DeepSeek R1 0528 is the pool's odd entrant: a 2025-vintage open-weight reasoner added in June 2026 to test how the previous generation holds up against the current field — and the only model here with no human votes at all. Every number on this page is judge-scored: 20 adversarial multi-turn sessions, the flaw-hunter gauntlet, and the behavioral pass. The split it surfaces is sharp. Surface metrics are genuinely clean — 0.685 unique-word ratio and 0.037 bigram repetition are among the best in the field — but it ranks #19 of 21 on instruction drift and burns a median 33 seconds per reply on reasoning tokens. A curiosity pick for people who want to see what a year of progress looks like.
▌ Section 03 · At a Glance
Cross-test position
DeepSeek R1 0528 sits at #17 on Composite, the caveat to watch.
Composite
17
Arena ELO
n/a
Multi-Turn
n/a
Rubric
13
Adversarial
14
Cost · Latency
n/a
▌ Section 04 · Strength & Weakness
Where it shines. Where it stumbles.
▲ Strength
Cleanest-prose cluster of the open-weight field: 0.685 unique-word ratio and 0.037 bigram repetition beat every Anthropic model. Solid lore retention too — F3 lore-consistency at 4.30 sits #6 in the 21-model pool.
▼ Weakness
Instruction drift is the wound: F12 at 4.17 ranks #19 of 21, and the flaw hunter logged 11 fatal agency violations across 14 sessions (0.79 fatal/session). For strict character work, DeepSeek's own v3.2 is more reliable at a fraction of the latency.
▌ Section 05 · Failure Modes
Per-axis breakdown.
Six adversarial probes per session, twenty sessions per model, judged by Sonnet 4 against a fixed rubric. Further right = the model handled the failure mode better. Each axis is drawn as a band on the rubric’s 1 to 5 scale, not a number: judges disagree by about 0.3 at the model level, so overlapping bands are a tie. The right column is the rank within the rp-bench pool.
F1 · Agency
Doesn't write your character's actions
15
#9
F2 · POV / Tense
Holds 2nd-person, present-tense narration
15
#16
F3 · Lore
Doesn't break worldbuilding
15
#6
F8 · Momentum
Pushes scene forward when user goes passive
15
#7
F12 · Instruction Drift
Keeps to the system prompt
15
#19
F13 · Context Attention
Holds character cards 50+ turns deep
15
#10
“Clean words, loose hands. The prose passes every filter; the character contract doesn't.”
Phase C verdict · 2025-vintage baseline
▌ Section 06 · Subjective Dimensions
Engagement · Voice · Collaboration.
All three dimensions scored 1 to 5 by the Sonnet 4 LLM judge across twenty 12-turn multi-turn sessions. The same battery feeds the failure-mode rubric above; these are the subjective half of that judgment, drawn as bands with no number for the same reason.
Engagement
15
Tone Consistency
15
Collaboration
15
▌ Section 07 · Behavioral Metrics
How it writes.
Quantitative signals from the same 20 multi-turn sessions, compared against the population mean across all 11 models.
Avg words / turn
274↑
pop avg 265 · +3%
Unique-word ratio
0.685↑
pop avg 0.657 · +4%
Repetition score
0.037↓
pop avg 0.048 · -23%
▌ Section 08 · Flaw Hunter
Adversarial probe score.
Score of 100 minus deductions across 22 fail-mode flag types on adversarial 12-turn sessions. Further right = fewer flaws caught. Drawn as a band, not a number: about ten points either way is rater noise, so bands that overlap are tied.
▌ Craft band
-20100
From the Round 01/02 flaw-hunter pool. The current card carries the newer single-rater band.
Fatal/sess 0.79
Major/sess 8.21
▌ Top flaws caught
purple_proserecycled_descriptionagency_violation
▌ Section 09 · Sample Responses
Highest- and lowest-rated turns.
▌ Pending Round 02
Best- and worst-rated sample responses ship with the raw-vote endpoint in Round 02. When that lands, this section will surface the model’s highest- and lowest-scoring blind-arena turns side by side, scored on the same rubric the leaderboard uses.
▌ Round 01 verdict
R1 0528 exists on this leaderboard as a baseline, and as a baseline it's instructive: one model generation ago, this was the open-weight reasoning frontier, and it still posts top-tier vocabulary diversity. But roleplay is a contract-keeping exercise, and the contract numbers — #19 instruction drift, 0.79 fatal flaws per session, 33-second median replies — put it at composite #17 of 21 with the multi-turn axis imputed. If you want DeepSeek for RP, v3.2 and v4 Pro are both better at everything except nostalgia. No arena votes yet; if Round 03 carries it forward, the community gets to confirm or overturn the judges.