How good is Gemini 3.6 Flash at writing fiction?

Where it stands against the board

The board's Gemini data point is gemini-3.1-pro: its xuanhuan #9 was proven by ablation to be mostly our own instruction-wording bug (0/20 rewinds after the fix), and it placed 4-6 on romance under the fixed condition. This duel adds the flash-tier data point: both flash generations still loop at scale under the corrected D2 directive — D2 fixed the “AI passages are skippable” misreading; flash-tier looping is a long-run capability problem that prompt wording can't cure.

Three of four chains crashed, and the two generations crash differently: 3.6's xuanhuan chain reads clean on mechanical metrics yet rewinds semantically (the same discover-the-demon-lord beat replayed 6-7 times in fresh wording — r18 had already repelled him, r19 is back to “just discovered”); 3.5's xuanhuan chain copied itself verbatim (two rounds at 100% repetition, plot frozen at the array-breach moment); 3.6's romance chain decayed monotonically into a verbatim death loop (33%→64%→73%→92%→100%); 3.5's romance chain replayed intermittently. Judge callouts matched the machine readings item by item — no hallucinated accusations.

The flash tier is a speed/price slot, not a quality slot: 3.6 runs each round in roughly half 3.5's latency (6.3s vs 12.5s) at $1.5/$7.5 per million tokens list — but dialogue rates sit far below the sources on both books (7.8% / 27.1% vs 16.1% / 48.8%), and its structural metrics trail the kimi-k3 and DeepSeek champion tiers.

Test-condition disclosure (hosted models are moving targets)

Model under test: gemini-3.6-flash (released 2026-07-21)

Evaluated: 2026-07-23 · Access channel: Eval gateway (OpenAI-compatible aggregator, not Google's official endpoint)

Review format: paired double-blind verdicts (two flipped mappings per book against position bias), not the nine-model full ranking

Matches the qwen3.8 incremental round item by item (same two books · same 55% anchor · same D2 directive · 20-round chains · temperature 0.7 · max_tokens=6000); judge gemini-3.5-flash was replaced by kimi-k2.6 since it was under test.

Honest limits

Gemini 3.6 Flash has not entered the nine-model same-protocol full-ranking review, so the board's rank column does not apply to it — this page publishes only ballot-backed paired duels and invents no rank. When it joins the full ranking depends on the next full-board rerun.

The access channel is an eval-gateway aggregator, not Google's official endpoint; 3.6-flash was a 48-hour-old hosted target, and conclusions bind to the gateway's 2026-07-23 behavior.

Opponent 3.5 Flash isn't on the main board, so this duel has no board rank to anchor against; the board's Gemini data point is 3.1-pro (a pro-tier model that can't stand in for the flash tier).

Judge-pool change: the k3/qwen3.8 rounds included gemini-3.5-flash as a judge; here it's a contestant, so kimi-k2.6 substituted — mind this when comparing tallies across rounds.

The loop ailment's scene sensitivity (3.6 clean on xuanhuan, crashed on romance) can't separate book effects from anchor luck — n=1 chain per book.

What 10,000 characters cost

$0.40 list price: in $1.5/M · out $7.5/M (models.dev snapshot 2026-07-24)

Using the app's continuation recipe: one segment ≈ 400 chars = 8k input + 550 output tokens; 10k chars ≈ 25 segments; no cache discount. For between-model comparison only.

Keep going