Anthropic's Opus 5 lands at half the cost of Fable, but skips the big leap

Anthropic just dropped Opus 5, and the pitch is blunt: near-Fable coding performance for roughly half the price. That’s the headline, but anyone expecting a seismic shift in agentic coding should temper their enthusiasm. The company’s own benchmarks tell a story of iteration, not revolution. Opus 5 nudges ahead of Opus 4.8 and OpenAI’s GPT-5.6-Sol across tests like Frontier-Bench and DeepSWE, and it trades blows with the much-hyped Fable model on coding tasks. But it’s not a clean win. The charts show a model that’s competitive, not dominant.

That cost-performance ratio is the real play here. If you’re running a dev team that’s been priced out of Fable for routine work, Opus 5 suddenly looks very interesting. It’s a practical offering for a market segment that’s tired of paying premium rates for marginal gains. Anthropic is betting that “just shy of Fable” is more than good enough when the alternative is burning through budget on a model whose full power you rarely need.

There’s a significant asterisk, though. Anthropic deliberately hamstrung Opus 5 on cybersecurity exploitation. The model can find vulnerabilities reasonably well, but the company says it’s “substantially behind Mythos 5” when it comes to actually using them. That’s a policy choice, not a technical limitation, and it means Opus 5 sidesteps some of the more controversial safety wrappers that dogged Fable — like the 30-day data review hold. Whether that’s a feature or a bug depends entirely on your threat model.