Inflection-2.5 takes on GPT-4 with 40% of the compute

Pi, the friendly personal AI, just decided it wants to compete with the heavyweights.

Inflection announced Inflection-2.5 to power its Pi assistant. The company claims the new model approaches GPT-4-level performance while using just 40% of the training compute.

Cutting the compute bill

The headline claim here is sheer efficiency. Inflection argues it reached frontier capabilities without the massive computational appetite of earlier giants.

Still, onlookers like developer Simon Willison quickly raised an eyebrow, noting that the exact compute used to train GPT-4 has never been public knowledge, making that 40% comparison tricky to verify.

Regardless of the exact math, the competitive field has shifted dramatically. In just a matter of weeks, the once-unreachable GPT-4 benchmark suddenly faces a crowd of peers, including Claude 3 Opus, Gemini 1.5, Mistral Large, and now Inflection-2.5.

Why it matters

We think this is a necessary leap for Pi. Until now, people saw it mostly as an empathetic sounding board rather than a workhorse. Giving it near-frontier smarts could turn it into something genuinely useful for day-to-day productivity without losing its conversational warmth.

For the broader industry, the GPT-4 moat has basically evaporated. The race is no longer about matching OpenAI's raw performance from last year, but doing it faster and with far fewer resources. However, without transparent compute figures, we'd take Inflection's exact efficiency claims with a grain of salt until real-world testing proves it.