Anthropic、画像理解能力で OpenAI に追いつく

On Tuesday, Anthropic released two new models — Claude Mythos 5 and Claude Fable 5. Under the hood, the two models are very similar. Both are variants of Claude Mythos Preview, the model Anthropic announced — but didn’t release publicly — two months ago. What differentiates them is how they’re being released.

The new version of Mythos, like the original, will only be available to handpicked organizations under Project Glasswing. These trusted partners will have relatively unfettered access.

Fable, in contrast, is available to the general public. But it comes with some significant restrictions. A new system will try to automatically detect when customers make dangerous requests (like hacking or designing a biological weapon) and automatically re-route them to the less powerful Claude Opus 4.8.

Mythos and Fable are a big step in coding abilities from previous models, a continuation of the trend of the last year. But there are other capabilities where models have made less progress.

For instance, frontier models have historically struggled to understand images, something I documented extensively in 2024 and 2025. Until recently, leading models struggled to perform simple tasks like reading an analog clock or counting the number of items in an image.

So as I was reading the official announcement post, this sentence caught my eye: “Fable 5 is the new state-of-the-art model for tasks involving vision.”

These tasks aren’t all that important in their own right, but they’re an interesting test case for a widely held assumption in the modern AI industry: that with enough data and computing power, frontier models will develop truly general intelligence. If new models are dramatically better at math and coding but only a little bit better at understanding images, that suggests that truly general intelligence might still be far away.

So I decided to evaluate the vision capabilities of Fable 5 and its main rivals, something I haven’t done since this August 2025 article about GPT-5.

I found that Claude Fable 5 and GPT-5.5 (though not Google’s Gemini models) can consistently solve many image-based problems that stumped last year’s top models. Fable 5 is arguably slightly better at these tasks than GPT-5.5, but it’s very close.

But these models haven’t made that much progress. GPT-5.5 and Claude Fable 5 continue to have geometric reasoning capabilities on par with young children. More fundamental architectural innovations may be needed to reach superhuman performance on this type of task.

Read more