Anthropic grows fast; OpenAI's GPT-5.6 still pulls more users

Anthropic's annualised revenue reached $65bn in July, up from $47bn in May. That is a striking clip. The Financial Times, drawing on people with knowledge of the matter and cited by Simon Willison's Weblog, also reports that Anthropic expects Q3 to be profitable and has 6,000 customers spending $100,000 or more annually. By the conventional scorecard, the company is winning.

Except OpenAI's annualised revenue has grown 35% in the current quarter alone and now exceeds $40bn, propelled by the July launch of GPT-5.6. That number is lower in absolute terms, but the trajectory is the more pointed fact. Anthropic leads on revenue; OpenAI is accelerating faster on users.

The enterprise counts, but the crowd sets the default

The 6,000 anchor customers are genuinely significant. At $100,000-plus annually, they represent a high-trust, high-commitment segment: precisely the procurement profile of a large industrial group, a multilateral institution, or a regulated financial services firm evaluating which AI vendor to standardise on. Claude's reputation for instruction-following, caution on sensitive topics, and longer context windows has made it the preferred choice for organisations whose legal and compliance teams review every new tool before deployment.

But enterprise preference does not automatically translate into the citation patterns that matter for brand visibility in LLM answers. The models most used for general information retrieval, the ones a policy analyst at a UN agency or a communications lead at a major bank reaches for first, tend to be the ones with the broadest consumer footprint. That footprint, at this moment, skews toward ChatGPT. GPT-5.6's launch was, by the FT's account, a jolt after a sluggish start to the year. Anthropic's flagship Claude model, meanwhile, is described as struggling to attract users even as revenue rises. Revenue and reach have decoupled.

That decoupling has direct consequences for how often each model's outputs shape perception of a given brand or topic. LLM citation behaviour correlates, imperfectly but observably, with training data recency, model deployment scale, and the volume of user interactions that generate reinforcement. A model with a smaller active user base processes fewer queries, draws on a narrower prompt distribution, and is less likely to surface a given organisation's content in the answers it generates at scale.

Cheaper tools, not better ones, are setting the pattern

The FT framing points to a third dynamic beyond the Anthropic-OpenAI binary: cheaper tools are thriving. The Ramp AI index, which Simon Willison flags in his post, tracks enterprise AI spending by product and reportedly shows users gravitating toward cost-efficient options. For brand visibility in AI-generated answers, this creates a more fragmented retrieval environment than anyone planned for. Organisations optimising their content for Claude may find their messaging surfaces less consistently in the cheaper, higher-volume models that a broader audience actually uses day to day.

For a communications team at, say, a major industrial group publishing technical guidance or sustainability reporting, the practical implication is uncomfortable. The content strategy that earns a citation in a Claude-generated answer may not be the same strategy that earns one in a GPT-5.6 answer, and neither maps cleanly onto the optimisation logic that governs retrieval in Gemini or Perplexity. Revenue share and user share now point in different directions, which means the audience a brand thinks it is reaching through AI channels and the audience it is actually reaching are not the same.

Anthropic's numbers are impressive by any prior benchmark. The harder question is whether impressive revenue from a concentrated enterprise base translates into the diffuse, high-frequency model usage that shapes what AI systems say about a brand when no one is specifically asking. On that measure, GPT-5.6's user jolt matters more than the revenue gap suggests.