OpenAI won one AI access argument on X. The evidence stops there.

On April 21, a Claude user reported that Anthropic's public pricing page appeared to remove Claude Code from its $20 Pro plan for prospective subscribers. Anthropic growth leader Amol Avasare said the change was a test affecting about 2% of new prosumer signups, with existing subscribers unaffected. Sixty-four minutes later, OpenAI product leader Thibault “Tibo” Sottiaux promised that Codex would remain available on the free and $20 Plus plans. Sam Altman amplified that promise 47 minutes after Sottiaux posted.

OpenAI won that access-and-communication exchange on X. It offered a simple plan-level promise while Anthropic was explaining a confusing experiment and later reverting public page changes. The evidence does not support a broader victory over Anthropic in market share, model quality, safety, revenue, or long-term strategy. In selected posts from this April 21–July 14 window, Anthropic displayed higher X view counts and published substantial first-party evidence about Claude Code use; its CEO also advanced a specific safety-policy proposal.

OpenAI made the April contrast easy to repeat

Anthropic's problem was not only the size of its pricing test. People outside the test saw a changed landing page and documentation, then learned from a user screenshot rather than a company announcement. Avasare clarified the 2% scope 76 minutes after the first report. Early the next morning, he said the public updates had confused the other 98% and that Anthropic had reverted them.

That response supplied context and a rollback. It also required readers to follow an explanation about cohorts, heavier agent use, plan economics, and future options. Sottiaux compressed the competitive contrast into one post: Codex would remain on free and Plus, OpenAI said it had the compute and efficient models to support that access, and important changes would be discussed with the community in advance.

The values language in Sottiaux's post went beyond what a subscription promise can establish. The communication advantage was still clear. A developer deciding where to spend $20 could understand the offer immediately, and Altman added executive reach without complicating it.

X's public counters also show the limits of this comparison. When rechecked on July 14, Sottiaux's April response displayed 1.7 million views and 577 replies; Avasare's clarification displayed 6.9 million views and 1,300 replies. At that snapshot, Anthropic's clarification displayed more views and replies, while OpenAI's response stated the access promise more simply. These mutable counters do not reveal whether people approved, objected, or intended to buy.

Sottiaux strengthened the case by explaining failures and repairs

OpenAI's stronger evidence came after the April exchange. Sottiaux repeatedly connected access claims to visible product operations.

On June 29, after users reported unexpectedly fast Codex usage consumption, he identified several contributing causes: more proactive auto-review, additional subagent work, duplicate background suggestions, retries, and incorrect usage reporting. OpenAI reset limits, credited another reset, reverted or fixed the affected behavior, and changed what appeared in usage charts. The post separated real consumption from display errors and warned that features designed to perform more work would still use more capacity.

After OpenAI launched GPT-5.6 Sol and ChatGPT Work, Sottiaux published another correction on July 10. He said high-compute settings had been too easy to select without making their effect on limits clear. He also named desktop-navigation problems, regressions in multi-agent workflows, plugin issues, and launch language that made some users think Codex was being phased out. OpenAI reset usage twice while changing defaults and preparing interface fixes.

Two days later, Sottiaux said OpenAI was temporarily removing the five-hour usage-limit restriction for Plus, Business, and Pro plans, announced another reset, and reported six million active users in a post about Codex and ChatGPT Work. The post did not define the active-user population or measurement window. The restriction removal was temporary and the user figure was company-reported; neither established a permanent price guarantee. The post still gave users a specific change they could check in their own accounts.

This communication pattern is OpenAI's strongest X argument: name the failure, explain the mechanism, describe the repair, and provide immediate relief. It is more useful than an executive slogan because customers can compare the explanation with the product they are using.

Anthropic's higher observed X view counts and product evidence block a broader verdict

OpenAI does not have an uncontested advantage in the selected X view counts. On July 12, Anthropic extended Claude Fable 5 across paid plans and kept Claude Code weekly limits 50% higher through July 19. When rechecked on July 14, X displayed 26.2 million views and 6,700 replies for that post. Sottiaux's temporary five-hour-limit removal, posted 58 minutes later, displayed 4.5 million views and 2,800 replies. Both offers were time-bound. At that snapshot, the Anthropic post displayed far more views than Sottiaux's post.

Anthropic has deeper product evidence than a promotional post. Its June 16 analysis of Claude Code use covered about 400,000 sessions from roughly 235,000 people between October 2025 and April 2026. Anthropic reported an average of 20 hours of use per week, a relationship between domain expertise and success, and a shift toward more complete agentic work.

The study is first-party research based on Anthropic's product and methods. It cannot independently establish market leadership. It does provide more detail about observed work than a growth number on X.

Anthropic also occupies a distinct safety-policy position. In June, CEO Dario Amodei argued for mandatory third-party testing of frontier models for cyber, biological, and autonomy risks, including authority to block or revoke deployments that pose catastrophic risk. A buyer may disagree with that proposal. It is still a concrete policy position, and it addresses a different trust question from subscription access or incident communication.

X shows the argument each company wants customers to hear

Across the April 21 to July 14 window, OpenAI leads on the clarity of the initial access response and the cadence of product-level incident updates. Among the posts examined, Anthropic recorded the highest observed X view count and supplied strong first-party evidence about Claude Code use. Its product distribution and its CEO's safety-policy proposal also give buyers reasons that cannot be reduced to a $20 subscription comparison.

Calling the exchange a cold war is useful only as a bounded metaphor for a public influence contest. The posts do not establish which company has better models, more customers, higher revenue, safer deployments, or the stronger long-term strategy. They show how each company wants access and trust to be interpreted: OpenAI emphasizes abundant use and fast repair; Anthropic emphasizes capable products, research, distribution, and its CEO's safety-policy proposal.

For buyers, X is an incident feed, not a procurement record. Verify which plan includes the product, how usage is measured, which limits are temporary, how much notice the provider promises, and whether the same terms appear in current documentation or a contract. If a workflow cannot tolerate a surprise change, define and test the fallback before another public argument begins. Our guide to model-access interruptions explains how to test that fallback, while the model bake-off guide covers the separate question of production performance.

In this exchange, OpenAI made the sharper case about access and repair. Anthropic's higher observed X view counts, product evidence, and policy differentiation mean that any broader victory claim would outrun the record.