OpenClaw's Creator Joined OpenAI — Then OpenAI Made OpenClaw Free. What's the Play?
Peter Steinberger built OpenClaw, then joined OpenAI. Days later, OpenAI made OpenClaw free for all paid users. Here's what that signals.
Peter Steinberger Built OpenClaw. Then He Joined OpenAI. Then OpenAI Made OpenClaw Free.
The timing is either a coincidence or a strategy. Either way, you should understand what happened.
Peter Steinberger, the creator of OpenClaw, is now working at OpenAI. Shortly after, Sam Altman stated publicly on May 1st that OpenClaw is “flat available under ChatGPT paid plans” — meaning every paid ChatGPT tier now subsidizes OpenClaw usage through Codex. The creator of the most popular open-source agent framework joined the company that then made that framework free to run on its infrastructure. If you build with OpenClaw, this changes your cost structure immediately. If you build AI products more broadly, it tells you something about how OpenAI thinks about agent distribution.
This is worth unpacking carefully, because the surface story (OpenAI made a thing free) is less interesting than the structural story underneath it.
The Sequence of Events
OpenClaw has been maturing fast. Through April 2026, the project shipped at a pace that would exhaust most funded product teams: task flow updates, memory updates, provider updates, channel integrations, code and automation tooling. The release notes alone read like a sprint log.
The most significant architectural addition is the task flow system — described in the docs as the orchestration layer above background tasks, managing durable multi-step flows with their own state and revision tracking while individual tasks remain detached units of work. That’s a meaningful shift from “agent that does stuff” to “runtime that manages work.” Webhook-triggered workflows, sub-agents that run their own sessions and report back, inspectable task queues — these are the boring words that signal a product becoming infrastructure.
Alongside this, OpenClaw shipped a provider manifest system that lets you swap the model brain at runtime without rebuilding the workflow. The supported providers now include GPT-5.5 via Codex, the Claude API, Gemini, DeepSeek, Open Router, Ollama, LM Studio, and Gemma 4. The workflow stays stable. The reasoning engine is pluggable.
That’s the product context. Now the business context.
In April, Anthropic moved to restrict the use of OAuth tokens from Claude Pro and Max subscriptions in third-party tools — including agent SDKs. The policy is defensible: agents aren’t normal chat users. They run longer, retry more, carry more context, sit in loops. A flat-rate consumer subscription was never designed to power always-on background agents at scale. Anthropic wanted Claude used as infrastructure to be paid for like infrastructure. Use the API, buy extra usage, stay inside official products.
The developer community disliked this intensely. And on the same day OpenAI took the opposite position.
OpenAI’s help docs now make Codex part of the ChatGPT subscription across all paid tiers. OpenClaw’s provider docs describe a Codex OOTH route alongside direct API usage. Sam Altman named OpenClaw specifically. The message was unambiguous: if you’re an OpenClaw user, route your work through Codex, and OpenAI will absorb the cost within your existing subscription.
Then there’s Peter Steinberger, who built OpenClaw, and who now works at OpenAI.
What This Actually Means for You
If you’re running OpenClaw workflows today, the immediate practical implication is straightforward: the Codex route is now the cheapest path for many workloads. You don’t need to restructure your workflow — the provider manifest lets you point at Codex without rebuilding anything. For builders who were routing through Claude subscriptions and just got cut off, this is a direct alternative.
But the more interesting implication is about distribution strategy, not pricing.
OpenAI is looking at OpenClaw’s user base — which is large, technically sophisticated, and actively building production workflows — and seeing a distribution channel. If OpenClaw users route work through Codex, Codex becomes more central to serious agent workflows. OpenAI’s infrastructure becomes the natural landing zone for open agent work. The “free” offer isn’t charity; it’s customer acquisition for the infrastructure layer.
This is a well-understood playbook in developer tools. You subsidize usage at the margin to win the architectural default. AWS did it with free tiers. Stripe did it with generous rate limits for early-stage companies. The cost of subsidizing a power user’s Codex calls is trivial compared to the value of that user’s workflow being built on your infrastructure.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
The Steinberger piece adds a layer to this. When the person who designed OpenClaw’s architecture is now inside OpenAI, the integration between OpenClaw and Codex is unlikely to be accidental or shallow. You’d expect the OOTH route to work well, to be maintained, to get better. You’d expect OpenClaw’s task flow system — with its durable multi-step flows, state tracking, and revision history — to be a natural fit for Codex’s agentic product surface. The creator of the runtime is now at the company running the compute.
For builders evaluating which model provider to anchor on, this matters. It’s not just about which model is best today. It’s about which provider has structural incentives to keep the integration working. The comparison between GPT-5.4 and Claude Opus 4.6 is a useful lens here — model capability differences are real, but they’re secondary to the question of which provider’s infrastructure you’re building your loop around.
The Non-Obvious Detail: Memory Is the Real Battleground
Here’s what gets underweighted in the “OpenAI vs. Anthropic” framing: neither of them owns your memory.
The provider manifest makes the model swappable. But if your agent’s operational context — project conventions, prior decisions, failure history, task state — lives inside a single provider’s product, you’ve just moved the lock-in from the model to the memory layer. You can swap the brain, but the brain took the knowledge with it.
This is why the OpenBrain project matters in this context. The open-source GitHub repo now includes a code review memory recipe that stores reusable lessons from PRs, a task flow worklog recipe that records what a long-running agent attempted and what blocked it, and a memory provenance recipe that labels where each piece of memory came from. The provenance labels are specific: observed from source, inferred by model, confirmed by user, imported from transcript. Those aren’t aesthetic choices — they’re the difference between memory you can trust as an instruction and memory that’s just accumulated sludge.
The architecture Nate Jones describes in his analysis is worth sitting with: OpenClaw gives agents an action layer, task flow gives work a durable loop, channels are where humans interact, and memory is the continuity layer. The model is the reasoning engine inside a much larger operating loop. When you can swap the engine, the loop itself becomes the valuable thing — and the loop’s memory needs to live outside any single engine.
For builders who want to build on top of this without writing the orchestration plumbing themselves, MindStudio handles multi-model routing across 200+ models with 1,000+ pre-built integrations and a visual builder for orchestrating agents and workflows — which means you can experiment with the same provider-agnostic pattern without standing up the infrastructure from scratch.
The practical implication: if you’re building serious workflows on OpenClaw right now, your memory architecture is more strategically important than your model choice. The model wars will continue. Anthropic will make more policy changes. OpenAI will make more offers. Google released Gemma 4 under Apache 2.0 explicitly positioned for agentic workflows and on-device use — a credible local branch of the runtime tree for steps that don’t need frontier model pricing. The model landscape will keep shifting.
Your memory layer, if you own it, doesn’t shift with it.
The Broader Signal for Agent Builders
There’s a pattern here that extends beyond OpenClaw specifically.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The labs are fighting over the agent runtime layer. Anthropic’s managed agents product — which provides sandboxing, state management, and error recovery as a service — is one approach. OpenAI’s Codex subscription integration is another. Google’s Gemma 4 positioning for edge and agentic use is a third. Each of these is a bid to become the default infrastructure for serious agent work.
What’s interesting about the OpenClaw situation is that it shows a different vector: instead of building the harness themselves, OpenAI is subsidizing the most popular open-source harness and hiring its creator. That’s a faster path to distribution than building from scratch, and it’s a more credible path to developer trust than a first-party product that competes with the tools developers already use.
For builders, the lesson is architectural. OpenClaw best practices from power users consistently point toward the same conclusion: don’t build your workflow around a single provider’s subscription terms. Build the runtime so the model can change. The April events — Anthropic’s restriction and OpenAI’s counter-offer landing on the same day — are a demonstration of why.
The breakdown of Anthropic, OpenAI, and Google’s agent strategies is useful context here. Each lab is making different bets on where the value accrues in the agent stack. OpenAI’s bet, evidenced by the Codex move, is that distribution through existing developer workflows is worth subsidizing. Anthropic’s bet is that controlling the terms of infrastructure use protects margins and capacity. Google’s bet, with Gemma 4 under Apache 2.0, is that open models at the edge create a different kind of lock-in — one based on deployment rather than subscription.
None of these bets are obviously wrong. All of them will affect your cost structure and your architectural choices.
What to Do This Week
The Codex OOTH route in OpenClaw’s provider docs is live now. If you’re running workflows that were previously anchored to Claude subscriptions and got caught by Anthropic’s April policy change, the path forward is concrete: update your provider manifest to point at Codex, test your existing task flows, and verify that your memory layer is external to the model rather than embedded in a single provider’s context.
If you haven’t looked at the OpenBrain recipes yet, the memory provenance recipe is the one to start with. The distinction between “observed from source” and “inferred by model” sounds academic until you’re debugging why your incident response workflow confidently recommended a rollback based on a model inference that was never confirmed. Provenance labels make memory auditable. Auditable memory makes workflows trustworthy enough to run unsupervised.
For the model routing question specifically: the right answer isn’t “use Codex for everything now that it’s free.” The right answer is the one Nate Jones articulates — use a local Gemma-class model for cheap background classification and low-risk triage, use GPT-5.5 through Codex for hard implementation work, use the Claude API when the judgment or architectural reasoning is worth the metered cost. The guide to running Claude Code free via Ollama and Open Router is a practical starting point for the local model side of that routing decision.
On the spec and tooling side: as workflows get more complex, the gap between “what the workflow is supposed to do” and “what the code actually does” becomes a real maintenance problem. Tools like Remy take a different approach to this — you write your application as an annotated markdown spec, and the full-stack app (TypeScript backend, database, auth, deployment) gets compiled from it. The spec stays as the source of truth; the code is derived output. For teams building workflow tooling on top of OpenClaw, that abstraction layer is worth understanding.
The Peter Steinberger situation is, at minimum, a signal that OpenAI is serious about the open agent ecosystem rather than just its own first-party products. Whether that translates into sustained investment in the Codex integration, or whether it’s a short-term distribution play, will become clear over the next few quarters. Watch the provider docs. Watch the rate limits. Watch whether the OOTH route gets better or gets quietly deprioritized.
The model wars will keep producing these moments — a policy change here, a free tier there, a hire that looks coincidental but probably isn’t. The builders who come out ahead are the ones who treat each of these as information about incentives rather than as permanent facts about the landscape.
Build the loop. Own the memory. Keep the brain swappable. The rest is noise.
