Is Phi 4 a good deal?
Whether anything in this catalogue beats Phi 4 on both quality and price, and what you give up if it does. A computation on the current catalogue, not an opinion.
12 undominated of 145 · Oct 6, 2026
As of Oct 6, 2026, Phi 4 is dominated for Balanced on LMArena. gpt-oss-120b scores 148.7 higher and costs 26% less, with a covering envelope. 12 of 145 rated, priced standard models are undominated.
Inspect model evidence Compare differences & requirements
gpt-oss-120b is both better and cheaper than Phi 4: 148.7 points higher on LMArena and 26% less per million tokens, $0.023 cheaper at this mix.
LMArena Elo · higher is better
Scale starts at 1040 Elo
1216.7
Effective $/M · Balanced · lower is better
$0.088/M
LMArena Elo · higher is better
Scale starts at 1040 Elo
1365.4
Effective $/M · Balanced · lower is better
$0.065/M
Phi 4 takes text, returns up to 14,745 tokens from a 16,384-token context, and is listed by 1 seller. Open weights, so it can also be self-hosted.
Compared against 145 rated, priced models on this lens: 5 models dominate it and give up nothing, none does so with a trade-off. Phi 4 scores 1216.7 at $0.088 per million tokens for this mix.
Envelope-safe replacements
Each row scores at least as high, costs no more, and covers this model’s context, output, modalities, tools, and reasoning. A cheaper narrower model is not listed here.
| Model | LMArena | Effective $/M | You save |
|---|---|---|---|
| gpt-oss-120b | 1365.4 +148.7 | $0.065/M | 26% |
| Gemma 3 12B | 1334.2 +117.5 | $0.075/M | 14% |
| Gemma 3 4B | 1290.8 +74.1 | $0.063/M | 29% |
| gpt-oss-20b | 1287.7 +71.0 | $0.036/M | 59% |
| Mistral Small 3 | 1233.6 +16.9 | $0.058/M | 34% |
The link keeps the model and mix, using the current catalogue. Monthly spend and switching cost stay in this browser tab and are left out of the link.
Saved decision references
Save the model, workload, catalogue date and capability-preservation rule in this browser. Spend, switching cost and bill contents are not saved. No account or notifications.
Constraint: replacements must preserve the model’s capabilities; any losses remain named trade-offs.
Historical decisions cannot be fully replayed from saved references: past prices, scores and capability evidence are not stored.