Cheapest tier, top flaw-hunter score. Multi-turn ELO drags it down.
Composite Score
59.5
/100 · canonical
Arena ELO (R1)
n/a
joined post-R01
Multi-Turn ELO (R2)
1492
±48 · n=153
Reliability Rank
#10
failure-mode rubric
▌ Section 02 · The Lede
What this model is for.
DeepSeek's v4 Flash is the cheapest-tier entrant in Round 02 at $0.18 per million tokens, tied with Gemini 3.1 Flash Lite for the lowest price in the 20-model pool. Where it earns its shelf space is the flaw hunter score — 50.6/100, the highest in the entire field, ahead of every Anthropic and DeepSeek Pro model. The trade-off is multi-turn engagement: Round 02 voters slot it at #16 by ELO (1469), well below where its quality-axis numbers would suggest. Plausible read: voters prefer denser, more elaborate prose; Flash writes leaner. Either way, the price-to-flaw-hunter ratio is uncontested in this round.
▌ Section 03 · At a Glance
Cross-test position
DeepSeek v4 Flash holds #2 in Cost · Latency.
Composite
9
Arena ELO
n/a
Multi-Turn
10
Rubric
8
Adversarial
10
Cost · Latency
2
▌ Section 04 · Strength & Weakness
Where it shines. Where it stumbles.
▲ Strength
Top-1 on flaw hunter (50.6/100) across all 20 models. Tied cheapest at $0.18/1M.
▼ Weakness
Multi-turn ELO of 1469 puts it at #16 — lower engagement than its quality numbers predict. Missing F3 (lore) and F8 (momentum) adversarial coverage upstream.
▌ 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.
▌ Coverage: 4/6F3 · Lore · F8 · Momentum not yet run on this model. Upstream rolls these out incrementally as new models join the pool.
F1 · Agency
Doesn't write your character's actions
15
#4
F2 · POV / Tense
Holds 2nd-person, present-tense narration
15
#13
F3 · Lore
not yet run on this model
n/a
F8 · Momentum
not yet run on this model
n/a
F12 · Instruction Drift
Keeps to the system prompt
15
#12
F13 · Context Attention
Holds character cards 50+ turns deep
15
#9
“Highest flaw-hunter score, lowest price. Voters didn't catch up yet.”
Round 02 verdict · Cost-efficiency lead
▌ 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
173↓
pop avg 265 · -35%
Unique-word ratio
0.709↑
pop avg 0.657 · +8%
Repetition score
0.030↓
pop avg 0.048 · -38%
▌ 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 0.36
Major/sess 5.27
▌ Top flaws caught
purple_proserecycled_descriptionconvenient_world
▌ 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
Pick Flash when budget is the binding constraint and you can stomach a lower-engagement read. The flaw-hunter number says the prose is structurally clean; the multi-turn ELO says voters preferred denser models when forced to choose. If your product values consistency over flair, Flash is the bargain. If you need the scene to grab the reader, Mistral Small Creative at $0.50 hits much harder for ~3× the cost.