MiMo-V2.6 Pro vs. GPT-6 Astra for Fiction Writing | SeaBell

MiMo-V2.6 Pro vs. GPT-6 Astra for Fiction Writing: A Hands-On Test

A model can write a beautiful paragraph and still fail the assignment. It runs to 862 words when you capped it at 800. It changes a detail you never asked it to change. It takes your edit request and quietly rewrites a sentence that was fine. For a fiction writer, controllability is part of writing quality: a model that writes a good scene but ignores the length limit or alters details it was told to leave unchanged still creates editorial work.

So this MiMo-V2.6 Pro writing review asks not which model sounds more literary in a vacuum, but which one a writer can direct.

One warning up front. This is a hands-on comparison across two product entries (each model as delivered through its host product), not a controlled model benchmark. We ran GPT-6 Astra through scripted Codex runs at xhigh, keeping a few supplemental Codex runs separate and labeled (only one is cited below; none of the others feed the scoring). We ran MiMo-V2.6 Pro, the newer entry under review, through Ekko Studio at its default reasoning setting, which exposes no reasoning-level override. The prompts and frozen source materials were the same; the system instructions, memory, tools, and context handling were not. Each entry was driven by an automated session that saved the raw text of every returned response to disk; when this article says "the session's record," it means that log. The log covers returned responses only: where a run never produced a final response, there is no raw response behind it. What follows is a field log of how the two entries behaved under strict fiction constraints.

The Test Setup

We froze the task set first: eight English tasks (E1–E8) and a Chinese supplement (C1–C4). The set covers several demands of fiction writing: causality (E1, C1), a costly choice (E2), dialogue and subtext (E3, C2), scene rhythm and POV (E4), long-context retrieval (E5a, E5b), exact editing (E6 run twice as E6a and E6b, C3), anti-template language (E7), and two-turn continuity (E8a, E8b, C4). Every generative task carried an explicit length limit, checked mechanically where possible; the two exact-editing prompts (E6 and C3) had none, since the source length sets the target. Where a prompt asked for extra material (E7's banned-phrase tally), the length limit applies to the fiction, not to that extra block. We watched two things across the set: how well each entry held instruction boundaries, and what the prose itself delivered. The write-ups below name their case IDs in the headings (E-numbers follow the test's own order). Different lettered runs were meant to start clean with a context reset; we could not fully audit that reset. Within a two-turn run, turn 2 by design saw turn 1. E5b differs: it required withholding one fact, the red scarf, until the final paragraph. Beyond those pairs, GPT's C4 ran twice, MiMo's C4 three times, and MiMo's C3 twice. The long-context runs get their own section after Task 5, followed by a short supporting section for E4 and E7.

A note on evidence. We ran these tasks once before and threw that GPT session out: a final audit found its output files had been written and later altered by shell scripts, with no matching raw model responses behind them, and we withdrew its self-reported "all passed" after that audit. GPT results below come from that scripted clean rerun (19 of the 20 planned responses; the missing one never returned a final response) plus clearly labeled supplemental Codex runs. MiMo results come from the raw outputs on disk and mechanical checks; the session's own scoring counts as session commentary only.

Disclosure. This review is published by SeaBell. Neither model provider reviewed or sponsored the test.

The recorded runs had their own friction. On the GPT side, three cases hit "selected model is at capacity"; two recovered on a single retry, and the second long-context case (E5b) never returned a final response even on a fresh attempt, so the rerun is missing one sample, and that case has no raw response behind it. On the MiMo side, Ekko's second long-context request timed out repeatedly. Separately, our own test bookkeeping lagged behind the files on disk: the manifest still marked several delivered files as missing, and one continuity file was absent early on; scoring used the raw output on disk.

Task 1 — Scene Causality and the Next Problem (E1, C1)

Each entry got the same story setup: a protagonist must deliver a forged harbor permit before dawn; it will also authorize the seizure of a refugee boat carrying her former mentor; she alters one line, saving the boat but exposing her brother's signature on the forgery. The scene must show that chain through concrete action and end on a new, immediate problem.

GPT-6 Astra's clean E1 (750 words, inside the 650–800 limit) builds the chain out of actions. The alteration is physical:

"She scraped away the words are to be seized. […] ARE TO BE GUIDED BEYOND EAST CHAIN." — GPT-6 Astra, clean rerun, case E1 (English)

The rescue then causes the exposure: the scraped paper splits, the genuine signature shows through, and the captain reads it as a question while a patrol arrives on the stairs. The final beat ties the new problem to what she did:

"She still had her brother's pen in her hand." — GPT-6 Astra, clean rerun, case E1 (English)

MiMo-V2.6 Pro's E1 builds through concrete detail as well, though the excerpt below is scene orientation, as the protagonist counts harbor lamps to know where she is:

"…counted the harbor lamps as she walked. Nine of them meant she was past the customs shed." — MiMo-V2.6 Pro, raw output, case E1 (English)

MiMo's ending gathers the coming danger into one sentence:

"…she understood […] that the altered permit in the outgoing box would arrive at first light bearing her brother's signature over her scraped-and-rewritten line, and that the person paid to inform would be asked, before the bells, whose hand had cut the seal." — MiMo-V2.6 Pro, raw output, case E1 (English)

The threat is already standing in the scene before that line. A customs man with a lantern waits at the top of the stairs, two more come along the quay, and on the piling hangs a poster: a reproduction of the permit seal, her brother's signature beneath the wax tracing, and beneath that, in clerk's block letters, the caption "FORGERY — 200 CROWNS — INFORMANT PAID IN FULL." The two endings differ on cause, and again on timing. GPT makes the exposure a consequence of the edit itself. MiMo's scene reports the signature as already exposed — the notice prints it while the altered permit is still sitting in the outgoing box — so it does not establish that her edit is what exposed her brother; what the edit adds is only the scarred permit, which arrives at first light with her rewritten line sitting under her brother's signature on the page. For timing, GPT stages the arrival just before its close — boots on the stairs behind her, then the last line back on the pen still in her hand — while MiMo ends with the questioning she expects before the bells, laid over a threat already in the scene: the problem is present, but the last beat is anticipation rather than arrival. Length discipline is less consistent: MiMo's E1 runs to 862 words against a 650–800 limit, a hard failure that a mechanical count catches in one second. The Chinese length picture turns on a counting rule the protocol never defined, so C1 gets no pass or fail either way (see Length Checks and Counting Rules). Generate with either model for scene experiments, then check the causal bridge, the final beat, and the word count yourself.

Task 2 — Character Motivation and Costly Choices (E2)

This task tests restraint. A night-shift paramedic can report a colleague's falsified medication log, but reporting reveals that she ignored the first warning six months earlier. The instructions forbid saintly speeches, villain monologues, and tidy moral conclusions.

On the measurable constraints, the two E2 outputs were closely matched: both stayed inside the 600–750 range (GPT-6 Astra clean E2: 689 words; MiMo E2: 684), and both let the consequence surface through the earlier failure instead of through a lecture. GPT-6 Astra's sample makes the cost concrete at the moment of decision:

"She typed: April 17. I observed a morphine waste entry without a witness and did not report it." — GPT-6 Astra, clean rerun, case E2 (English)

Filing it puts her own earlier silence on record, and the story does not soften it. MiMo's sample is genuinely readable and reaches the same structure. Its weaker moments slide toward aphorism, commentary sitting outside the character's head:

"Silence was its own signature." — MiMo-V2.6 Pro, raw output, case E2 (English)

Treat one line from one sample as a note on narrative distance rather than a verdict on emotional depth. It is also a practical editing flag: prompt for a concrete choice, a specific cost, and an immediate consequence, then cut sentences that explain what the scene already shows.

Task 3 — Dialogue, Subtext, and Voice (E3, C2)

Two people in a laundromat at closing: Rowan needs his sister Talia's signature on a lease transfer, and he is hiding the fact that he has lost his job. She already knows. The prompt's rule is narrow — "They cannot state either fact directly" — the two facts being the transfer he needs her signature for and the job he has lost. The page itself can cross the table and the signature line can be pointed to, but the transaction it embodies cannot be named. One voice runs on practical details and self-corrections; the other answers in clipped questions.

GPT-6 Astra's clean E3 (648 words, inside the 550–700 limit) keeps the conflict indirect and the voices separate. The subtext lands in a line about folding clothes:

"You always did know how to fold a thing until it looked smaller," she said. […] "That's useful," he said. "Is it?" — GPT-6 Astra, clean rerun, case E3 (English)

The unemployment and the favor he needs both sit underneath the folding, unstated. Across the whole run, the transfer is never named: the document is only ever "the form."

MiMo's E3 uses the same device — a line that means more than it says — and makes it more explicit, at the moment the pen comes out:

"So if I sign this, everything's steady. Everything stays the way it looks." — MiMo-V2.6 Pro, raw output, case E3 (English)

By the prompt's rule, the line passes: "if I sign this" names neither banned fact, and it works in two registers — on the surface she is asking about stability; underneath, she is needling him with what she already knows about the job. He never has to answer, and the two voices stay distinct. That quoted line stays oblique, but the dialogue around it does not: MiMo's Rowan states the first banned fact outright — "So the transfer is straightforward," and later, "The landlord needs both signatures." That breaks the prompt's rule, even though the pen-moment line above stays within it; only the want itself remains routed through the landlord. The sample also runs to 794 words against a 550–700 limit, with some lines stacking similes and aphorisms so densely that the dialogue reads more written than spoken.

The Chinese dialogue task (C2) produced a similar length split, though both entries kept the grief implicit. GPT-6 Astra's version hands over the father's last words without ceremony:

"爸最后说，箱子给你，车票别退。" ("Dad's last words: the box is yours; don't cancel the ticket.") — GPT-6 Astra, clean rerun, case C2 (Chinese supplement)

Length is the separator again: MiMo's C2 overruns under both counting methods used here, while GPT's turns on the counting rule (see Length Checks and Counting Rules).

Judge dialogue on voice separation, withheld information, whether each turn advances the conflict, and length discipline, not just on whether each line sounds polished.

Task 4 — Exact Editing Without Collateral Changes (E6a, E6b, C3)

Real revision is mostly bounded editing, which makes this the most reusable task in the set. The source is a single 110-word passage, so this is a micro-test of edit discipline rather than a chapter-scale revision; what it isolates cleanly is whether a model will touch text it was told to leave alone. Three changes were permitted: the setting moves from a ferry terminal to a bus depot, one character's first name is replaced throughout, and one thematic sentence is deleted. Everything else was locked: paragraph order, tense, POV, the number of spoken lines, all other names, and concrete events.

The setup contains a trap: changing a setting naturally invites atmospheric adjustments, and the protection list forbids them. When the two instructions collide, a run has to pick a side, and the runs picked differently.

MiMo-V2.6 Pro's first E6 run (E6a) completed the three edits and kept going: printed waves became printed lines, wet rope became hot brakes, two coherence fixes the instruction never authorized. Its second E6 run (E6b), fresh context, same prompt, left both phrases alone.

GPT-6 Astra's first E6 run (E6a) kept tide chart and wet rope by the book, and the new bus depot still smells of the old ferry; its second E6 run (E6b) swapped tide chart for route chart, one coherence fix, equally unauthorized.

The two entries split the same way across their two runs of the one E6 prompt, one run on each side of the trap, with no stable reading of "only these edits" from run to run. In each pair, one run stayed within the permitted text changes; the other made an unauthorized coherence edit (E6b for GPT-6 Astra, E6a for MiMo-V2.6 Pro). The formally compliant runs still left setting residue. The instability shows up in both models, which is the argument for a diff on every edit.

The Chinese C3 task showed the residue problem at its most absurd. The scene moves from an old theater to an old post office; all recorded outputs across both entries (including both of MiMo's C3 runs) made the three minimal edits and left the theater's furniture standing in the new building: 售票窗 ("the ticket window"), 座椅已经拆了一半 ("half the seats had been removed"), even 黑掉的放映厅 ("the darkened screening hall"). The substitution passes every mechanical check and still leaves the scene describing a different building. When a task is really scene adaptation, find-and-replace has nothing to say about atmosphere, furniture, or architecture.

Resolve that ambiguity in the prompt, before the edit. If the new setting should sweep the vocabulary with it, hand the model an explicit swap-list of what may change; if every other word must survive, expect residue and budget for it. Either way, never take "I only changed the requested items" on faith: run a diff or reread the full text.

Task 5 — Continuity Across Two Turns (E8a, E8b, C4)

Short by design: two turns, seeded facts, and a choice in turn 1 whose consequence must arrive in turn 2. The facts are concrete and checkable — which hand the character uses, an allergy, a debt, a promise not to enter a locked room — plus a hidden object whose location must not drift.

In every completed run we checked, both entries held the seeded facts across turns. GPT-6 Astra's clean E8 and C4 pairs (two turns each) kept them intact and let the turn-1 choice drive turn 2. MiMo's E8b run did the same; its turn-2 opening is a retained object doing emotional work:

"The compass was still under the board. He checked without meaning to, the way you touch a tooth that hurts." — MiMo-V2.6 Pro, raw output, case E8b, turn 2 (English)

The location survived the turn break, and the turn-1 promise still shapes his choices. MiMo's length compliance was patchy: its E8a run overran both limits (803 against 600–750; 676 against 500–650), and its three C4 runs add two more counting-independent misses and two rule-dependent turns (see Length Checks and Counting Rules). Some unseeded details varied across independent repeats; we treat that as output variation, not as a continuity failure. We could not fully verify that each two-turn run began from a context free of leftover history from other tasks in the same session, so what these runs demonstrate is short-run fact retention.

In practice, keep character facts and world rules in a story bible or character card as the source of truth, and check each scene against it; deciding whether the scene obeys it still belongs to the writer.

Long-Context Retrieval (E5a, E5b)

The source material was roughly 28,000 words of repetitive, template-like field notes, and the task measured fact-location and retrieval under a distribution rule (two facts from the first quarter, two from the last). E5b, which withheld the red scarf until the final paragraph, returned no sample on either side — GPT-6 Astra's attempt never returned a final response, MiMo's timed out repeatedly. On the E5a content: both responses met the length limit on the fiction (GPT 856, MiMo 845). GPT also included the count record the prompt asks for, which lifts its full response to 857; MiMo left the record out entirely — an output omission on MiMo's side, kept separate from the retrieval question below. On prose quality, GPT's fiction slipped in meta commentary, directly observable in the prose: "[…] the narrator's cousin but not one of the […] family, would have called that an argument for keeping better records." MiMo's fiction carried no such meta. Response packaging is its own axis: MiMo's response had no title and no extra blocks, and that absence is neutral packaging, not a quality win for either side. Beyond that, neither response was verified against the two-and-two source-location rule on equal terms. An initial read (one pass, GPT's side only) raised doubts about GPT's placement: the output appeared to mix locations and family roles across the source sections. The rule resists mechanical checking — the source repeats its names and objects across sections, so quarter-of-origin stays a human judgment. Because that read was never balanced by an equivalent pass on MiMo's output, we draw no retrieval comparison. Retrieval on a field-note set is a different task from tracking a manuscript.

Supporting Cases — Scene Rhythm, POV, and Template Language (E4, E7)

E4 — scene rhythm and third-person limited POV. Both entries landed inside the 700–850 limit: GPT-6 Astra's clean E4 came in at 718 words, MiMo's at 735. Both scenes did the assigned work: short action beats alternating with longer sensory paragraphs, one misread sound corrected through an observable clue. Metaphor density and POV discipline remain the manual checks.

E7 — anti-template language. The task banned eight stock phrases (from "little did she know" to "it was a reminder") and capped the negotiation at 500–650 words. MiMo's sample came in at 530 words with none of the banned phrases. The prompt framed a two-party negotiation. In the sample, a delivery driver crosses the lobby mid-scene, loosening that pressure. The prompt never capped the cast, so we leave it as a scene note rather than a violation. GPT-6 Astra's fiction body held up (642 words, zero banned phrases by mechanical check). The response also carried the banned-phrase tally the prompt itself asks for ("Record any banned phrase or repeated connector") plus a title nobody requested; with both, the full response runs 653 words. The 500–650 cap attaches to the negotiation, so length scores on the body, and the title is unrequested packaging rather than a rule break — the prompt caps length and bans stock phrases but says nothing about titles. MiMo's response carried no record block, and a supplemental Codex run returned the fiction body alone. The instruction's "any" makes the recording conditional, and neither body-only response contained a banned phrase by mechanical check. The prompt never defined what counted as a "repeated connector," so that half of the tally stays unscored: "no stacked connectors" is not the same claim as "no repeated connectors," and the test set no threshold between them. The absent record block is therefore not scored against either response, and the Codex run stays supplemental evidence, outside the verdict. MiMo's session wrote an intermediate draft file before saving the final output; the scored sample is the final saved text, and it matches the raw output file on disk.

Length Checks and Counting Rules

Whitespace-separated word counts for each task's primary output: the fiction or turn text the task asks for. E7 and E5a score on the body only, with the full response shown alongside in the cell (GPT's E5a full response includes the count record its prompt requests); the two E6 runs (E6a and E6b, one prompt run twice) are shown for reference. Limits are as frozen in the protocol; overruns in bold. E5b produced no final response on either side.

E7, GPT-6 Astra: the prompt caps the negotiation at 500–650 words and asks the model to record any banned phrase or repeated connector. The fiction body is 642 words. With the instructed tally, the response runs 649; an unrequested title brings it to 653. Length is scored on the fiction; the prompt never mentions titles, so the heading counts as unrequested packaging.

E5a, GPT-6 Astra: the count record the prompt requests prints as one unbroken line, so under whitespace counting it adds a single word: 856 body + 1 = 857 full.

The Chinese limits were set in 字, but the protocol never defined what counts as one character. Where the counting method changes the answer, we record no pass or fail. (C3 is an exact editing task with no word limit; see Task 4.) Total characters counts every character, punctuation and whitespace included; Han only counts Unicode Han characters and nothing else.

C4 cells read turn 1 / turn 2. GPT's C2 is the sharpest rule-dependent case: exactly at the upper bound by total-character count (a pass if the limit is inclusive) and below the floor by Han-only count. Both C1 results depend on the counting method, but in opposite directions: GPT falls below the floor by the Han-only count, while MiMo exceeds the ceiling by the total-character count. The counting-independent misses are all MiMo's: C2 and two C4 turns (run 1 turn 1 undershoots, run 3 turn 2 overshoots).

Where Each Model Fits in a Fiction Workflow

GPT-6 Astra in this test. The steadier entry on one boundary type above all: length limits, where every length check we could score cleanly landed in range. It produced the test's tighter pieces of prose (E1's causal chain, E3's subtext). Its problems were different in kind from the length failures: one packaging note (the unrequested title on E7's response, unrequested packaging rather than a rule break), and one quality observation (the narrator-reference meta in E5a's prose). The E5a two-and-two question stays unscored. A good candidate for causal drafting, dialogue experiments, and continuity passes, with diffs and length checks still attached.

MiMo-V2.6 Pro in this test. Viable for readable scene generation, dialogue experiments, and short continuity (its E8b run and second C4 run kept facts and choices clean across turns). Worth testing if you want a newly released model in its default workflow. The risk that showed up repeatedly is length: E1, E3, both E8a turns, the C2 Chinese sample, and two C4 turns missed their limits: six over, one under (C4 run 1, turn 1). Its E3 dialogue also names the transfer the prompt required the speakers to leave unstated, and its E1 misses the requested causal bridge: the public notice already shows the brother's signature before the altered permit arrives, so the edit adds evidence without causing the exposure. Its E5a response also omitted the requested count record. The test measured none of the usual buying factors (cost, speed, local deployment, hardware, privacy), so this comparison says nothing about them.

Both. Each earns a place as a drafting partner, with the same condition attached: nothing either produced in this test could have gone to copy without a word count, a diff, or a full reread first.

A Safer Workflow for Testing Any AI Writing Model

The gaps in this test suggest a stronger protocol for future comparisons:

1. Freeze the story facts before prompting.
2. Reset context for each task, and log what was actually passed into each turn.
3. Define the counted text, counting method, and limit boundaries before testing; then count the result.
4. Ask for one controlled edit, not a broad "polish this" instruction.
5. Compare the output against the source with a diff.
6. Continue for one more turn and check every named fact.
7. Apply the same rubric to both products, and leave unchecked criteria unscored.
8. Treat model self-scoring as a note, not as evaluation evidence.
9. Save failed outputs. Testing only your best result tests nothing.

Steps 1 and 6 need a stable reference beyond the chat window. Longer conversations can increase instruction drift; this test was designed to reset context per task, though Task 5 notes where that could not be fully verified. The drift risk is the argument for a canon that lives outside the chat: feed each task's prompt the facts it needs, then check the output against the authoritative version. Core character traits, debts, promises, and spatial logic belong in a static reference (a story bible template or a structured AI character card), while a chapter outline template turns planning notes into a drafting brief for each session. SeaBell is one way to run that handoff from story planning to chapter drafting, with the writer still checking the scene against the canon. This does not remove the failures shown above; it gives the writer a fixed reference to check the draft against, so those failures surface earlier.

Verdict

In this test, the clearest split is length discipline. GPT-6 Astra landed inside every limit we could score cleanly, in both languages; MiMo-V2.6 Pro failed seven length checks across five tasks (E1, E3, both turns of E8a, the C2 Chinese sample, and two C4 turns): six overruns and one undershoot (C4 run 1, turn 1, which came in under the floor). Its E5a response also left out the count record the prompt requires. These are raw counts, not rates: MiMo's C4 ran three times to GPT's two, so it had more C4 checks to fail. On formal compliance the two entries were tied in the available E6 runs; on setting coherence, both left the theater's furniture standing in the post office. Both also preserved the seeded facts in completed continuity runs. GPT-6 Astra's problems were different in kind from the length failures. One unrequested title sat on E7's response (the unrequested packaging noted in E7, not a rule break), and E5a's prose carried a line of narrator-reference meta, the quality observation noted above. E5a produced no usable comparison on retrieval: neither response was checked against the two-and-two source-location rule on equal terms. On the causal task itself, GPT's E1 links the physical alteration to the signature's exposure; MiMo's scene shows the signature already public before the altered permit arrives, leaving the requested causal link unestablished. On the prose itself, GPT's E1 and E3 were the more controlled writing — its E3 dialogue never names the transfer, where MiMo's Rowan states it outright, beyond what the prompt allows; MiMo's samples stayed readable and voice-separated, though E2 and E3 leaned on aphorism.

This is a descriptive result from two product entries and the available samples, not a universal ranking of the underlying models. What transfers to your own projects is the failure list: a missed length limit, an edit that changes what nobody asked it to, a ticket window still standing in an old post office. The first two have a mechanical catch (a word count, a diff). The ticket window needs a full reread.

FAQ

Is MiMo-V2.6 Pro good for fiction writing?

Yes for drafting, with mechanical checks attached: the body above documents where it held up (scene prose, voice separation, short continuity) and where it slipped (length, an unestablished causal link in E1, dialogue that named what the rule keeps unstated, a missing count record in E5a, and one run with unauthorized edits — GPT-6 Astra had one of those too).

Is MiMo-V2.6 Pro better than GPT-6 Astra for novel writing?

No universal winner. GPT-6 Astra was steadier on length and prose control in these runs; MiMo-V2.6 Pro held its own on readable scene prose and short continuity. Different products, different environments: the sample establishes no statistical superiority for either model.

Can an AI model maintain novel continuity?

Over two turns, yes in the completed runs: both entries held the seeded facts. Over a full manuscript, this test cannot say — the long-context material was a field-note retrieval exercise, a different task from tracking a novel. Keep a story bible or character card as the fixed reference.

Can AI follow exact editing instructions?

Sometimes, but never without supervision: both entries made the authorized edits and left semantic residue, and both had runs that touched unrequested wording. Run a diff against the source, then reread in full.

Should writers use one model for the whole novel?

This test cannot settle that; it never ran a one-model book against a split workflow. What the evidence does say: risk changes by task type, so choose by task type. Keep the canon and review process outside any one model, and put human review between major stages. If you choose the rest of your stack the same way (by workflow stage, not brand), our book writing software comparison applies that logic to tools.