Anthropic Restricts Third-Party Agents, OpenAI Opens Up: Which Provider Should You Build On?
Anthropic locked down always-on agent subscriptions. OpenAI opened Codex to everyone. Here's how to pick the right provider for your agentic workflow.
Two Providers, Two Bets, One Architecture Decision You Have to Make
Anthropic and OpenAI made opposite calls in April 2026, and if you’re building agentic workflows on top of either provider, you need to understand what each call actually means for your system. Anthropic restricted subscription use for always-on third-party agents at scale. OpenAI opened Codex to all paid ChatGPT plans and added a Codex OOTH route to OpenClaw’s provider docs. Same month, opposite directions.
The choice isn’t just “which model is better.” It’s “which provider’s policy assumptions does my architecture depend on?” Those are different questions, and conflating them is how you end up rebuilding your agent infrastructure six months from now because a subscription policy changed.
Here’s how to think through it.
Why the Policy Gap Opened Up Now
Anthropic’s position is coherent, even if it was unpopular. Claude subscriptions were priced for human users having conversations — not for background agents that run loops, retry tool calls, carry large context windows, and generate intermediate work that no human ever sees. An agent doing serious work consumes tokens at a fundamentally different rate than a person asking for help drafting an email. Anthropic wants that usage priced like infrastructure, not like a seat license.
The developer community reacted badly. But Anthropic was also operating under real compute constraints — a company in hypergrowth mode making hard choices about where to allocate capacity. The restriction wasn’t arbitrary. It was a margin decision under pressure.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
OpenAI’s move was the mirror image. Sam Altman stated explicitly on May 1st that OpenClaw is now available under all paid ChatGPT plans. The Codex OOTH route appeared in OpenClaw’s provider docs alongside direct API usage. OpenAI has more online compute right now, and they’re using that surplus to pull agent workloads toward their infrastructure. If OpenClaw users route work through Codex, OpenAI’s agent strategy gets reinforced. It’s a distribution play disguised as a policy decision.
The Peter Steinberger detail matters here. When the creator of OpenClaw joins OpenAI and OpenAI immediately makes Codex available to all OpenClaw users, that’s not coincidence — it’s alignment of incentives made visible. Anthropic had the model many early OpenClaw users loved. OpenAI now has Codex, subscription access, and a strong structural reason to make these workflows feel native on their infrastructure.
The Five Dimensions That Actually Determine Your Choice
Before the side-by-side, you need a framework. The model quality debate — Claude vs. GPT — is real but secondary. The dimensions that determine which provider you should build on are different.
1. Pricing model fit. Are you running always-on background loops, or are you running on-demand tasks triggered by user actions? Always-on loops at scale are exactly what Anthropic restricted. On-demand tasks fit the subscription model fine.
2. Subscription vs. API dependency. If your workflow depends on a flat-rate subscription to stay economically viable, Anthropic’s April move broke that assumption. If you’re already on the API with metered usage, the policy change is largely irrelevant to you.
3. Task type and judgment requirements. Some steps in an agentic workflow genuinely need high-judgment architectural reasoning — the kind of careful, nuanced output Claude has been known for. Other steps need fast, cheap classification. The right provider for step one is not necessarily the right provider for step seven.
4. Runtime durability. What happens to your workflow when a provider changes their policy? If the answer is “it breaks,” your architecture has a single point of failure that isn’t technical — it’s contractual.
5. Ecosystem lock-in. Memory, tool permissions, channel integrations, task state — if any of these live inside one provider’s product, you’ve traded portability for convenience. That trade has a cost that becomes visible exactly when you don’t want it to.
Anthropic: Premium Judgment, Metered Access, Deliberate Constraints
Claude’s strengths for agentic work are real and specific. For high-judgment tasks — architectural review, nuanced writing, sensitive decisions where the reasoning chain matters — Claude API remains a strong choice. The model’s tendency toward careful, hedged reasoning is a feature when you’re doing the kind of work where confidently wrong is worse than slow.
The constraint is the pricing model. If you want Claude in your agent workflow, you’re on the API now for any serious background processing. That means metered costs, which means your economics look different than they did when you could run Claude loops against a flat subscription. For workflows that call Claude frequently on cheap classification tasks, this gets expensive fast. For workflows that call Claude selectively on the hard steps, the metered model is fine — you’re paying for what you use, and what you’re using is genuinely worth paying for.
The practical implication: Claude becomes a premium, selective component in a multi-model workflow. Not the always-on substrate. Not the cheap background brain. The model you route to when the step actually requires its particular strengths.
There’s also a strategic consideration. Anthropic is making choices that reflect compute constraints and margin pressure. That’s not a criticism — it’s a fact about their current situation. But it means the policy environment around Claude subscriptions is likely to remain restrictive, and possibly tighten further, until their compute position improves. Building an architecture that depends on Anthropic’s subscription terms staying stable is a bet on their infrastructure situation resolving favorably.
For builders doing real-world coding comparisons between GPT-5.5 and Claude Opus, the token efficiency gap compounds this: Claude Opus uses significantly more output tokens on the same tasks, which matters a lot when you’re paying per token on the API.
OpenAI: Distribution Play, Subscription Access, Agentic Infrastructure Push
OpenAI’s April posture is the opposite of Anthropic’s. Codex is already an agentic product. Making it available under all paid ChatGPT plans, and adding the Codex OOTH route to OpenClaw’s provider docs, signals that OpenAI wants to be the default infrastructure for agent workflows — not just the model provider.
The /goal feature in Codex is the clearest expression of this. It keeps a goal alive across turns and doesn’t stop until the goal is achieved. A16Z’s Andrew Chen ran it on an eGPU plus Mac device driver project for 14 hours unattended. Alex Finn built a complete extraction shooter video game in over an hour of continuous operation. The “Ralph loop” framing — Philip Corey’s term for it at OpenAI — describes a system that can run for days without stopping. That’s not a chat feature. That’s infrastructure behavior.
The subscription access matters for economics. If you can run Codex-backed agent loops under a paid ChatGPT plan, your cost structure for background processing is fundamentally different than metered API calls. For high-volume, long-running workflows — the kind that OpenClaw’s TaskFlow orchestration layer is designed to support — this is a meaningful advantage.
The risk is the same risk that always comes with distribution plays: the terms can change. OpenAI is making this move because it serves their strategy right now. If the strategy shifts, the terms shift. The Anthropic restriction is a reminder that provider policies are not permanent architectural commitments.
GPT-5.5 through Codex is also the right choice for hard implementation and complex repo work, according to the model routing framework that’s emerged from serious OpenClaw builders. Not because it’s categorically better than Claude, but because it’s the right tool for that specific step — and the subscription economics make it viable to use it there at scale.
For a broader look at how these three major labs are positioning their agent strategies differently, the Anthropic vs OpenAI vs Google agent strategy comparison is worth reading alongside this.
The Routing Framework That Makes the Choice Less Binary
Here’s the opinion: the Anthropic vs. OpenAI framing is the wrong frame for most builders. The right frame is which model handles which step, and why.
The routing recommendation that’s emerged from serious OpenClaw users in April 2026 is specific: use a local Gemma-class model (Google released Gemma 4 under Apache 2.0 explicitly for agentic and edge use cases) for cheap background classification, duplicate detection, and low-risk triage. Use GPT-5.5 through Codex for hard implementation and complex repo work. Use Claude API when the judgment, the writing style, or the architectural reasoning is worth the metered cost. Use cheaper hosted models for bulk summarization and formatting.
This isn’t a hedge. It’s an acknowledgment that no single model is optimal for every step in a complex workflow. The Anthropic restriction and the OpenAI opening are both easier to navigate if your workflow was never dependent on one provider to begin with.
The memory layer is where this gets concrete. If your agent’s memory lives inside one provider’s product — inside a chat transcript, inside a model’s context, inside a subscription-backed session — then your workflow is locked to that provider’s policy decisions. OpenClaw’s memory direction with memory wiki, active memory, and providence-rich recall points toward a different model: memory as an operational layer that lives outside any single brain.
The memory provenance labels matter here. Observed from source, confirmed by user, inferred by model, imported from transcript — these distinctions determine whether memory is trustworthy enough to act on. Without clear provenance, agent memory becomes a pile of accumulated assertions with no way to distinguish reliable context from confident hallucination. With good provenance labels, memory becomes the continuity layer that lets you swap models without losing the workflow’s history.
Platforms like MindStudio handle this kind of orchestration across providers: 200+ models, 1,000+ integrations, and a visual builder for chaining agents and workflows — which means the routing decisions described above don’t require you to write the orchestration code from scratch.
Verdict: Use Anthropic If X, Use OpenAI If Y
Use Claude API if: Your workflow has specific steps that require high-judgment architectural reasoning, nuanced writing, or careful analysis where the cost of confident wrongness is high. You’re already on metered API pricing and the economics work. You’re building a selective, multi-model workflow where Claude handles the hard steps and cheaper models handle the rest. You want the model that has historically been strongest at the kind of careful reasoning that matters for sensitive decisions.
Use OpenAI/Codex if: You’re running always-on background loops that need to stay economically viable under subscription pricing. Your workflow involves long-running autonomous tasks — the kind /goal is designed for, where the agent works for hours or days without stopping. You’re doing hard implementation work on complex repos where GPT-5.5 through Codex is the right tool for the step. You want the provider that’s actively incentivized to make agent workflows feel native on their infrastructure right now.
Use neither exclusively if: You’re building anything that needs to survive a provider policy change. The Anthropic restriction in April is a data point, not an anomaly. Provider policies will keep changing as compute economics shift, as frontier models get more expensive to serve, and as labs make strategic choices about where they want agent workloads to land. The durable architecture routes different models to different steps, owns its memory independently of any provider, and treats model choice as a routing decision rather than an architectural commitment.
For builders evaluating sub-agent model choices specifically, the GPT-5.4 Mini vs Claude Haiku sub-agent comparison covers the cost and performance tradeoffs at the cheaper end of the model stack — which is where most of your agent’s actual token volume will land if you’re routing correctly.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
The question “which provider should I build on” has a clean answer only if you’re building something simple. For serious agentic work — the kind that OpenClaw’s TaskFlow orchestration layer, durable multi-step flows with their own state and revision tracking, is designed to support — the answer is: build the runtime so the model can change, own the memory so the user controls it, and route based on what each step actually needs.
The labs will keep making policy decisions that reflect their incentives. Some paths will open, some will close. The builder response is architecture, not loyalty.
If the broader question is how to go from workflow spec to deployed application, tools like Remy take a different approach: you write your application as annotated markdown — a spec where readable prose carries intent and annotations carry precision — and it compiles into a complete TypeScript backend, database, auth, and deployment. The spec is the source of truth; the generated code is derived output. It’s a different layer of the same abstraction problem that model routing solves.
For builders who want to go deeper on how these models actually compare on the tasks that matter for agentic coding, the Qwen 3.6 Plus vs Claude Opus 4.6 agentic coding comparison and the three-way benchmark across GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro are worth reading before you finalize your routing decisions. The benchmark results often surprise people who’ve been relying on reputation rather than data.
The policy gap between Anthropic and OpenAI is real, and it matters. But it’s most useful as a forcing function to build the architecture you should have built anyway — one where the model is a component, not the foundation.
