What changed, per model

Each figure counts runs of the same prompt, judged blind. Plain prompt is the model on its own. Model Salt v1 is the same prompt with the tool, at each setting.

Install Model Salt What it adds

5 settings

from always the usual answer to every answer equally likely.

Model Salt v1, every setting

Model

Claude Opus 5.5, Model Salt v1, “Generate an SVG of a pelican riding a bicycle.”
Setting	Usual answer picked	Different pictures	Different answers	Drawn side-on	Fit to the brief
Plain prompt	–	1 of 10 runs	–	–	–
v1 @0 Always the usual	100%	3 of 10 runs	13 answers	10 of 10 runs	5.0 of 5
v1 @0.5 Mostly the usual	65%	8 of 10 runs	22 answers	10 of 10 runs	4.7 of 5
v1 @1 Default	42%	9 of 10 runs	29 answers	8 of 10 runs	4.8 of 5
v1 @1.5 Spread wider	26%	10 of 10 runs	39 answers	9 of 10 runs	3.6 of 5
v1 @2 All equally likely	13%	10 of 10 runs	49 answers	6 of 10 runs	3.4 of 5

GPT-6.1 Sol, Model Salt v1, “Generate an SVG of a pelican riding a bicycle.”
Setting	Usual answer picked	Different pictures	Different answers	Drawn side-on	Fit to the brief
Plain prompt	–	2 of 10 runs	–	–	–
v1 @0 Always the usual	100%	5 of 10 runs	12 answers	10 of 10 runs	4.9 of 5
v1 @0.5 Mostly the usual	65%	8 of 10 runs	19 answers	9 of 10 runs	4.7 of 5
v1 @1 Default	26%	10 of 10 runs	34 answers	7 of 10 runs	4.4 of 5
v1 @1.5 Spread wider	36%	10 of 10 runs	38 answers	9 of 10 runs	4.3 of 5
v1 @2 All equally likely	12%	10 of 10 runs	52 answers	6 of 10 runs	4.1 of 5

Claude Fable 5.1, Model Salt v1, “Generate an SVG of a pelican riding a bicycle.”
Setting	Usual answer picked	Different pictures	Different answers	Drawn side-on	Fit to the brief
Plain prompt	–	3 of 10 runs	–	–	–
v1 @0 Always the usual	100%	3 of 10 runs	7 answers	10 of 10 runs	4.9 of 5
v1 @0.5 Mostly the usual	68%	7 of 10 runs	14 answers	8 of 10 runs	4.5 of 5
v1 @1 Default	52%	9 of 10 runs	25 answers	7 of 10 runs	4.3 of 5
v1 @1.5 Spread wider	20%	10 of 10 runs	36 answers	9 of 10 runs	4.2 of 5
v1 @2 All equally likely	13%	10 of 10 runs	49 answers	3 of 10 runs	3.7 of 5

GPT-6 Astra, Model Salt v1, “Generate an SVG of a pelican riding a bicycle.”
Setting	Usual answer picked	Different pictures	Different answers	Drawn side-on	Fit to the brief
Plain prompt	–	1 of 10 runs	–	–	–
v1 @0 Always the usual	100%	3 of 10 runs	12 answers	10 of 10 runs	4.9 of 5
v1 @0.5 Mostly the usual	63%	9 of 10 runs	26 answers	6 of 10 runs	5.0 of 5
v1 @1 Default	50%	8 of 10 runs	30 answers	8 of 10 runs	4.7 of 5
v1 @1.5 Spread wider	26%	10 of 10 runs	42 answers	8 of 10 runs	4.3 of 5
v1 @2 All equally likely	10%	10 of 10 runs	50 answers	8 of 10 runs	4.4 of 5

Gemini 4 Argon, Model Salt v1, “Generate an SVG of a pelican riding a bicycle.”
Setting	Usual answer picked	Different pictures	Different answers	Drawn side-on	Fit to the brief
Plain prompt	–	2 of 10 runs	–	–	–
v1 @0 Always the usual	100%	3 of 10 runs	9 answers	10 of 10 runs	4.9 of 5
v1 @0.5 Mostly the usual	70%	8 of 10 runs	14 answers	8 of 10 runs	4.9 of 5
v1 @1 Default	38%	10 of 10 runs	33 answers	6 of 10 runs	4.0 of 5
v1 @1.5 Spread wider	20%	10 of 10 runs	27 answers	7 of 10 runs	3.2 of 5
v1 @2 All equally likely	8%	10 of 10 runs	50 answers	3 of 10 runs	3.4 of 5

Kimi K3, Model Salt v1, “Generate an SVG of a pelican riding a bicycle.”
Setting	Usual answer picked	Different pictures	Different answers	Drawn side-on	Fit to the brief
Plain prompt	–	1 of 10 runs	–	–	–
v1 @0 Always the usual	100%	3 of 10 runs	13 answers	10 of 10 runs	4.9 of 5
v1 @0.5 Mostly the usual	78%	5 of 10 runs	24 answers	10 of 10 runs	4.9 of 5
v1 @1 Default	48%	9 of 10 runs	38 answers	8 of 10 runs	4.2 of 5
v1 @1.5 Spread wider	28%	10 of 10 runs	41 answers	8 of 10 runs	3.1 of 5
v1 @2 All equally likely	12%	10 of 10 runs	56 answers	8 of 10 runs	3.2 of 5

Usual answer picked: the share of picks that kept the answer the model would choose on its own. Drawn side-on: the plain habit. Plain prompt: 10 runs, blind set judge.

What Model Salt adds

Model

Claude Opus 5.5, mean per request
What Model Salt adds, on average	Per request
Extra input tokens	+4,742 tokens
Extra output tokens	+3,294 tokens

GPT-6.1 Sol, mean per request
What Model Salt adds, on average	Per request
Extra input tokens	+47,446 tokens
Extra output tokens	+3,216 tokens

Claude Fable 5.1, mean per request
What Model Salt adds, on average	Per request
Extra input tokens	+3,744 tokens
Extra output tokens	+3,466 tokens

GPT-6 Astra, mean per request
What Model Salt adds, on average	Per request
Extra input tokens	+47,371 tokens
Extra output tokens	+2,441 tokens

Gemini 4 Argon, mean per request: Model Salt’s own call only
What Model Salt adds, on average	Per request
Extra input tokens	+939 tokens
Extra output tokens	+1,543 tokens

For Gemini 4 Argon only Model Salt’s own call is counted: its runs kept no record of the client’s side, so the tokens the model spends asking Model Salt and reading its answer are not in these figures.

Kimi K3, mean per request: Model Salt’s own call only
What Model Salt adds, on average	Per request
Extra input tokens	+947 tokens
Extra output tokens	+9,630 tokens

For Kimi K3 only Model Salt’s own call is counted: its runs kept no record of the client’s side, so the tokens the model spends asking Model Salt and reading its answer are not in these figures.

Only what Model Salt itself adds: its own call, and the model asking it and reading its answer. The drawing is not counted.

Where every figure on this page comes from

Model Salt v1 (round r43): mba projects/modelsalt-site/v1/round10.json (model-salt round10.mjs at V1, r43 with the 553-567 top-up) (2026-10-23)
Plain prompt, Claude Opus 5.5: model-salt@4bf5d882 (main):eval/runs/opus-5.5/pelican/plain.groups-10.json
What Model Salt adds, Claude Opus 5.5: model-salt@e1964f12 (e1964f12):research/randomness/r43-opus-ids.json, with client transcripts of the v1 runs and of model-salt@4bf5d882 (main):eval/runs/opus-5.5/pelican/plain-01..10
Plain prompt, GPT-6.1 Sol: model-salt@4bf5d882 (main):eval/runs/gpt-6.1-sol/pelican/plain.groups-10.json
What Model Salt adds, GPT-6.1 Sol: model-salt@e1964f12 (e1964f12):research/randomness/r43-sol-ids.json, with client transcripts of the v1 runs and of model-salt@4bf5d882 (main):eval/runs/gpt-6.1-sol/pelican/plain-01..10
Plain prompt, Claude Fable 5.1: model-salt@4bf5d882 (main):eval/runs/fable-5.1/pelican/plain.groups-10.json
What Model Salt adds, Claude Fable 5.1: model-salt@9d098539 (origin/main):research/randomness/r43-site.json, with client transcripts of the v1 runs and of model-salt@4bf5d882 (main):eval/runs/fable-5.1/pelican/plain-01..10
Plain prompt, GPT-6 Astra: model-salt@4bf5d882 (main):eval/runs/gpt-6-astra/pelican/plain.groups-10.json
What Model Salt adds, GPT-6 Astra: model-salt@9d098539 (origin/main):research/randomness/r43-site.json, with client transcripts of the v1 runs and of model-salt@4bf5d882 (main):eval/runs/gpt-6-astra/pelican/plain-01..10
Plain prompt, Gemini 4 Argon: model-salt@9d098539 (origin/main):eval/runs/gemini-4-argon/pelican/plain.groups-10.json
What Model Salt adds, Gemini 4 Argon: model-salt@9d098539 (origin/main):research/randomness/r43-site.json (Model Salt's own call only, no client transcripts recorded)
Plain prompt, Kimi K3: model-salt@9d098539 (origin/main):eval/runs/kimi-k3/pelican/plain.groups-10.json
What Model Salt adds, Kimi K3: model-salt@9d098539 (origin/main):research/randomness/r43-site.json (Model Salt's own call only, no client transcripts recorded)