Matched or beat all three opponents

236 of 240 head-to-head judgments. The percentage combines 30 wins and 206 functional ties. It is not a win rate.

Interruption sensitivity: excluding the three hosted-delivery-failure judgments (Q59, Q61, Q75 versus Claude), the record is 233 of 237, or 98.31%. The headline does not depend on them.

What changed

Qwen 3.8 27B ran locally on the same operator-controlled workstation and was compared with DeepSeek, Claude Opus 5, and Qwen 3.5. It led the four-model field while keeping the benchmark matters and prompts off outside AI infrastructure during local generation.

See the four systems

See the complete result Inspect all 80 matters

New leader · 240 judgments involving Qwen 3.8

Equal or ahead in 236

Both bars use the same 0 to 240 scale. Select either bar to inspect the underlying comparisons.

Qwen 3.8 matched or beat30 wins plus 206 functional ties across all three opponents236judgments Opponent winsFour judgments where another model finished ahead4judgments

Exact value: 98.33%, displayed as 98%.

How 98% relates to the earlier 96%

Original headline result: DeepSeek versus Claude Opus 5Equal or better on 77 of 80 matters, displayed as 96%.

New August 15 result: Qwen 3.8 leads the four-model fieldEqual or better in 236 of 240 judgments, displayed as 98%.

Different denominators: matters in one head-to-head comparison versus judgments across three opponents.

New leaderQwen 3.8 27B30 wins, 4 losses, 206 ties

Benchmark scope4 models · 80 mattersSix pairings per matter

Blind judgments480 totalSame published judge prompt

ExecutionLocal, one GPUNo tools or follow-up repair

Head-to-head record

Qwen 3.8 led every comparison

The new model was never worse than either local predecessor and finished ahead of Claude on the full delivered-system record.

Decisive wins by opponent

Each row covers the same 80 legal matters.

0 to 80 scale

vs DeepSeek: 5 wins75 ties, 0 losses5wins vs Qwen 3.5: 19 wins61 ties, 0 losses19wins vs Claude: 6 wins70 ties, 4 losses6wins

Evaluation sequence

From closed record to blind verdict

The expanded release uses the same method, the same 80 matters, and the same published judge prompt.

Generate

Each of four models answered the same fictional matter and supplied law without tools or browsing.

Pair

All six model pairings were created for every matter, producing 480 comparisons.

Judge blind

A rival-lab judge saw concealed A/B labels and applied the same material-outcome instrument.

Publish

Every answer, judgment, assignment, prompt, and integrity record is available for audit.

Use the exact judge prompt Read the methodology

Audit path

Move from claim to source

Start with the standings, open any matter, then inspect the outputs and blind judgment behind it.