Four models read the same bills

Every score on this site is an AI judgment. Four models scored the same 443 bills against the same 24 issues, independently. Here is what each one is, what it cost, and how often they agreed.

The lineup

Claude Opus 5.5
Anthropic · A Claude Code agent session
Reads each bill and writes −2 to +2 scores directly, each with a reason and a confidence level.
Bills with a score
362 of 443
Issue scores
564 1.6 per bill
Helps / hurts
71% / 29%
Cost
≈ $4.92
Estimate. No usage was logged, so input is sized from Gemma’s real prompt per bill and output from what the model wrote. An agent session’s own overhead would add to this. 1,075,037 input and 30,979 output tokens, at $4 in and $20 out per million.
TypeSafe jev-1.13.0
TypeSafe · TypeSafe API
Answers, as probabilities, whether a bill affects each issue and how much. An issue counts when the chance it’s affected is at least 0.6 and the effect at least 0.5; effects of 1.75 or more score ±2.
Bills with a score
289 of 443
Issue scores
554 1.9 per bill
Helps / hurts
59% / 41%
Cost
$0.14
Logged input tokens times the listed price. Output is free on this model. 3,287,102 input and 0 output tokens, at $0.042 in and $0 out per million.
Gemma 4 31B
Google, open weights · OpenRouter, mostly served by DeepInfra
Writes −2 to +2 scores directly into a strict JSON schema, each with a reason and a confidence level.
Bills with a score
372 of 443
Issue scores
796 2.1 per bill
Helps / hurts
80% / 20%
Cost
$0.16
Actual cost reported by OpenRouter for the whole run. 1,291,015 input and 66,893 output tokens.
GPT-6 Astra
OpenAI · An agent annotation run
Writes −2 to +2 scores directly, each with a reason.
Bills with a score
300 of 443
Issue scores
536 1.8 per bill
Helps / hurts
80% / 20%
Cost
≈ $12.90
Estimate. No usage was logged, so input is sized from Gemma’s real prompt per bill and output from what the model wrote. An agent session’s own overhead would add to this. 1,075,037 input and 43,038 output tokens, at $10 in and $50 out per million.

All four runs together cost about $18.12, most of it the two frontier models.

Which issues they flagged

Each count is a bill and issue pair, like “H.R. 82 affects seniors”, that one or more models flagged. Of 1,150 pairs flagged by anyone, 222 were flagged by all four; 497 were flagged by only one.

Every overlap as a table

Claude Opus 5.5	TypeSafe jev-1.13.0	Gemma 4 31B	GPT-6 Astra	Pairs

Head to head

Overlap is the share of pairs either model flagged that both flagged. Same direction is, of the pairs both flagged, how often they agreed the bill helps or hurts.

Models	Both flagged	Overlap	Same direction
Claude Opus 5.5 and TypeSafe jev-1.13.0	339	44%	93%
Claude Opus 5.5 and Gemma 4 31B	438	48%	92%
Claude Opus 5.5 and GPT-6 Astra	321	41%	99%
TypeSafe jev-1.13.0 and Gemma 4 31B	386	40%	97%
TypeSafe jev-1.13.0 and GPT-6 Astra	295	37%	99%
Gemma 4 31B and GPT-6 Astra	390	41%	98%

The board and member pages currently rank with Claude Opus 5.5 and TypeSafe scores (averaged where both scored a bill). Gemma and Astra are compared here but not yet counted. How scoring works.

The lineup

Claude Opus 5.5

Anthropic · A Claude Code agent session

Reads each bill and writes −2 to +2 scores directly, each with a reason and a confidence level.

Bills with a score: 362 of 443
Issue scores: 564 1.6 per bill
Helps / hurts: 71% / 29%
Cost: ≈ $4.92

Estimate. No usage was logged, so input is sized from Gemma’s real prompt per bill and output from what the model wrote. An agent session’s own overhead would add to this. 1,075,037 input and 30,979 output tokens, at $4 in and $20 out per million.

TypeSafe jev-1.13.0

TypeSafe · TypeSafe API

Answers, as probabilities, whether a bill affects each issue and how much. An issue counts when the chance it’s affected is at least 0.6 and the effect at least 0.5; effects of 1.75 or more score ±2.

Bills with a score: 289 of 443
Issue scores: 554 1.9 per bill
Helps / hurts: 59% / 41%
Cost: $0.14

Logged input tokens times the listed price. Output is free on this model. 3,287,102 input and 0 output tokens, at $0.042 in and $0 out per million.

Gemma 4 31B

Google, open weights · OpenRouter, mostly served by DeepInfra

Writes −2 to +2 scores directly into a strict JSON schema, each with a reason and a confidence level.

Bills with a score: 372 of 443
Issue scores: 796 2.1 per bill
Helps / hurts: 80% / 20%
Cost: $0.16

Actual cost reported by OpenRouter for the whole run. 1,291,015 input and 66,893 output tokens.

GPT-6 Astra

OpenAI · An agent annotation run

Writes −2 to +2 scores directly, each with a reason.

Bills with a score: 300 of 443
Issue scores: 536 1.8 per bill
Helps / hurts: 80% / 20%
Cost: ≈ $12.90

Estimate. No usage was logged, so input is sized from Gemma’s real prompt per bill and output from what the model wrote. An agent session’s own overhead would add to this. 1,075,037 input and 43,038 output tokens, at $10 in and $50 out per million.

All four runs together cost about $18.12, most of it the two frontier models.

Which issues they flagged

Every overlap as a table

Claude Opus 5.5	TypeSafe jev-1.13.0	Gemma 4 31B	GPT-6 Astra	Pairs

Head to head

Overlap is the share of pairs either model flagged that both flagged. Same direction is, of the pairs both flagged, how often they agreed the bill helps or hurts.

Models

Both flagged

Overlap

Same direction

Claude Opus 5.5 and TypeSafe jev-1.13.0

339

44%

93%

Claude Opus 5.5 and Gemma 4 31B

438

48%

92%

Claude Opus 5.5 and GPT-6 Astra

321

41%

99%

TypeSafe jev-1.13.0 and Gemma 4 31B

386

40%

97%

TypeSafe jev-1.13.0 and GPT-6 Astra

295

37%

99%

Gemma 4 31B and GPT-6 Astra

390

41%

98%

The board and member pages currently rank with Claude Opus 5.5 and TypeSafe scores (averaged where both scored a bill). Gemma and Astra are compared here but not yet counted. How scoring works.