Anthropic vs. OpenAI on Agent Token Access: Two Opposite Bets on the Same Day
On the same day Anthropic banned subscription tokens in third-party agents, OpenAI made Codex free for all paid users. Here's what each bet means.
The Day Anthropic and OpenAI Made Opposite Bets on Your Agent
On the same day in late April 2026, Anthropic told you that using OAuth tokens from a Claude Pro or Max subscription inside third-party agent tools — including the agent SDK — was no longer permitted. OpenAI, on the same day, made Codex freely available across all paid ChatGPT tiers. Same moment. Opposite directions.
That’s not a coincidence. It’s a strategy collision, and you’re the one who has to build around it.
If you’ve been running OpenClaw with Claude as the default brain, this is the policy change that forces a real architectural decision. Not “which model do I prefer” — that’s a taste question. This is “what happens to my workflow when a provider changes the rules,” and the answer you give now determines how fragile your agent stack is for the next eighteen months.
What Each Company Is Actually Betting On
The surface read is that Anthropic is being restrictive and OpenAI is being generous. That’s not quite right.
Anthropic’s position is coherent. Subscription plans were priced for human users having conversations — not for always-on agents running background loops, retrying tool calls, carrying large context windows, and generating intermediate tokens a human never sees. An agent running inside OpenClaw doesn’t behave like a Pro subscriber. It behaves like infrastructure. Anthropic wants infrastructure priced like infrastructure: use the API, buy metered access, stay inside official products.
Dario Amodei put some numbers around why this matters. In Q1 2026, Anthropic saw 80x annualized growth in revenue and usage. They were already straining under compute constraints severe enough that daily Claude usage was getting interrupted. The subscription token policy wasn’t just terms enforcement — it was capacity management under genuine pressure.
OpenAI’s position is also coherent, just from the opposite angle. They have more compute online. Sam Altman said explicitly on May 1st that OpenClaw is “flat available under ChatGPT paid plans.” OpenClaw’s provider docs now describe a Codex OOTH route alongside direct API usage. If you route agent work through Codex, you’re routing it through OpenAI’s infrastructure — and OpenAI gets distribution inside one of the most popular open-source agent frameworks in the world.
Two rational companies, same moment, opposite calls.
The Dimensions That Actually Matter for Builders
Before picking a side — or more accurately, before deciding how to architect around both — here are the dimensions worth evaluating.
Cost model predictability. Subscription access is flat-rate until it isn’t. API access is metered but predictable at scale. Anthropic is pushing you toward metered. OpenAI is subsidizing flat-rate through Codex. Both will change. The question is which change is easier to absorb.
Workflow lock-in. If your OpenClaw task flow — the orchestration layer above background tasks that manages durable multi-step flows with state and revision tracking — is wired to one provider’s auth model, a policy change breaks the workflow, not just the model call. That’s a much bigger problem than swapping a model.
Model quality per step. Not every step in a durable workflow deserves frontier model pricing. A local Gemma 4 model (Apache 2.0, explicitly positioned for agentic workflows and on-device use) can handle cheap background classification. GPT-5.5 via Codex can handle hard implementation work. Claude API, metered, can handle the high-judgment architectural passes where its writing and reasoning are genuinely worth the cost. The routing question is more interesting than the “which model is best” question.
Provider manifest portability. OpenClaw’s provider manifest lets you swap the model brain at runtime without rebuilding the workflow. The supported providers now include GPT-5.5 via Codex, Claude API, Gemini, DeepSeek, Open Router, Ollama, LM Studio, and Gemma 4. That list is the answer to provider lock-in — but only if your workflow is designed to use it.
Memory independence. This is the one most builders underweight. If your agent’s operational memory lives inside a single provider’s product, you don’t have a portable workflow. You have a workflow that works until the provider changes something. The OpenBrain project’s memory provenance model — labeling memories as observed from source, inferred by model, confirmed by user, or imported from transcript — exists precisely because agent memory needs to survive model swaps.
Anthropic’s Bet: Premium, Metered, Controlled
Anthropic is making a specific argument: Claude is good enough to be worth paying for like infrastructure.
That’s not a weak position. Claude’s judgment, writing quality, and architectural reasoning are genuinely differentiated. For the steps in a workflow where those qualities matter — sensitive email triage, high-stakes code review, nuanced incident postmortems — Claude API access at metered cost is defensible. You’re paying for a specific capability, not a subscription.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
The problem is the transition cost. A lot of OpenClaw builders started with Claude as the default brain because it was accessible through a subscription they already had. The policy change doesn’t just add a cost — it changes the default architecture. Claude becomes a premium, intentionally-invoked component rather than the always-on substrate for background loops.
For builders who were using Claude subscription tokens as cheap infrastructure, this is a hard reset. For builders who were already treating Claude as a high-value metered component for specific workflow steps, the policy change barely registers.
The deeper issue is developer goodwill. Anthropic knew this would be unpopular. They made the call anyway, which is the right call when you’re running 80x annualized growth into a compute wall. But “correct under constraint” and “popular with the developer community” are different things, and Anthropic is paying a real cost in the latter even if the former is justified.
The Anthropic vs OpenAI vs Google agent strategy comparison is worth reading alongside this — the subscription token policy fits a broader pattern of Anthropic trying to control the terms under which Claude is used as infrastructure, not just as a chat assistant.
OpenAI’s Bet: Distribution Through Subsidy
OpenAI’s move is simpler to describe: they have compute, they want OpenClaw users, and they’re willing to subsidize access to get them.
The Codex OOTH route in OpenClaw’s provider docs is not a coincidence. When the creator of OpenClaw — Peter Steinberger — is now working at OpenAI, and OpenAI is simultaneously making Codex free across all paid ChatGPT tiers, the power dynamic shifts. Anthropic had the model many early OpenClaw users loved. OpenAI now has Codex, subscription access, and a strong incentive to make open agent workflows feel native on its infrastructure.
This is distribution strategy dressed as generosity. It works because it’s genuinely useful — Codex via ChatGPT paid plans is real access to a capable model for agent work, not a degraded tier. GPT-5.5 handles hard implementation and complex repo work well. The real-world coding performance comparison between GPT-5.5 and Claude Opus 4.7 shows the tradeoffs: GPT-5.5 uses significantly fewer output tokens on the same tasks, which matters for agent loops where token efficiency compounds.
The risk for builders who go all-in on the Codex route is the same risk Anthropic just demonstrated: provider policies change. OpenAI is subsidizing this access because it serves their distribution goals right now. When those goals change, the subsidy can change. The architecture lesson from Anthropic’s April move applies equally to OpenAI’s May move.
What This Means for Your Workflow Architecture
The wrong takeaway from this story is “OpenAI won, use Codex.” The right takeaway is that both companies just demonstrated, on the same day, that provider policies are a runtime variable — and your workflow architecture should treat them that way.
OpenClaw’s provider manifest is the practical tool here. It enables swapping the model brain at runtime without rebuilding the workflow. That’s not a convenience feature. It’s the architectural answer to exactly what happened in April. If your task flow — with its durable multi-step state, revision tracking, and detached background tasks — is wired to a specific provider’s auth model, you’ve built fragility into the foundation.
One coffee. One working app.
You bring the idea. Remy manages the project.
The incident response workflow pattern makes this concrete. You have logs, dashboards, Slack, GitHub, runbooks, and deployment history all feeding into one OpenClaw loop. A fast model handles log analysis. A cheap model drafts status updates. A deep inference model does root cause analysis. The workflow doesn’t care which brain handled which substep — and it shouldn’t. If Anthropic changes a policy or OpenAI changes a pricing tier, you reroute one step, not the whole loop.
The same logic applies to the email triage workflow: segregate sensitive emails with high judgment, draft replies, run a QA pass, handle attachments — multiple models, multiple routing decisions, one durable loop. The loop is the product. The model is a component.
For teams building these kinds of multi-model workflows without writing orchestration code from scratch, MindStudio offers a no-code path: 200+ models, 1,000+ integrations, and a visual builder for chaining agents and workflows across providers. The provider-agnostic architecture that OpenClaw is building toward in open-source form is the same principle MindStudio applies to the broader agent-building space.
Memory is the other half of this. If your agent’s operational context — project conventions, prior decisions, past failures — lives inside one provider’s product, you have the same lock-in problem as subscription token dependency. The OpenBrain recipes for OpenClaw address this directly: a code review memory recipe that stores reusable lessons from PRs, a task flow worklog that records what a long-running agent attempted and what blocked it, and a memory provenance recipe that labels where each memory was observed and confirmed. When you can swap models, you need memory that travels with the workflow, not with the brain.
The spec-driven approach to building is relevant here too. Tools like Remy take the position that the source of truth should be a structured spec — annotated markdown that carries intent and precision — compiled into a complete TypeScript stack. The principle is the same as provider-agnostic agent architecture: the source of truth should be something you own, not something that lives inside a vendor’s product.
Which Route to Take, Given What You’re Building
Use Claude API (metered) if the specific workflow steps you’re routing to Claude are ones where its judgment, writing quality, or architectural reasoning are genuinely differentiated and worth the cost. Don’t use it as cheap background infrastructure — that’s what Anthropic just told you they don’t want, and they’re right that it was never priced for that.
Use Codex via ChatGPT paid plans if you’re doing hard implementation work, complex repo tasks, or anything where GPT-5.5’s token efficiency on coding tasks gives you a real cost advantage. The sub-agent model comparison between GPT-5.4 Mini and Claude Haiku is a useful reference for thinking about which tier of model to route to which tier of task.
Use Gemma 4 or local models via Ollama or LM Studio if you’re doing background classification, duplicate detection, low-risk triage, or any step where frontier model pricing is genuinely wasteful. Google released Gemma 4 under Apache 2.0 with explicit positioning for agentic workflows and on-device use. That’s a credible local branch of the runtime tree, not a fallback.
Use Open Router or DeepSeek if you want hosted model access without committing to a single provider’s pricing structure. The provider manifest supports them both.
Don’t use any single provider as your only option. That’s the lesson April 2026 delivered, and it was delivered by both companies simultaneously.
The Actual Stakes
Day one: idea. Day one: app.
Not a sprint plan. Not a quarterly OKR. A finished product by end of day.
The model war over OpenClaw’s brain is going to continue. Frontier models are expensive to train and strategically critical. Every major lab has an incentive to make their model the default brain inside the most popular open-source agent framework. Some paths will open, some will close, and the timing will rarely be convenient for you.
The builders who come out of this well aren’t the ones who picked the right provider in April. They’re the ones who built workflows that survive provider decisions — with a runtime abstraction that can swap brains, memory that lives outside any single product, and routing logic that matches model capability to task cost.
Anthropic and OpenAI made opposite bets on the same day. The correct response isn’t to pick a winner. It’s to build so that you don’t have to.
