Amit Gupta

[Amit Gupta @amitgupta2108]

190 GPQA Diamond questions. Our own runs; half-informed model-rating tuning. Held-out scores were within ~2 points of the full-set charts.

Routing also lost: Sonnet 5.5 beat Artemis Quality-first on accuracy and cost.

Results: https://t.co/dx60c8W3lf

[Amit Gupta @amitgupta2108]

“Use the strongest model” should earn its place.

Our GPQA run:
GPT-5.5: 92.6%, $5,740/100k
NuRoute Artemis Balanced: 88.9%, $137/100k
42× the cost for +3.7 accuracy points.
Half the questions informed tuning.

Would that gain justify the bill in your product?

[Amit Gupta @amitgupta2108]

Using the same LLM for every task is expensive engineering.
We built Artemis Task at https://t.co/BoSEMUBcv1 to change that.
On 426 held-out business prompts, our Quality-first router achieved:
→ 72.1% lower inference cost
→ 99.1% of GPT-5.5's quality
Model selection should happen per request, not per application.
https://t.co/CGptwCUh1N