On the evening of November 18, 2026, Weibo's hot search carried a question with a number in it: which AI can draw 13 eggs correctly? [1] Behind it was a failure anyone could replay. A user asked Doubao, ByteDance's AI assistant, for an image of a box holding 13 eggs — and got 12, then 9, through more than a dozen rounds of corrections that never landed on the requested count. The session ended with Doubao giving up on the image entirely and typing the answer as text: rows of 4, 4 and 5 [2][3].

The failure exposes a boundary the current generation of image generators shares: a number written into a prompt is a weak constraint [1][4]. The most systematic multi-model test the episode produced — AI blogger @零重力瓦力's nine-model run, published by a Chinese community aggregation on November 19 — found exactly four of nine mainstream image models drawing the required count: GPT Image 2.6, Nano Banana 2.1, Ideogram 4.5 and Luma UNI-1 [1]. If a viewer will count anything in your image — inventory shots, seating charts, product lineups, hands — treat every generated count as an unverified draft, and treat the model's own "I counted" as no verification at all.

Four of nine passed — in one blogger's single test

The test behind those four names was deliberately strict. One unified prompt — a top-down shot of 13 white eggs with visible gaps between them, plus an instruction to count before finishing — went to all nine. The names come from a single blogger's single round, not from a benchmark; they are one unusually well-documented data point, not a ranking.

The failure behind the hot search is better documented. The original on-site record is a Weibo post from user 一号哨位, published at 11:18 on November 18 with screenshots of the multi-round exchange; by that evening it had collected close to 2,000 likes [2]. The screenshots show Doubao apologizing fast and promising the fix in full, then miscounting again — the apology instant and complete, the count still wrong [2]. Asked to number the eggs, the model produced two labeled "3" and skipped 11 [1][3]; marked with red dots instead, the count fell back to 12 [3]. In one round it declared the count fixed while the image still showed 12 [3]. One user reported the same wall on paid plans, and one documented success came only on the 17th attempt [2]. The counts in circulation — 12, 9, 11, in one account 15 — come from the poster's and commenters' retellings of the screenshots and don't all match in detail; the through-line does: a first image of 12, a dozen-plus failed corrections, a text list as the fallback [2][3].

The instability is the finding. Yuanbao and Qwen, two other Chinese AI assistants, failed the same instruction for some users [1], while other testers saw Qwen and ChatGPT pass on the first try and Doubao succeed on its fourth [1]. Blogger RIAN's cross-model round concluded that none of the models drew it right; in another test reported in the same coverage, Qwen drew 15 eggs, then changed to 16 after being told empty slots were allowed [4]. Same instruction, different runs, different outcomes. The round's most revealing detail is what did pass: DeepSeek delivered 13 eggs by writing HTML code that the browser rendered — handing the counting to deterministic code instead of drawing eggs at all [4].

Image models render what counts look like — they don't count

Asked directly, almost no language model gets "13 eggs" wrong; what fails is the step that turns the number into an image [4]. Understanding a count, drawing a count and checking a count are three different tasks — and Doubao's own screenshots show the checking part failing on its own: in one round it generated three candidates, the middle one held exactly 13 eggs, and the model confidently recommended the wrong first one [4]. It is the same shape as the "finger problem" that circulated in late 2025 — a model counting five fingers on a hand drawn with six, with "six fingers" written into the prompt [2].

Diffusion models generate an image step by step out of learned visual regularities. There is no "count, then place" pass, and every step is a fresh sample: the model never carries a checklist forward to reconcile against the last version [2]. The benchmark record puts measured numbers on that ceiling. T2ICountBench (v2, January 2026) found that all state-of-the-art diffusion models fail to generate the correct number of objects, that accuracy drops significantly as the requested count rises, and that prompt refinement generally does not improve counting [5].

A study posted November 4, 2026 held the object and the starting noise fixed and varied only the requested count, then had people who never saw the request count 3,088 images: every open model tested followed the number word only up to a point — SD3.5-M and FLUX.1-dev stopped soon after ten, while FLUX.2 [klein] and Qwen-Image followed it loosely up to twenty-five [6]. NumBench (August 2026) scaled the question to 640,000 prompts across 1,600 categories and counts from 1 to 100: across five commercial systems and two open models, performance declines sharply with the requested count, and every evaluated method is weak above 50 objects [7].

Accuracy degrades continuously as counts rise [5][7] — thirteen is only where this event landed. Where the papers and the event agree is on what helps: presenting the count as a layout. In the same November study, people found the right count in 59% of images where the count was shown as a noised grid in the starting latent, against 4% when it was stated only in words [6]; NumBench found grid guidance the strongest of the guided layouts [7]; and the community's fallback in this episode — rows of 4, 4 and 5 instead of a bare 13 — points the same direction [2]. Chinese coverage of the episode also cited an ICML 2026 study from Fudan University, the Shanghai Innovation Institute and other institutions reporting that, in multimodal diffusion architectures, word-level information degrades as the network deepens, with number words among the first to fade [4]. The community's other explanation — that the egg cartons an image model learns from mostly have 12 slots, pulling 13 back toward 12 — is explicitly interpretation: no vendor has published technical detail on the failure [4].

The newest release promises stability, not counting

Nine days before the hot search, on November 9, ByteDance released Seedream 5.5, its newest image-generation model, across Doubao, Doubao Work and its Jimeng creative suite [8][9]. State-media reports of the launch list the headline claims: faster generation, stronger multi-round editing stability, native transparent assets, and better realism and small-text rendering [8][9]. By the vendor's account, the model takes a hundred consecutive rounds of edits while the untouched parts of the image stay pixel-consistent [8].

Seedream 5.5's announcement covers editing stability; the egg test probes holding a count while painting — different capabilities, and only one of them is claimed: the event's coverage called counting the one problem no product launch has solved [2]. That distinction matters for anyone picking a tool in 2026: structure, style and editing stability are the dimensions vendors compete on and claim loudly [8][9], while exact counting carries no equivalent claim in any release this episode produced [2][8].

Treat every generated count as a draft until you check it

Put together, the evidence supports a narrow answer. Drawing a specified number of objects past roughly ten is not a dependable capability of any mainstream image generator on today's record: the benchmarks measure the ceiling directly [5][6][7], and the event shows the ceiling operating inside commercial tools, with run-to-run variation wide enough that any single pass or fail can flip [1][4]. The four model names above passed one community test, one prompt, one round — a data point, not a leaderboard.

The next step costs nothing. For any image where a viewer will count, verify the output yourself before it ships, or make the count part of the layout from the start — the direction both the event's 4+4+5 rows and the benchmarks' grid-guidance results point [2][6][7]. The egg test marks which instructions image models can actually hold — and number words past roughly ten are still on the wrong side of it.