GPT-6 Astra uses about 21K tokens per task versus roughly 64K for Fable 5.1, post says
A post argues Astra's real threat to Anthropic is token efficiency rather than benchmark scores, and that this gap could explain why Claude Code limits run out quickly.
OpenAI could be creating a serious problem for Anthropic. And it might have less to do with benchmarks than people think. The real advantage could be efficiency. GPT-6 Astra reportedly uses around 21K tokens to complete a task, while Fable 5.1 can use roughly 64K. That’s a massive difference. If both can get to a similar result, Astra is doing it with far fewer tokens. And that could explain why some tasks that take hours with Fable can be completed much faster with Astra. Even the Claude Code limits start to make more sense. Maybe you’re not actually using Claude too much. Maybe Fable is just burning through tokens much faster. If Astra can consistently deliver similar results while using a fraction of the tokens, that’s more than a benchmark win. It’s a major cost and efficiency advantage. And that’s something Anthropic can’t afford to ignore.