How good is Grok 4.6 at writing fiction?

Where it stands against the board

The headline is verified generational repair: both board-documented killers of grok-4.5 — the mid-window verbatim loop (rounds 9-11, an identical 271 characters ×3) and romance first-person density at half the source's (5.2‰ vs gold 9.0) — vanished in 4.6's same-protocol retest: cross-round 12-gram peaks of just 2.8% (fantasy r14) / 2.1% (romance r13), first-person back at 11.0‰, level with the champion; round-by-round reads show linear plot on both books, with the romance chain finally leaving the incident day. The combined 24:0 with all twelve stable verdicts at high confidence beats the previous record, Opus 5's single-book 12:0.

Against the incumbents, read it two ways: same tier as Flash on fantasy (8:4 raw; the series' heaviest position bias voided eight ballots, leaving a 4:0 stable component all for 4.6), one tier short of Pro on romance (2:10, strong signal). It becomes the third genre-specialist specimen in the “no universal style model” picture — Flash rules fantasy, Pro rules romance, 4.6 is fantasy-strong romance-weak. Half-width-quote artifacts: zero on both books (the K3/GLM/Opus-4.8 family of stubborn habits never appears in the Grok line), and its low dialogue rate (8.9% vs gold 16.1% on fantasy) is real behavior, not a detector artifact.

The new ailment, outline-itis, carries a product-grade paradox: 4.6 is the only chain on the table that held the 80-220-character length spec every round, and the same 150-character rounds read as “plain-spoken, high-action-density, a perfect fit” to fantasy judges and as “anemic, the long-form texture lost” to romance judges. Spec compliance and genre expectation point in opposite directions — pick models per book, not per leaderboard. The residual habit is image-level reuse (“rust-iron tang” nine times, a well-water motif across five rounds), downgraded from disease to fondness — with the same image praised by one judge as the original's spirit and filed by two others under repetition.

Cost and latency: list price $2/M input, $6/M output (cached input metered at $0.5; the double-priced faster tier unused), and a metered bill of $2.0955 (≈¥15) for all 40 rounds — a list price 3-10× the domestic models' yet the cheapest run in the series (K3's round: ¥12.21; Opus 5's: $12.54), precisely because it writes 150-character rounds and reasoning is metered separately. Reasoning takes 95.5-95.6% of output tokens, the highest share recorded (K3 held 81-86%), at 39-44 seconds per round. And the “statistically closer ≠ reads closer” family gains a reversed entry: predecessor 4.5's romance distance of 0.259 was the closest in the field — better than the champion's — yet it lost 12:0; its disease lived in verbatim loops, which an eight-metric average cannot see.

Test-condition disclosure (hosted models are moving targets)

Model under test: grok-4.6 (released 2026-08-12)

Evaluated: 2026-08-13 · Access channel: xAI official API (reasoning on by default)

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

Matches the k3/opus5/gemini36 incremental rounds item by item (same two books · same 55% anchor · same D2 directive · 20-round chains · temperature 0.7 · max_tokens=2800; xAI meters reasoning separately, so it doesn't eat the prose budget); judge deepseek-v4-pro recused as an opponent this round, with gemini-3.5-flash returning per precedent.

Honest limits

Grok 4.6 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.

Paired blind duels ran only against the predecessor and the two champions' archived snapshots — it never entered the nine-model full ranking, so its record can't be hard-compared against board ranks (different review formats: full ranking there, paired verdicts here).

The two fantasy opponent chains carry the old-directive condition (hedged by the DeepSeek-family ablation's zero regressions, still strictly a mixed-in variable); all opponents are 2026-07-16 snapshots, not same-day re-runs — and the romance champion's same-named live model was hot-swapped to 0813 on 08-13 (see the DeepSeek V4 Pro 0813 duel page).

Single book, single chain, single anchor (n=1 per book); the outline-itis verdict's anchor sensitivity is untested; grok-4.6 was a ~24-hour-old hosted target, and conclusions bind to the official API's 2026-08-13 behavior.

Pricing: grok-4.6 isn't in the models.dev list-price snapshot (captured 2026-08-12), so this page's cost section is honestly absent; list price and the per-round metered bill are in the positioning notes and the write-up.

Keep going