Deciding which AI model to use.

A model is on the 3D Pareto frontier if no other model beats it on score, cost, and time. The switch fades everything else and joins the frontier with a dotted line. I use that line as a shortlist: start cheap, then move up if the model struggles.

The v4.3 rebuild moved nearly every score and recalculated task costs, so the old frontier no longer applies. GPT-6 Luna now owns the low-cost end. At max it matches GPT-5.6 Luna’s 37.3 for about $0.07 per task instead of $0.18. GPT-6 Sol scores about the same as GPT-5.6 Sol at every effort setting for roughly half the cost. In the middle, the Muse Spark 1.3 Contributor pricing scenarios reuse scores of 45 to 48 at a derived cost of about $0.06 per task. Contributor max eligibility remains unverified; that point is hypothetical pricing, not a confirmed available tier. Contributor usage permits model training.

Claude Opus 5.5 at max now has the highest plotted score at 57.6, for about $5.98 and 13.3 minutes per task. At xhigh it scores 56.0 for about $3.46 in 8.0 minutes. At high it scores 53.6 for about $1.82, just above Claude Fable 5.1 at max, which costs about $7.63 and takes 12.0 minutes. GPT-6 Astra at xhigh reaches 52.4 for about $2.31 in 5.1 minutes. Claude Sonnet 5.5 at max matches Opus 5.5 at xhigh with 56.0, but costs about $7.60, more than twice as much. Artificial Analysis has not published Sonnet 5.5’s time per task yet, so it is not plotted. Models without current cost or timing data remain in the full score table, so a missing point does not mean a missing result.

CursorBench ranks Opus 5.5 first: 57.8% at max for $13.43 per task, and 56.0% at both xhigh and high, where high costs $3.97 instead of $6.98. Sonnet 5.5 at xhigh scores 53.1% for $3.88, above every Fable 5.1 setting. Fable 5.1 at max tops out at 51.8% for $17.28.

On DeepSWE, GPT-5.6 Luna at max reaches 67.2 for about $0.54 per task. Gemini 3.8 Flash at high reaches 73.8 for about $2.36. GPT-6 Astra at xhigh leads at 74.1 for about $4.10. DeepSWE has no results for Opus 5.5, Sonnet 5.5, GPT-6 Sol, or GPT-6 Luna yet.