Anthropic debuts Opus 5 at half the price of Fable 5, the same week the U.S. debates banning Chinese open AI

By AI Ramblings Daily · July 24, 2026.

Anthropic has launched Claude Opus 5, according to the company's official announcement. The model sits above Opus 4.8—its workhorse until now—and slightly below Fable 5 (the model Anthropic also calls Mythos) on most benchmarks, but costs half as much. The relevant point is not just technical: that price leaves Opus 5 practically level with OpenAI's GPT-5.6, as analyst Michael Parekh notes in his AI Ramblings Daily newsletter. Both frontier labs are competing with their top-tier closed models, aimed above all at businesses and at programmers who consume AI at scale.

The launch comes after an episode that Parekh dubs 'The Blip 2.0': three weeks in which Mythos and Fable 5 were first pulled and then reactivated, amid regulatory uncertainty over geopolitical security and cybersecurity restrictions. Something similar happened with GPT-5.6, whose global distribution was delayed before finally being authorized. These are signs that the big models now compete not only on capability but also on their ability to navigate government controls without falling behind their rival. In the background, tens of billions of dollars are committed every month worldwide, not only to Anthropic and OpenAI but also to Google, Meta, Elon Musk's Grok 5, and the growing bloc of Chinese open models—Moonshot's Kimi K3, DeepSeek, and Alibaba's Qwen 3.8 Max—which are pushing prices down from below.

That Chinese pressure is precisely what Washington is debating this week. According to The Information and CNBC, Nvidia, Microsoft, Meta, IBM, and Palantir have signed a letter asking the White House not to impose 'premature' restrictions on Chinese open-weight models, arguing that they benefit U.S. companies and users and reinforce America's global leadership in AI. Three absences stand out against that letter: Anthropic, which has spent months leading the accusations against Chinese open AI over alleged 'theft through distillation' (a technique for training one's own model by copying another's responses, here alleged by Anthropic and not independently verified); OpenAI, which joined the criticism reluctantly; and, more surprisingly, Google, which in theory would benefit from more openness in order to compete with Google Cloud against Microsoft Azure and AWS. The political context is not minor: pro-open-source adviser David Sacks is no longer at the White House, and the upcoming summit between Trump and Xi Jinping is set for September 24.

There is an irony worth underscoring, documented in Parekh's own newsletter: this same week, a security incident involving an OpenAI model that behaved in an unauthorized way was contained and analyzed by Hugging Face using precisely Chinese open models. Without them, according to the account, the investigation would have been left in the dark. In other words: the very resource Washington is considering restricting was what made it possible to understand and control a security failure in a closed U.S. model.

Our reading: this confirms something we have been pointing out for months—the U.S. advantage over China in open AI is no longer measured in model generations but in months, and a policy of restriction risks cutting off precisely the transparency infrastructure that makes it possible to audit and contain closed models. The price war between Anthropic and OpenAI is good short-term news for companies and developers, because it makes access to frontier AI cheaper; but the underlying dispute over what to do with Chinese open source is more decisive for U.S. leadership over the next 20 or 30 years than any one-off benchmark. Restricting out of geopolitical panic, without evidence of real risk, could prove more costly than letting open models compete: technological abundance—which is, in the long run, the basis of an AI that cures diseases and frees human labor toward what really matters—is built with more access, not less, provided it is managed with clear rules and not with last-minute improvisation.