate Actually Tells Us
The figure is Steinberger’s estimate, reported from his conference remarks. It is not an independently audited spending statement. The Register attributes the funding to investors, but its account does not provide an allocation by model, workload or provider, or establish how the estimate maps to actual invoices.
Those missing details matter because token consumption, nominal token value and cash paid are different measurements.
A token is a unit that a model processes or generates. Providers can charge different rates for fresh input, cached input and output. The cost of a workload therefore depends on the model, the composition of its traffic and the applicable pricing arrangement. A dollar figure alone cannot reveal how many tokens were processed.
The report also leaves the useful-work denominator unknown. It does not say how many coding tasks were completed, how many experiments were abandoned, or how much output required correction.
That makes several tempting interpretations unsupported:
- It is not a typical OpenClaw operating cost. The report describes Steinberger’s intensive usage, not a published budget for an ordinary deployment.
- It is not a measured waste total. He said some usage was not useful, but did not quantify that portion.
- It is not a productivity benchmark. There is no corresponding count or assessment of completed work.
- It is not evidence that he personally paid a $10 million bill. The reporting points to investor funding and supplies no invoice-level accounting.
I’d treat the figure as evidence of how far a well-funded developer can push agent workloads, not as a forecast for what a development team should spend. Its usefulness is in prompting better accounting: what did all that activity produce?
Some unsuccessful experiments can still have value. A failed approach may reveal a limitation or eliminate an unproductive direction. Steinberger’s admission does not separate that kind of learning from avoidable repetition, runaway attempts or discarded output. The available evidence supports the existence of unproductive usage, but not a percentage or dollar amount.
Cloud Execution and Cloud Model Spending Are Separate Choices
Steinberger questioned why agent workloads still needed to use his computer and argued that infrastructure had not caught up with how he wanted to work, The Register reported.
OpenClaw’s own documentation helps unpack that argument. It describes the software as a self-hosted gateway connecting messaging services to AI coding agents. The gateway runs on a user’s machine or a server, manages sessions and routing, and connects to the agent runtime. The documented setup also calls for an API key from the chosen model provider. See the OpenClaw documentation.
That architecture separates three decisions that are easily bundled together under the word “cloud”:
| Layer | What it does | The practical decision |
|---|---|---|
| Gateway | Connects chat channels, routes messages and manages sessions | Where the control service runs and who can reach it |
| Agent execution | Carries out work using tools and a working environment | Which files, commands and systems the agent can access |
| Model inference | Processes input and generates responses | Which model service handles requests and how usage is charged |
Running a gateway locally does not, by itself, make model inference local or eliminate API charges. Equally, moving an execution environment to a remote server does not require handing that environment unlimited access to other systems.
OpenClaw’s documentation explicitly allows the gateway to run on a personal machine or a server. The consequential question is therefore broader than whether a laptop or a cloud machine hosts the process. It is how the deployment divides control, execution, data access and model calls.
Cloud execution can be attractive when jobs need to continue without depending on a developer’s laptop. It can also provide a separate environment for agent work. Those are architectural advantages, not evidence of lower total cost. Compute, storage and operational overhead can remain alongside model charges.
My view is that “move it to the cloud” is a reasonable infrastructure preference, but an incomplete cost strategy. It changes where work happens. It does not automatically reduce how much work an agent attempts or how many model calls it makes.
Why Agent Activity Can Grow Faster Than Useful Output
A coding request can trigger much more than a single model response. In a typical agent workflow, the system may inspect files, plan changes, call tools, read results, revise its approach and check the result. A multi-agent arrangement can add delegation and review.
Each stage can be useful. Each can also introduce more model traffic.
Consider a conceptual example: an agent attempts a fix, receives a failing test result, revises the patch and tries again. The second attempt may be necessary progress. Repeating an ineffective approach several more times adds cost without necessarily bringing the task closer to completion. This example explains the accounting issue; it is not a reconstruction of Steinberger’s usage.
The same distinction applies to context. A large amount of processed text may consist of material the system has already seen, such as source files, instructions and earlier tool results. That traffic still needs to be counted, but its cost can differ from newly supplied input or generated output.
A separate MindStudio write-up on orchestrated coding-agent teams illustrates how difficult that accounting can become. In its vendor-published comparison of five lead-and-worker model pairings across six tasks, MindStudio reported approximately 936 million recorded tokens and about $405 in recorded orchestration spending.
Its qualifications are more important than the headline totals:
- Only 8 of 30 task runs had complete captured usage data.
- Most recorded input consisted of repeated cached context.
- Some runs were interrupted by quota or provider errors.
- The teams did not receive equal computation budgets, and there was no solo-model baseline.
MindStudio consequently described its recorded spending as a floor, not a final bill. These figures neither corroborate Steinberger’s estimate nor provide a like-for-like comparison with his workloads. They do show why raw token totals and partial cost logs are poor foundations for declaring an agent setup efficient.
A task interrupted early can look cheap because it stopped before finishing. A successful task can look expensive because the logging captured every attempt. Without completion status and consistent accounting, a cost ranking may reward missing data.
Permission-Brokering Agents Still Need Enforced Boundaries
Steinberger’s other proposal concerned authority. Instead of giving a task agent every permission it might need, he suggested having another agent decide whether to grant access. The task agent would then negotiate with that permission-controlling agent, sometimes finding a simpler way to proceed within the guardrails.
Frequently Asked Questions
4 questions
1Did Peter Steinberger Personally Pay $10 Million?
The available reporting does not establish that. The Register attributes the funding to investors and reports Steinberger’s estimate without an invoice-level breakdown.
