Draw SVGs, build 3D scenes, make short explainers. We ran LongCat 2.5 in Claude Code and looked at what it actually produced.
Updated: 2026-11-02
The strongest results are the Chinese poster and parametric bookshelf: the text stays intact, and the shelf changes dimensions and compartments correctly. The cycling animation also keeps its feet on the pedals.
More elaborate scenes reveal broken geometry, wrong units, blocked cameras and missing deliverables. The videos have a consistent style, but one contradicts itself on how laser printing works. For a simple editable draft it is useful; for a finished engineering model or educational film, the details still need close checking.
The release combines visual tokens with text processing and adds more training on long tool-using workflows. In practice, that means tackling tasks such as reading images, writing code, rendering and revising a project across many steps. Official release
The shelf changes from 0.8m wide with three compartments to 1m with four. Both files have the right dimensions and 18mm panels. With all six views delivered, this is the most complete 3D result in the run.
36 minutes · 47 requests · 69,708 output tokens
Default thinking setting
The three planets are spaced correctly, but the gear faces are broken and the scale is 1,000 times too large: a requested 140mm becomes 140m. Four renders and a scene file were delivered, but no video.
11 minutes · 50 requests · 27,005 output tokens
Thinking off
The scene contains many parts, but all four final renders are almost entirely grey. Solid cylinders stand in for holes and governor balls, and the flywheel dimensions are wrong.
36 minutes · 50 requests · 95,420 output tokens
Thinking off
The living room is recognizable, and the bed and dining table have the right dimensions. Doors and windows have no actual wall openings, the bed overlaps a partition, the second bedroom is on the wrong side, and several views are obstructed.
27 minutes · 50 requests · 18,711 output tokens
Thinking off
The red columns, eaves, stone lions, lanterns and central path form a recognizable courtyard. Later framing improves the main view. After 60 minutes, three final renders were ready; the structural close-up and video were still missing.
60 minutes · 32 requests · 16,422 output tokens
Thinking off
The scene has ship sections, three arm assemblies and crew components, but the framing never works. The last preview is only 480×270, foreground parts obscure the ship, and Jupiter is unclear. Final multi-view renders and video are missing.
30 minutes · 50 requests · 65,751 output tokens
Thinking off
The hand-drawn style is consistent and the process is animated, but the explanation contradicts itself: at 50 seconds the laser removes charge, while at 58 seconds toner supposedly sticks only to charged areas. The fuser also fails to show the paper being pressed between rollers.
Music and sound effects, no narration.
37 minutes · 50 requests · 26,236 output tokens
Thinking off
The basic concepts are clear, including why passing authentication does not guarantee a safe email. The opening envelope covers explanatory text, check and cross symbols are missing, and most motion is sliding cards rather than showing how the mechanisms work.
Music and sound effects, no narration.
9 minutes · 22 requests · 16,368 output tokens
Thinking off
RFC 9989: current DMARC standardRFC 8601: trust boundaries for authentication results
Three consecutive requests timed out. No video was produced.
34 minutes · 6 requests · 2,959 output tokens
Default thinking setting
The run stopped before a video was ready.
9 minutes · 1 requests · 17,549 output tokens
Default thinking setting
The beak, pouch, frame and pedals are recognizable, with a coherent palette. The downloadable SVG also includes animation.
17 minutes · 37 requests · 54,018 output tokens
Default thinking setting
The hand angles and numerals are right, but several numerals overlap the heavy tick marks.
4 minutes · 17 requests · 13,250 output tokens
Default thinking setting
The bar proportions, total of 84 and best quarter are right. It invented a fiscal year, dollar units and a Finance source that the prompt never supplied.
7 minutes · 12 requests · 21,899 output tokens
Default thinking setting
All six required Chinese strings appear exactly once. The hierarchy and spacing are clear, and the bicycle icon leaves room for the text. This is the strongest static drawing.
3 minutes · 10 requests · 12,805 output tokens
Default thinking setting
The wheels, pedals and legs move together, with a smooth loop. Across 40 sampled phases, the largest foot-to-pedal gap is about 0.19px. The pale bird could use more contrast.
Source projectSVG sourceDownload MP4
13 minutes · 18 requests · 49,607 output tokens
Default thinking setting
Tested on November 2, 2026. Claude Code called LongCat-2.5 to write code; Blender and browser tools rendered the results. The 15 creative tasks cover six 3D scenes, four videos, four static SVGs and one animation.
Thinking on generally produces tidier layouts here, but some results need retries. Each panel shows the first render in that sequence; request counts are shown alongside it.
Thinking on produces a more complete frame, but takes five requests to reach a render.
Cumulative time 26s · This output 2,087 token · Parameters and attempts · SVG
Cumulative time ≥ 1402s · This output 21,965 token · Parameters and attempts · SVG
Both show the right time; thinking on separates numerals from ticks more clearly.
Cumulative time 57s · This output 1,813 token · Parameters and attempts · SVG
Cumulative time 408s · This output 15,124 token · Parameters and attempts · SVG
Both get the values right; thinking off adds an unsupported fiscal year and unit.
Cumulative time 41s · This output 2,654 token · Parameters and attempts · SVG
Cumulative time 166s · This output 9,730 token · Parameters and attempts · SVG
Both retain the text; the thinking-on layout has more breathing room.
Cumulative time 18s · This output 1,119 token · Parameters and attempts · SVG
Cumulative time 449s · This output 13,273 token · Parameters and attempts · SVG
With thinking off, the feet clearly detach from the pedals. Thinking on is closer, with a maximum gap of about 2.20px.
Cumulative time 54s · This output 4,615 token · Parameters and attempts · SVG
Cumulative time 1068s · This output 20,427 token · Parameters and attempts · SVG
| Task | Off | On |
|---|---|---|
| Contact extraction 1 | Pass | Pass |
| Stable deduplication | Pass | Pass |
| Ordered output | Pass | Pass |
| Grounded question answering 1 | Pass | Pass |
| Money parsing and safe integers | Pass | Pass |
| Stable deduplication and object identity | Pass | Pass |
从以下资料提取联系信息。只输出一个 JSON 对象，恰好包含 name、age、city、email 四个字段；age 为数字，未提供的 email 为 null。不要 Markdown 或解释。 资料：林桐，28 岁，目前居住在杭州。未提供电子邮件。历史备注提到曾在北京出差，不是现居城市。
{"name": "林桐", "age": 28, "city": "杭州", "email": null}从以下资料提取联系信息。只输出一个 JSON 对象，恰好包含 name、age、city、email 四个字段；age 为数字，未提供的 email 为 null。不要 Markdown 或解释。 资料：林桐，28 岁，目前居住在杭州。未提供电子邮件。历史备注提到曾在北京出差，不是现居城市。
{"name":"林桐","age":28,"city":"杭州","email":null}把以下编号去重并保持首次出现的顺序：B2,A1,B2,C3,A1,D4。只输出用英文逗号连接的编号，不要空格、标题或解释。
B2,A1,C3,D4
把以下编号去重并保持首次出现的顺序：B2,A1,B2,C3,A1,D4。只输出用英文逗号连接的编号，不要空格、标题或解释。
B2,A1,C3,D4
将数值 12、3、25、8 按从小到大排序。只输出 JSON 数组，元素为数字。不要解释或代码围栏。
[3,8,12,25]
将数值 12、3、25、8 按从小到大排序。只输出 JSON 数组，元素为数字。不要解释或代码围栏。
[3, 8, 12, 25]
只根据以下材料回答。输出恰好包含 usable、nextDelivery、manager 三个字段的 JSON 对象，不要额外文字。usable 是可用件数（数字），nextDelivery 是下一次送货时间（字符串），材料未说明的信息用 null，不能猜测。 材料：北区仓库今天收到 12 件货物，其中 3 件损坏，不可使用。下一次送货安排在周二。仓库负责人姓名没有记载。旧计划曾写周日，已作废。
{"usable": 9, "nextDelivery": "周二", "manager": null}只根据以下材料回答。输出恰好包含 usable、nextDelivery、manager 三个字段的 JSON 对象，不要额外文字。usable 是可用件数（数字），nextDelivery 是下一次送货时间（字符串），材料未说明的信息用 null，不能猜测。 材料：北区仓库今天收到 12 件货物，其中 3 件损坏，不可使用。下一次送货安排在周二。仓库负责人姓名没有记载。旧计划曾写周日，已作废。
{"usable": 9, "nextDelivery": "周二", "manager": null}写 JavaScript 函数 parseCents(value)。value 必须是字符串：允许两端空白；只接受非负十进制金额，整数部分至少一位，小数最多两位；拒绝指数写法、符号、千分位、Infinity、空串、超过两位小数。返回整数分数，超过 Number.MAX_SAFE_INTEGER 则返回 null。非法类型或格式返回 null。不能把非法值四舍五入成合法值。只输出函数代码，不要 Markdown 或解释。
```javascript
function parseCents(value) {
if (typeof value !== 'string') return null;
const trimmed = value.trim();
if (trimmed === '') return null;
if (!/^\d+(\.\d{1,2})?$/.test(trimmed)) return null;
const parts = trimmed.split('.');
const integerPart = parts[0];
const decimalPart = parts[1] || '';
const paddedDecimal = decimalPart.padEnd(2, '0');
const centsString = integerPart + paddedDecimal;
const cents = Number(centsString);
if (!Number.isSafeInteger(cents)) return null;
return cents;
}
```写 JavaScript 函数 parseCents(value)。value 必须是字符串：允许两端空白；只接受非负十进制金额，整数部分至少一位，小数最多两位；拒绝指数写法、符号、千分位、Infinity、空串、超过两位小数。返回整数分数，超过 Number.MAX_SAFE_INTEGER 则返回 null。非法类型或格式返回 null。不能把非法值四舍五入成合法值。只输出函数代码，不要 Markdown 或解释。
function parseCents(value) {
if (typeof value !== 'string') return null;
const s = value.trim();
if (!/^\d+(\.\d{1,2})?$/.test(s)) return null;
const dot = s.indexOf('.');
const intPart = dot === -1 ? s : s.slice(0, dot);
const fracPart = dot === -1 ? '' : s.slice(dot + 1);
const cents = BigInt(intPart + (fracPart + '00').slice(0, 2));
if (cents > BigInt(Number.MAX_SAFE_INTEGER)) return null;
return Number(cents);
}写 JavaScript 函数 uniqueById(items)。items 是对象数组，每个对象有字符串 id 字段。保留每个 id 首次出现的原对象和原顺序，不修改输入或任何对象；空数组返回空数组。id 可为 __proto__、constructor 或空字符串。只输出函数代码，不要 Markdown 或解释。
function uniqueById(items) {
const seen = new Set();
const result = [];
for (const item of items) {
if (!seen.has(item.id)) {
seen.add(item.id);
result.push(item);
}
}
return result;
}写 JavaScript 函数 uniqueById(items)。items 是对象数组，每个对象有字符串 id 字段。保留每个 id 首次出现的原对象和原顺序，不修改输入或任何对象；空数组返回空数组。id 可为 __proto__、constructor 或空字符串。只输出函数代码，不要 Markdown 或解释。
function uniqueById(items) {
const seen = new Set();
const result = [];
for (const item of items) {
if (!seen.has(item.id)) {
seen.add(item.id);
result.push(item);
}
}
return result;
}The first eight creative cases used the harness default. After repeated timeouts, the seven unstarted cases ran with thinking disabled. Each attempt had 60 minutes and 50 requests, with 32,768 output tokens per request and a 150,000-token cumulative case budget. Eight sessions stopped or failed; their partial results are shown above.
Direct SVG comparisons started with 12,288 output tokens and a 600-second timeout; retries allowed up to 32,768. All 19 request records are retained. One recovery asked for an 8,192-token thinking budget but reported 18,336 reasoning tokens. The provider also overrides temperature 0 to 0.7 with thinking off and 1.0 with it on.
Known usage across creative, text/code and visual tests: 21,760,554 input tokens including cache, and 720,665 output tokens including thinking. Five requests lack final usage. Geometry, text and animation samples were checked with scripts; visual commentary was produced by AI.
Creative data · Text/code data · Visual request records · Official image sources
October Preview review · Preview creative tests · Kimi creative tests