On DeepSeek V4 Flash official API day, we let the in-app Agent choose and rewrite scenes inside a 385-chapter novel
On the day DeepSeek V4 Flash entered official API public testing, we re-ran the same 385-chapter urban novel and user prompt inside the Foreverse App Agent: read relevant chapters, judge which scenes deserve expansion, then rewrite. The book survived, write failures hit zero, chapter-tool calls numbered about 174, and whole-file edits stayed at zero. Thirteen chapters grew by ≥300 characters; the run took about 19 minutes. Versus a July 21 baseline on the same model ID, tool discipline improved; throughput did not.
On the afternoon of July 31, the same urban webnovel was still open in the emulator: 385 chapters, about 3.39 MB at baseline. The user message was one sentence—starting from chapter 50, expand the scenes that deserve more ink—no tool names, no chapter list. The agent read relevant chapters, judged, rewrote, and sealed after 62 steps. The book survived. Anchors held. No chapter collapsed into an empty shell.
That same day DeepSeek’s official changelog said V4 Flash had entered official API public testing. The request ID is still deepseek-v4-flash. We pulled a whole-book rewrite stress run from July 21—same ID, same book, same prompt—and asked a narrower question: after the official release, is the in-app long-novel agent more obedient?
What did we test for DeepSeek V4 Flash official as a novel agent?
This is not “continue one more segment.” It is the Foreverse agent with real read/write tools: the novel is a file, rewrites must land on disk. The protocol stays blank on purpose—zero tool hints—so the model must choose which chapter to open and which edit entry to use. The baseline is the July 21 run on the same book and prompt; both sides used BYOK against the official channel with the same model string.
The accurate framing is read-then-judge rewrite. The run logged 168 chapter reads and concentrated its work between chapters 50 and 132. It operated inside a 385-chapter book; the evidence does not show a sequential read of chapters 1–385. That distinction matters more than a punchier benchmark headline.
Which numbers moved versus two weeks ago?
| Metric | 2026-07-21 baseline (three-run totals; speed from matched turn 2) | 2026-07-31 official day (one run) |
|---|---|---|
| Book survived | Yes (chapter count held) | Yes (385 chapters held) |
| Write failures | 10 across three baseline turns | 0 |
| Chapter / whole-file write calls | 18 / 76 (19% chapter-scoped) | 174 / 0 (100% chapter-scoped) |
| Chapters with ≥300-char growth | ~15 in the continuous stretch | 13 |
| Steps / wall clock | 72 / ~530s | 62 / ~1163s |
| Prompt / completion tokens | ~5.79M / ~44k | ~5.22M / ~119k |
The middle rows are the story. Two weeks ago it preferred bulk-editing the whole manuscript file, and failures piled up there. Today it almost only touched chapter-scoped entries and failed zero times. That is a hard signal for long unsupervised work: choose the right tool, or the error-recovery loop never gets a clean chance.
The ~174 figure is a chapter-tool call count, not 174 different chapters or 174 visible outcomes. The analyzer identified 87 successful landing points, and repeated edits on the same chapter still inflate that number. We diff by chapter and count only ≥300 characters of net growth: 13 chapters. Same order of magnitude as the baseline’s ~15—better discipline, not double the yield.
Why was the official-day run slower?
1,163 seconds versus 530; completion tokens from ~44k to ~119k. It may have read in finer slices and edited more often, or the endpoint’s load and sampling behavior that day may have changed. We did not run a second machine A/B, so we only report the observation: same prompt, not faster.
For writing products that means Flash can carry a read-then-rewrite loop, but do not treat it as a batch engine by default. Budget time and tokens for tens of minutes and six-figure completion; leave the room only if task recovery exists—do not expect a two-minute handoff.
How should this sit next to the fiction continuation test?
The same day we published the official fiction 20-round and reasoning-effort field test: less mechanical looping, while high/max can still spend the output budget on hidden reasoning. That run is direct API continuation; this one is an in-app tool loop over whole-book files. The tables are not interchangeable. Flash as a budget fantasy continuer does not automatically mean “the only correct long-rewrite agent.”
The evergreen hub best model for continuing novels still owns genre picks; agent ability needs tool failure rate, book invariants, and chapters with real growth. For more agent tasks, see five tasks on a bookshelf agent, including a self-correction after a failed edit.
Should you run long-novel agents on Flash official now?
If the job is plain urban prose and selective scene expansion, the official-day numbers are usable: the book does not explode, tools behave, and a dozen chapters show real growth. If you need ornate palace diction, heavy planning, or a five-minute deadline, do not treat this post as a green light—switch models, split the job, or let it propose candidate chapters before you approve writes.
One boundary stays honest: the official catalog has no second Flash ID. All we can claim is a time-point comparison between the release-day endpoint and the July 21 snapshot. If weights move again without a rename, the numbers need a re-run. We will not brand a changelog date stamp as a new model name.
FAQ
How is DeepSeek V4 Flash official better as an agent than two weeks ago?
On the same book, prompt, and model ID, write failures fell from a cumulative 10 on the baseline to 0. Almost all rewrites used chapter-scoped tools instead of bulk-editing the whole manuscript file. Content-wise, 13 chapters grew by ≥300 characters—same order of magnitude as the baseline’s ~15-chapter stretch. Discipline improved; output volume did not explode.
What does “read-then-judge rewrite” mean versus dumb expansion?
The user gives one goal sentence: no tool names, no chapter checklist. The agent must read the TOC and relevant prose, decide which scenes deserve ink, then write chapter by chapter. We score book survival, write failures, and chapters with real growth—not how many times a tool reported success. This run logged 168 chapter reads, but that does not prove a chapter-by-chapter read of all 385 chapters.
How long did the official-day run take, and how many tokens?
62 steps, about 1,163 seconds (~19 minutes), ~5.22M prompt tokens and ~119k completion tokens. The July 21 same-prompt baseline was ~72 steps / 530 seconds. Today was slower with higher completion. Flash can finish the loop; it is not automatically fast or cheap.
Is there a separate public model ID like Flash 0731?
No. The stable request ID remains deepseek-v4-flash. The date stamp in DeepSeek’s changelog is a hosted version label. This article compares the official-release-day endpoint to a July 21 snapshot, not two different model strings.
Flash wins fantasy continuation—should it be the default long-rewrite agent too?
Continuation blind ranks and long tool loops are different scoreboards. Flash often remains a budget pick for plain-spoken fantasy continuation. For read-then-rewrite jobs, watch tool choice, book invariants, and chapters with real growth. For ornate registers or heavy planning, switch models per task instead of making Flash the only agent default.
Questions or ideas? Join our Discord →