Grok 4.5's case file, retried under the original protocol

The ugliest file on our nine-model continuation board belongs to Grok 4.5: #8 on the fantasy epic, a unanimous #9 on the palace novel, bottom of both genres. Three counts on the charge sheet: mid-chain it looped the same two passages verbatim, three times over; its palace chain spent twenty rounds unable to leave the afternoon of the inciting incident; and its first-person density ran at half the original author's, losing the narrator of a first-person book. That file has sat in the archive for almost a month.

On August 12 the Grok 4.6 API opened — same base model, post-training upgrade, per xAI. Grounds for a retrial. Within roughly 24 hours we reconvened: identical protocol, each old charge re-examined, plus the customary bout against the reigning champions. The verdict up front: all three old charges dropped, 12:0 and 12:0 against the predecessor — and one new charge filed, which cost it the romance bout 2:10.

Terms of the retrial

The protocol matches the K3, Opus 5 and Gemini 3.6 Flash rounds item for item: the same two books (an 8.9-million-character Chinese fantasy epic, and Empresses in the Palace), the same anchor at roughly 55% depth, the same 16k-token rolling window, 20 rounds per chain at temperature 0.7, the model as the only variable. The defendant ran on xAI's official API (api.x.ai/v1; the model ID is grok-4.6 with a dot — the hyphenated form 404s), tested 2026-08-13, no protocol deviations. The aggregators hadn't picked the model up yet, which conveniently forced us onto the cleanest possible channel.

Every opponent is archive footage, re-run zero times: Grok 4.5's two chains, fantasy champion DeepSeek V4 Flash's chain and romance champion V4 Pro's chain, all from the July 16 benchmark archive. Two condition gaps, disclosed in the open: the fantasy opponents ran under our older context-labeling wording (the palace-side chains all share the current wording, so those pairings are clean), and the opponents are snapshots rather than today's live weights — a caveat that acquired new weight the very same day; we'll get there at the end. Judges: six models across five vendors; deepseek-v4-pro recused itself as a contestant's stablemate and gemini-3.5-flash took the seat per precedent; no xAI model judges. All 48 verdicts returned clean.

Charge one: mid-chain verbatim looping — dropped

Old evidence first. Grok 4.5's fantasy rounds 9, 10 and 11 are an identical 271 characters — the same two passages, three times — re-verified byte for byte before this retrial. The palace chain's image swamp was also measured to the bottom: calamus 21 times across 19 rounds, “musk mingled with” 13 times. Six judges flagged both without knowing whose chains they were reading.

New evidence: Grok 4.6's cross-round 12-gram repetition peaks at 2.8% (fantasy, round 14) and 2.1% (palace, round 13) — far under the 30% alarm line, not one round flagged. A round-by-round human read finds linear plot on both books: the fantasy chain runs from blood-cocoon siege through breakout to regrouping without one backtrack. Charge dropped.

Charge two: the stalled story clock — dropped

Grok 4.5's palace chain never left the afternoon of the incident; both blind reviewers independently wrote “plot rewind” in the original benchmark. Grok 4.6 moves the clock: the accusation scene, house arrest, a demotion edict, aftermath maneuvering, a fresh plot hook planted for the next arc — twenty rounds that actually leave the incident day. gemini-3.1-pro's line from the predecessor bout reads like a closing statement: “B's pacing is composed, its sentences fall in varied lengths, and scenery like the brick seams rinsed with well water carries the original's cool, fine-grained spirit” — the quoted image exists verbatim in round 17. Charge dropped.

Charge three: first-person collapse — dropped

This one has a blunt number. The source writes “I” at 9.0 per thousand characters; Grok 4.5 managed 5.2, a first-person book slowly losing its narrator. Grok 4.6 reads 11.0 — level with the romance champion's archived chain, back in the source's range. Of the three convictions this is the cleanest repair; no commentary required.

The verdict: 12:0, and another 12:0

Against the predecessor: fantasy 12:0 (windows 12:0 / 12:0 / 12:0), palace 12:0 (windows 10:2 / 12:0 / 12:0), and all twelve mapping-stable verdicts at high confidence. The combined 24:0 is the most lopsided generational result this pipeline has produced — the previous record was Opus 5's single-book 12:0. xAI's “post-training upgrade” claim is, in the fiction-continuation domain, measurably true.

Two verified compliments from the fantasy bench: gemini-3.5-flash — “the original is a minimalist, action-heavy, description-light fast webnovel; A fits this plain-spoken, high-action-density register perfectly”; claude-4.6-sonnet — “short-sentence attack, breath-level detail riding the action rhythm ('the blood cocoon's surface shivered', 'rust-iron tang'), dialogue brief and restrained” — both quoted phrases exist verbatim in round 1.

The new charge: outline-itis

The romance bout against reigning champion V4 Pro (0716 snapshot) went 2:10, and it is a strong signal: five judges mapping-stable for the champion, glm-5.2 alone stable for Grok 4.6 (a taste split), zero contradictory ballots. The five wrote variations of one sentence. gemini-3.1-pro: “too few words, too fast — the long-form texture of the original is entirely lost.” gemini-3.5-flash: “severe outline-style shrinkage late in the chain, anemic narration, no layered detail.” kimi-k2.6: “degrades into outline prose late, losing the winding delicacy.”

The mechanical exhibit pins the number: about 150 characters per round from Grok 4.6, about 475 from the champion. The irony deserves its own line — Grok 4.6 is the only chain of the six on the table that held the 80-220-character length spec every single round; the champion runs 475, the predecessor 239-267, all out of spec. The same 150 characters that fantasy judges praised as punchy rhythm, romance judges read as poverty. Rule-keeping is a virtue in one genre and a charge in the other — for anyone wiring a continuation model into a reader, that finding is worth more than the score.

The championship bout: unresolved

Fantasy, against reigning champion V4 Flash (0716 snapshot): raw ballots 8:4 for Grok 4.6. But this bout drew the heaviest position-bias contamination in the series: four of six judges cast eight ballots that contradicted themselves under the mapping flip, all voided. The surviving stable component runs 4:0 for Grok 4.6 — claude and kimi, with no judge stable for the champion. Direction favorable; confidence insufficient. Stack on the old-directive condition gap in the opponent chain, and our own rules say: no crown changes hands. Logged as “same tier as the champion, stable ballots 4:0 in its favor.” The champion's chain took its own verified hits — judges flagged late-window padding collapsing into log-keeping prose, with the recurrence counts checked one by one.

Grok 4.6's residual record, also in the open: image-level reuse. “Rust-iron tang” nine times across nine rounds, “chill” twelve times; on the palace side a metallic-sweet taste six times in eight rounds and the well-water motif across five. The two-sidedness is real — the same “brick seams rinsed with well water” that one judge praised as the original's spirit, two others filed under repetitive imagery. Downgraded from verbatim looping to a fondness for certain images: that is what remains after the generational repair.

Release terms, and the bill

List price (xAI's model page, verified 2026-08-13): $2 per million input tokens, $6 per million output; cached input metered at $0.5 in our runs; a double-priced faster tier we didn't touch; 500K context; reasoning on by default. The bill: $2.0955 across both chains as metered round by round, about $2.10 with probes — roughly ¥15, the cheapest run in this series despite a list price 3-10× the domestic models' (K3's round: ¥12.21; Opus 5's: $12.54). The reason is the new charge itself: 150-character rounds, with reasoning metered separately. That reasoning runs 95.5-95.6% of output tokens, the highest share we've recorded (K3 held the record at 81-86%); visible prose is about 130 tokens per round, the thinking about 21 times the prose; latency 39-44 seconds per round, and inside a reader that wait is real.

Boundaries: one prompt framework; two books, both Chinese; one chain per book, single anchor; the fantasy opponents carry the old-directive condition; Grok 4.6 was about 24 hours old and hosted behavior may drift; the outline-itis verdict's sensitivity to the anchor point is untested, and n=1 stays n=1. One boundary gained weight the same day: the opponents are July 16 snapshots, not live weights — and on August 13 DeepSeek hot-swapped V4 Pro to its GA build under the same model ID, a build that lost 1:11 to the very snapshot Grok couldn't beat. The romance winner is now as unreachable as the loser's old case file. The same day handed the industry both halves of the lesson: one vendor's new version swept its predecessor 12:0 twice; the other's lost to its own preview with zero stable ballots.

The practical ruling: for fast-paced fantasy, Grok 4.6 belongs on your shortlist today; for dense period prose it is one tier of narrative supply short — check the buying guide for the current callable list. Whether to switch is not a question for us: same opening, three continuations each, read them shuffled and flip the cards. This file reopens the day Grok 4.7 ships.