We benchmarked Gemini on two novels. Then we priced the free tier

Gemini is the strangest entry in our continuation benchmark: the only model we had to disqualify, and the only one whose failure turned out to be entirely our fault. Both halves of that sentence are useful if you are deciding whether to write fiction on Google’s free tier, so this is the full log — the 20/20 failure, the one-sentence fix, the re-run that landed mid-table, and then the free-tier fine print, which has its own catch.

Run one: twenty rounds, twenty restarts

The setup, briefly: nine models each continued two Chinese novels for 20 consecutive rounds, every round’s output fed back into a fixed 16k-token context window, results ranked by double-blind review. The protocol and full tables live in the benchmark write-up. On the fantasy epic, Gemini 3.1 Pro did something no other model did: it ignored everything generated so far and restarted from the original text’s ending. Every round. Twenty phrasings of the same opening instant, a story that never took its second step.

Both reviewers marked it last without hesitation — you do not need a rubric to notice a book that refuses to move. If we had published that table and stopped, the takeaway would have been “Gemini cannot continue fiction”, and it would have been wrong.

Why did Gemini keep restarting the story?

Because one sentence in our context told it to. Our assembled context labels which passages are original prose and which are earlier AI continuations, and the explanation line described AI passages as “for plot continuity reference”. Gemini read “reference” as “not canon” and skipped them all — every round, it was faithfully continuing the last passage it considered real, which was the original ending. A restart loop is the most common failure in our long-run failure taxonomy, and this was its cleanest specimen.

Three ablations pinned the cause. Old wording: 20/20 restarts. Label with no explanation at all: still roughly 6 of 8 rounds restarting. Explanation rewritten as two explicit statements — these passages are canonical events that already happened; continue from the very last paragraph — zero restarts in 28 rounds across both books. One sentence was the entire distance between disqualification and a normal ranking, and that rewritten sentence now ships inside Foreverse’s continuation pipeline. If you assemble context by hand in a chat window, it is also the sentence you should steal.

Run two: mid-table, and the most literary voice of nine

With the label fixed, we ran Gemini 3.1 Pro against the palace-intrigue novel alongside eight other models. It placed fourth to sixth of nine — genuinely mid-table, zero restarts, and the reviewers’ note is worth quoting because it describes a style, not a defect: still the most literary-flourished voice of the field. Gemini decorates. On an ornate first-person court novel that reads as richness; on the plain, fast fantasy epic it would read as perfume. Neither reviewer ever mistook it for the original author, which is the honest ceiling: a capable continuer with a recognizable accent, not a mimic.

The re-run also settles a question the disqualification had left open: nothing about the 20/20 failure was a capability problem. Same model, same books, one sentence changed, and the placement moved from dead last to the middle of a nine-model field. We have watched people abandon a model over exactly this symptom, one they could have fixed in their prompt in under a minute.

What does the AI Studio free tier actually cover in 2026?

The reason Gemini keeps coming up in fiction threads is not the ranking above — it is that Google still runs the most usable genuinely free API tier of any frontier lab. The shape of it, as of July 2026: free means flash-class. Pro-class models went paid-only in April 2026, so the free key covers Gemini’s fast models, not its flagship. Quotas are enforced per Google Cloud project, not per key, and Google’s rate-limit page publishes the live numbers inside AI Studio per project rather than as one fixed table — the working figures are on the order of 10 requests a minute and a few hundred requests a day for flash-class models, with daily caps resetting at midnight Pacific.

Translate that into reading terms and the daily cap is roomier than it sounds. One continuation is one request. A long evening might spend thirty. The wall you can actually hit is the per-minute cap, and the way you hit it is regeneration sprees — rerolling a scene five times in ninety seconds. The other line item is quieter: Google’s pricing terms state that free-tier prompts and responses may be used to improve its products. Paid-tier traffic is treated differently. Continuing a published webnovel, you may shrug; feeding it your unpublished manuscript deserves a pause.

The mismatch nobody prints

Read the two halves of this post against each other and there is an inconvenient gap: the model we benchmarked — 3.1 Pro — is not the model the free tier gives you. Our blind data covers the paid flagship; the free key covers flash-class models we have not blind-ranked for continuation. So the honest claim is narrower than either half suggests: the free tier is a real $0 on-ramp with real quota, and the ranked, known quantity costs money. Anyone telling you “Gemini placed mid-table and it’s free” is splicing two different models into one sentence.

The workaround is to run the audition yourself, and the free quota is exactly enough for it: take one fixed passage from your book, have a flash-class model continue it for several consecutive rounds, do the same with one or two rivals, then read the outputs in a shuffled order before checking which was which. That is our benchmark protocol at kitchen scale, it costs zero on this tier, and it answers the only question a ranking never can — whether the accent works on your book.

Getting a free key into a reader

The setup is the least eventful part of the log. Sign in at aistudio.google.com, open the API keys page, create a key in a new or existing Google Cloud project, and copy the string — AI Studio lets you view it again later, which OpenAI and Anthropic do not. Two dated footnotes from our provider directory: keys created in AI Studio now are auth keys by default, and Google’s legacy standard keys stop working from September 2026, so new setups are unaffected but a key from an old tutorial may have a shelf life.

In Foreverse, Settings → AI models & services → Model providers → Gemini, paste, save, and run the built-in capability test — one real call that settles whether key and endpoint agree before you are three chapters deep; the provider directory keeps the per-provider fine print, including the free-tier note. Set Gemini as the default text model or pick it per request on the continuation form — per-request switching is how you route court scenes to one model and battles to another inside the same book. Import a book, long-press a paragraph, continue. The key stays encrypted on the device, requests go straight to Google, and every call logs its model and token counts in the app’s request history, so you can audit exactly what the free tier absorbed. If it earns its keep, flipping billing on in AI Studio upgrades the same key rather than starting over — and if you would rather skip Google accounts entirely, Foreverse’s metered official channel carries Gemini entries alongside the DeepSeek tiers, with 5,000 signup credits covering roughly 260 continuations before anything costs money.

Where the log ends up: we disqualified Gemini, retracted the disqualification after three ablations, and watched it take a calm mid-table run once our own sentence stopped sabotaging it. The free tier is the cheapest way to find out whether its accent suits your book — just know that the accent you audition on flash-class is not the voice we ranked, and that twenty restarts in a row is, occasionally, the prompt’s fault and not the model’s.