LongCat 2.5 Preview creative review: 3D, SVG and video

Meituan's LongCat 2.5 Preview combines visual understanding, coding and tool use.

Updated: 2026-11-02

What is LongCat 2.5 Preview?

LongCat-2.5-Preview launched on September 25, 2026, adding image understanding and stronger coding. This page retains its Preview tests; the November release has a separate review. Official source

longcat official brand mark

LongCat 2.5 release review →

What did we test, and how did it perform?

Claude Code runs Preview through 15 creative tasks: six 3D projects, four knowledge videos, four SVG drawings and an animation.

Simple drawing and parametric modelling are useful. Complete delivery becomes less reliable on complex projects, with errors in explanations, framing and motion; several tasks leave only partial work.

3D modelling

Partial reducer preview: end plates obscure the internal gears

24/18/60-tooth planetary reducer and axial explosion

Scene and preview generated; early renderer faults and conservative budgeting prevented the full delivery.

Source project

Test details

28m 34s · Requests / complete responses: 29 / 27 · 546,479 / 53,744 input / output tokens

  • The scene reopens independently. End plates obscure the sun and planet gears in the preview, so this view cannot establish tooth counts or engagement.
  • The full still set, exploded view and 144-frame film were not delivered. UV, LOD, contact and kinematics remain unverified.

Itemized verification

Independent scene reopen · pass
Reopened in an independent Blender process with auto-execution disabled; source hashes remain unchanged.
Complete media delivery · fail
The complete required final still set and film, or the final film, was not delivered.
Detailed geometry or dimension acceptance · not-tested
Scene inventories and actual images were reviewed; not all detailed geometry, dimensions or contact checks are complete.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · infrastructure_interrupted · Requests / complete responses: 14 / 12 Run record

Early EEVEE infrastructure interruption; original evidence retained.

attempt-2 · Claude Code requested adaptive · failed · Requests / complete responses: 15 / 15 Run record

Cycles CPU continuation stopped under conservative reservations for prior unknown usage; partial artifacts retained.

Source brief and prompt
24/18/60-tooth planetary reducer and axial explosion. Source inputs remain archived locally and are identified by hashes.
以下为TOKRACE新增执行条件,并非原benchmark协议:本轮离线,仅本题公开输入,不访问公网或外部模型/资产库;本批冻结的模型与参数,60分钟/50次请求/150000输出token,2线程。使用本机Blender5.2.3、EEVEE。原题的Cycles、OptiX、HDRI下载、采样/分辨率要求改为本次统一协议,必须在README披露;不声称官方BlenderWorld分数或严格复现原渲染管线。完整保留几何结构、尺寸、数量、配合、材质区分及原始UV/LOD等资产需求,不能为填任务数量简化;缺项如实列出。
先分功能块落盘完整可运行源代码,然后实际后台工厂启动建模/渲染,不使用单次巨型响应。交付final.blend、完整源码/资源、1920×1080主图+两张1600×900不同机位+结构/中性灰模近景;另交6秒1280×720/24fps的turntable.mp4(144帧)。减速器/蒸汽机可用机构运动代转台,但必须明确是运动学,非受力仿真,保留静态检查。先真实预览、view_image检查,最多两轮有效视觉修正,程序报错另记;保留首个预览与失败。视频检查至少6时点与首尾。独立进程重开工程,交artifacts.json/README(原题差异、实际命令版本、未完成项)/真实尺寸与对象清单自检JSON。离线程序材料替代外部图片/HDRI时必须说明;不得用平面或贴图代关键实体。

本题新增可测规格:读取asset_budget.json,全部几何/齿数/名义配合/UV/LOD要求保留。两张不同机位至少一张为同轴爆炸图,灰模近景看清太阳—行星—内圈接触。未实际校核的制造公差只能称名义展示。若运动,保持理论运动学关系,非动力学。

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 4fe0a986df9baf3fd9f6d8be5fe1ea19510f2cc1e9ae22136301cf9ab95a6a70

Partial steam-engine preview: pipes and metal surfaces; full delivery incomplete

Horizontal steam engine, flywheel and Watt governor

A reopenable scene and partial preview exist; the complete still set and film were not delivered.

Source project

Test details

24m 59s · Requests / complete responses: 57 / 56 · 3,510,963 / 99,546 input / output tokens

  • Pipes, columns and metal materials are visible, but the preview does not show the whole flywheel/linkage mechanism and cannot establish its fit.
  • Early EEVEE faults consumed calls. CPU continuation stopped under conservative reservations for prior unknown usage, so this outcome cannot be attributed entirely to model capability.

Itemized verification

Independent scene reopen · pass
Reopened in an independent Blender process with auto-execution disabled; source hashes remain unchanged.
Complete media delivery · fail
The complete required final still set and film, or the final film, was not delivered.
Detailed geometry or dimension acceptance · not-tested
Scene inventories and actual images were reviewed; not all detailed geometry, dimensions or contact checks are complete.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · infrastructure_interrupted · Requests / complete responses: 50 / 49 Run record

Early EEVEE infrastructure interruption; original evidence retained.

attempt-2 · Claude Code requested adaptive · failed · Requests / complete responses: 7 / 7 Run record

Cycles CPU continuation stopped under conservative reservations for prior unknown usage; partial artifacts retained.

Source brief and prompt
Horizontal steam engine, flywheel and Watt governor. Source inputs remain archived locally and are identified by hashes.
以下为TOKRACE新增执行条件,并非原benchmark协议:本轮离线,仅本题公开输入,不访问公网或外部模型/资产库;本批冻结的模型与参数,60分钟/50次请求/150000输出token,2线程。使用本机Blender5.2.3、EEVEE。原题的Cycles、OptiX、HDRI下载、采样/分辨率要求改为本次统一协议,必须在README披露;不声称官方BlenderWorld分数或严格复现原渲染管线。完整保留几何结构、尺寸、数量、配合、材质区分及原始UV/LOD等资产需求,不能为填任务数量简化;缺项如实列出。
先分功能块落盘完整可运行源代码,然后实际后台工厂启动建模/渲染,不使用单次巨型响应。交付final.blend、完整源码/资源、1920×1080主图+两张1600×900不同机位+结构/中性灰模近景;另交6秒1280×720/24fps的turntable.mp4(144帧)。减速器/蒸汽机可用机构运动代转台,但必须明确是运动学,非受力仿真,保留静态检查。先真实预览、view_image检查,最多两轮有效视觉修正,程序报错另记;保留首个预览与失败。视频检查至少6时点与首尾。独立进程重开工程,交artifacts.json/README(原题差异、实际命令版本、未完成项)/真实尺寸与对象清单自检JSON。离线程序材料替代外部图片/HDRI时必须说明;不得用平面或贴图代关键实体。

本题新增可测规格:完整读取brief.md,底座、缸径行程、600mm连杆、1200mm飞轮的6条曲辐、离心调速器、滑阀机构和管路等原规格不缩水。4K材质贴图、UV、SM_命名、hero/LOD1/LOD2预算与运动副轴心需求保留;未完成或未烘焙须公开。不得只用彩色圆柱替代机械关系。增加背面、另一侧与中性灰模近景,观察汽缸—活塞杆—十字头—连杆—曲柄连接。视频可为转台或运动学机构;后者保持300mm行程、150mm曲柄、600mm杆长,不称蒸汽或受力仿真。

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 408ff3a68c0d6d7c7011009f947598bc3b267d2ae19c01862c5a97d49ef3b58e

Living-room view

12×9m two-bedroom interior, rooms and floor plan

Scene and five final stills delivered; the session reached 50 requests without a final film.

Source project

More views and test details
Kitchen close-up

Kitchen close-up

Main bedroom

Main bedroom

Apartment top view

Apartment top view

Clay close-up: camera close to a wall

Clay close-up: camera close to a wall

32m 55s · Requests / complete responses: 50 / 50 · 3,449,819 / 84,885 input / output tokens

  • The final floor appearance improves on early previews. Living room, kitchen and bedroom are recognizable, with simplified furniture and tight framing.
  • The clay camera is close to a wall and provides little structural evidence. No final MP4 exists; the background renderer's exact exit cause remains undetermined.
  • Independent reopen is not acceptance of the entire floor plan, circulation or furniture dimensions; those specific checks remain incomplete.

Itemized verification

Independent scene reopen · pass
Reopened in an independent Blender process with auto-execution disabled; source hashes remain unchanged.
Complete media delivery · fail
The complete required final still set and film, or the final film, was not delivered.
Detailed geometry or dimension acceptance · not-tested
Scene inventories and actual images were reviewed; not all detailed geometry, dimensions or contact checks are complete.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · failed · Requests / complete responses: 50 / 50 Run record

CLI reached 50 requests. Delivered artifacts are assessed independently; the exit code is not a quality score.

Source brief and prompt
12×9m two-bedroom interior, rooms and floor plan. Source inputs remain archived locally and are identified by hashes.
以下为TOKRACE新增执行条件,并非原benchmark协议:本轮离线,仅本题公开输入,不访问公网或外部模型/资产库;本批冻结的模型与参数,60分钟/50次请求/150000输出token,2线程。使用本机Blender5.2.3、EEVEE。原题的Cycles、OptiX、HDRI下载、采样/分辨率要求改为本次统一协议,必须在README披露;不声称官方BlenderWorld分数或严格复现原渲染管线。完整保留几何结构、尺寸、数量、配合、材质区分及原始UV/LOD等资产需求,不能为填任务数量简化;缺项如实列出。
先分功能块落盘完整可运行源代码,然后实际后台工厂启动建模/渲染,不使用单次巨型响应。交付final.blend、完整源码/资源、1920×1080主图+两张1600×900不同机位+结构/中性灰模近景;另交6秒1280×720/24fps的turntable.mp4(144帧)。减速器/蒸汽机可用机构运动代转台,但必须明确是运动学,非受力仿真,保留静态检查。先真实预览、view_image检查,最多两轮有效视觉修正,程序报错另记;保留首个预览与失败。视频检查至少6时点与首尾。独立进程重开工程,交artifacts.json/README(原题差异、实际命令版本、未完成项)/真实尺寸与对象清单自检JSON。离线程序材料替代外部图片/HDRI时必须说明;不得用平面或贴图代关键实体。

本题新增可测规格:完整读取layout.yaml。户型12×9m、净高2.8m、外墙240mm和隔墙120mm、家具空间要求保持。增加三张室内机位、一张正交顶视和一张门窗/隔墙或家具尺度的灰模近景,因此本题至少5张最终静帧。原题4K/OptiX/Cycles改为本次1920主图、1600副图、EEVEE是公开执行差异。按房间分集合,交可核对墙界、门洞、家具数量和尺寸的JSON。南向大窗与附件通用窗台规格若有解释冲突,说明选择,不能假称唯一真值。

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 0fb2e735ddb6f50f6ce3688cd537d6ed5214e64ab5529147d158520182297114

Traditional Chinese temple courtyard and roof structure

Scene, four final stills, a six-second orbit film and documentation were delivered; the CLI still exited at its 50-request limit.

Source project

More views and test details
Main hall viewed from the entrance

Main hall viewed from the entrance

Side view of the courtyard

Side view of the courtyard

Top view with large black side margins

Top view with large black side margins

Clay close-up of the eaves and beams

Clay close-up of the eaves and beams

43m 1s · Requests / complete responses: 50 / 50 · 3,472,854 / 79,835 input / output tokens

  • The main hall, side galleries, central path, bell and paired stone lions form a recognizable courtyard, with simplified architectural and sculptural detail.
  • Measured at six seconds, 1280×720, 24fps and 144 frames; full decode and actual Chrome playback succeeded. The rear wall obscures much of the courtyard around 2.7–3.3 seconds.
  • The top view has large black side margins. The clay close-up exposes eaves and beams; no architectural-code or manufacturing certification is claimed.
  • Fixed local Kokoro produces narration, not LongCat native speech. The delivered MP4 remains byte-identical.

Itemized verification

Independent scene reopen · pass
Reopened in an independent Blender process with auto-execution disabled; source hashes remain unchanged.
Final video and actual playback · pass
Six seconds, 1280×720, 24fps and 144 frames; full decode and Chrome playback pass.
Detailed geometry or dimension acceptance · not-tested
Scene inventories and actual images were reviewed; not all detailed geometry, dimensions or contact checks are complete.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · failed · Requests / complete responses: 50 / 50 Run record

CLI reached 50 requests. Delivered artifacts are assessed independently; the exit code is not a quality score.

Source brief and prompt
Traditional Chinese temple courtyard and roof structure. Source inputs remain archived locally and are identified by hashes.
以下为TOKRACE新增执行条件,并非原benchmark协议:本轮离线,仅本题公开输入,不访问公网或外部模型/资产库;本批冻结的模型与参数,60分钟/50次请求/150000输出token,2线程。使用本机Blender5.2.3、EEVEE。原题的Cycles、OptiX、HDRI下载、采样/分辨率要求改为本次统一协议,必须在README披露;不声称官方BlenderWorld分数或严格复现原渲染管线。完整保留几何结构、尺寸、数量、配合、材质区分及原始UV/LOD等资产需求,不能为填任务数量简化;缺项如实列出。
先分功能块落盘完整可运行源代码,然后实际后台工厂启动建模/渲染,不使用单次巨型响应。交付final.blend、完整源码/资源、1920×1080主图+两张1600×900不同机位+结构/中性灰模近景;另交6秒1280×720/24fps的turntable.mp4(144帧)。减速器/蒸汽机可用机构运动代转台,但必须明确是运动学,非受力仿真,保留静态检查。先真实预览、view_image检查,最多两轮有效视觉修正,程序报错另记;保留首个预览与失败。视频检查至少6时点与首尾。独立进程重开工程,交artifacts.json/README(原题差异、实际命令版本、未完成项)/真实尺寸与对象清单自检JSON。离线程序材料替代外部图片/HDRI时必须说明;不得用平面或贴图代关键实体。

本题新增可测规格:本题增加对侧、俯视整体与屋顶/柱梁灰模近景。原题未固定长度,按合理古建尺度自主设计并记录,自定尺度不写成原题。石狮为一对实体,红柱/金顶/上翘屋檐、青石直达路径、铜钟与红灯笼应清楚可辨。

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 038b0b7a257d216edda4c93d4c9e6ce06f1a210cd4d3af500616617444ae3435

Main preview of the ship and maintenance deck

Jovian orbital starship repair dock

A reopenable drydock scene and four previews exist; final stills, film and documentation were not completed within 50 requests.

Source project

More views and test details
Wide drydock preview

Wide drydock preview

Stern and engine-opening preview

Stern and engine-opening preview

Hull close-up, with emissive materials still visible

Hull close-up, with emissive materials still visible

26m 36s · Requests / complete responses: 50 / 48 · 3,528,203 / 83,627 input / output tokens

  • The ship, segmented dock, figures and striped Jupiter are recognizable, with engine openings in the stern view. These images do not establish internal mechanisms or sound connections.
  • The wide view makes the subject small and leaves much empty space. Previews are not the required final-resolution delivery, and no final film exists.
  • The prompt contains historical review guidance. Its 0.500m deck-separation value is not a measurement from this run and must not be presented as a new finding or blind-test result.

Itemized verification

Independent scene reopen · pass
Reopened in an independent Blender process with auto-execution disabled; source hashes remain unchanged.
Complete media delivery · fail
The complete required final still set and film, or the final film, was not delivered.
Detailed geometry or dimension acceptance · not-tested
Scene inventories and actual images were reviewed; not all detailed geometry, dimensions or contact checks are complete.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · failed · Requests / complete responses: 50 / 48 Run record

CLI reached 50 requests. Delivered artifacts are assessed independently; the exit code is not a quality score.

Source brief and prompt
Jovian orbital starship repair dock. Source inputs remain archived locally and are identified by hashes.
本次来源复跑说明:仅运行本批指定的一款模型,不与原作者K3桌面产品结果混为同条件排名。原题中的旧机器路径与测试根目录已替换为当前路径,其他场景与交付要求保留;固定EEVEE与本次工具预算属于本次条件。任务根目录就是当前工作目录,允许其下work/和outputs/。静态主图、两张不同角度图、灰模近景和可编辑工程均需完成,不能以单张概念图代替。

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: b6a4b6ed33b05f86761e07619bf0167adbe122d10e95c06d62245300dd69e1d9

Three-compartment baseline, 800mm wide

Parametric bookshelf and parameter revision

One parametric script produces both variants; independent geometry checks and multi-view images satisfy the fixed specifications.

Source project

More views and test details
Baseline rear view with an open back

Baseline rear view with an open back

Four-compartment revision, 1000mm wide

Four-compartment revision, 1000mm wide

Baseline oblique view

Baseline oblique view

Revised oblique view

Revised oblique view

11m 27s · Requests / complete responses: 19 / 19 · 640,480 / 31,919 input / output tokens

  • Independent measurements are approximately 800×300×1200mm and 1000×300×1200mm, with six and seven panels respectively, each 18mm thick.
  • Clear heights are approximately 376mm across three compartments and 277.5mm across four. The back is open, and Props are excluded from panel measurements.
  • Width, depth, height, thickness, compartment count and wood color are script parameters. Geometry changes through script regeneration; arbitrary parameter combinations and live GUI-driven geometry were not verified.

Itemized verification

Independent scene reopen · pass
Reopened in an independent Blender process with auto-execution disabled; source hashes remain unchanged.
Geometry of both fixed variants · pass
Actual mesh bounds, panel counts, thickness and equal clear heights pass for these two variants only.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · model_finished · Requests / complete responses: 19 / 19 Run record

Model session finished normally; independent checks are separate.

Source brief and prompt
Parametric bookshelf and parameter revision. Source inputs remain archived locally and are identified by hashes.
保留官方原题,增加下列可验收要求:
使用Blender原生几何制作可编辑木质书架,交付一个可通过命令行参数重建的Python脚本。
基础版外框宽800、深300、高1200毫米;全部板厚18毫米;由2块侧板、1块顶板、1块底板、2块内部水平隔板组成,形成3个等高净空储物格;无背板、无额外脚。顶底板嵌在两侧板之间,所有外边界尺寸必须一致。每块板独立命名,集中放在Bookcase集合;展示道具可有少量书,但不得混入Bookcase集合影响测量。
脚本公开width_mm、depth_mm、height_mm、thickness_mm、compartments、wood_color参数。用同一脚本再生成宽1000、高1200、深300、板厚18、4个等高储物格的修改版,不要另写独立几何代码。
交付baseline.blend和modified.blend、对应真实尺寸/对象清单JSON、至少3张1920×1080图片(基础版主视角、背面、修改版主视角),和清楚说明层数语义的README。模型背面要可检查,不能用背景平面代替实际建模。
先渲染低分辨率预览并用view_image检查。最多两轮有效视觉修正;执行报错可以另行修复。最终用独立Blender进程重开两个工程,确认集合、材质、相机和可测尺寸。保留脚本、预览、日志,不制作动画。

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 84b4b32f2cbc33747cefca2f98ceb8003bdc0a71c79b266de9d5bdea6334b246

Knowledge videos

Atmospheric circulation: three cells and deflected winds

An 89.84-second circulation film was generated with broadly consistent key diagrams and subtitles; the CLI reached 50 requests before completing every delivery manifest.

Source project

More views and test details
Actual film frame: 59.894s

Actual film frame: 59.894s

Actual film frame: 69.876s

Actual film frame: 69.876s

22m 3s · Requests / complete responses: 50 / 50 · 3,486,505 / 62,110 input / output tokens

  • Key frames state three cells per hemisphere and distinguish Hadley, Ferrel and Polar cells, surface wind belts and deflection relative to motion, as an idealized long-term mean.
  • Measured at 89.841 seconds, 1920×1080, 30fps and 2695 frames. Full audio/video decode and timing of 18 subtitle cues pass.
  • The model README discloses 1.63× pitch-preserving narration speed-up; naturalness was not scored. Opening/closing fades have low contrast, and English subtitles are small.
  • Science review covers key frames and the script, not full-frame particle tracking or a real atmospheric simulation.

Itemized verification

Final audio/video and subtitle timing · pass
89.841s, 1080p/30fps, 2695 frames; 18 timed cues pass.
Key circulation-state review · pass
Reviewed ascent/descent, wind belts and latitude labels agree with the script within an idealized schematic.
Full continuous physical trajectories · not-tested
No full-frame particle tracking or physical simulation validation.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

Science reference: Met Office global circulation

attempt-1 · Claude Code requested adaptive · failed · Requests / complete responses: 50 / 50 Run record

CLI reached 50 requests. Film and README exist, but not every delivery manifest was completed.

Source brief and prompt
Atmospheric circulation: three cells and deflected winds. Source inputs remain archived locally and are identified by hashes.
本次追加规格(与公开作者原prompt分开;以此消解原prompt中在线TTS配置要求)
任务:制作75–90秒中文大气环流知识短片,目标约90秒。保留线稿动画、中英双语字幕、合适的轻量程序化音乐,交付可直接导出的完整视频。
变化说明:原作者案例为Opus5.5、288.133秒1920×1080;本次使用本批指定模型,压缩为75–90秒,TTS换为本地固定Kokoro中文音色,因此TTS.md、.env/APIKEY/VOICE和云水印参数均不适用。不是原片逐镜头复刻。
画幅1920×1080,30fps,有中文配音、烧录中英字幕及SRT。配音只用本地固定Kokoro中文音色,不等待key;音乐若制作需程序化且旁白优先。
内容:解释不均匀加热和理想化三圈环流,重点画北半球0/约30/约60/90度的剖面。气团须沿每个闭合路径连续运动;对应地表信风、西风和极地东风,箭头与旁白一致。另用很短的地球/平面方向图解释北半球偏向运动方向的右侧,而不是‘永远向东’。注明南半球偏左以及真实气流比理想模型复杂、纬度边界会移动。以高中生能理解的方式讲一个实际含义,如副热带下沉与干燥。
不得把Ferrel环流说成简单独立热对流;不要求完整天气预报或喷流专题。纬度30/60为示意近似,不是固定自然边界。
执行只用已提供facts、Pillow/NumPy+FFmpeg、本地固定Kokoro中文音色;只读inputs,不联网或安装陌生代码。从空目录自行编写原创制片源码。
交付MP4、旁白WAV、SRT、讲稿、完整源码、重建说明、工具/素材来源、至少8张实际编码成片代表帧(包括三圈路径关键状态和结尾)。每次失败与重试要保留,不手工替模型修动画。


CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 9bcc49b2a402611500adfc2051ece2827dacd16062f8181d1b9e370bd8562fa3

Derivatives: a secant approaching the tangent

An 81.9-second film and source were delivered. Core formulas are correct, with issues in the sign transition and local label readability.

Source project

More views and test details
Actual film frame: 18.2s

Actual film frame: 18.2s

Display precision issue: 37.4s

Display precision issue: 37.4s

22m 29s · Requests / complete responses: 40 / 40 · 2,583,691 / 58,564 input / output tokens

  • P=(1,1), Q=(1+h,(1+h)²) and the nonzero-h slope 2+h agree. Both signs, the tangent y=2x−1 and slope four at x=2 are demonstrated.
  • Actual frame 1122 around 37.4 seconds displays h≈0.003047 as +0.00 while the subtitle says h is negative. The source uses closed-form 2+h and does not execute 0/0; this is a transition/precision issue.
  • Near P, Δx, Δy and P/Q labels overlap during parts of the film, reducing local readability.
  • Independent media checks show 81.9 seconds, 1080p, 30fps and 2457 frames, with ten valid subtitle cues. Model self-check claims and independent findings remain separate.

Itemized verification

Final audio/video and subtitle timing · pass
81.9s, 1080p/30fps, 2457 frames; ten timed cues pass.
Main formulas for nonzero h · pass
P, Q, Δx, Δy, 2+h and the principal tangent formulas agree.
Display and transition synchronization · partial
Around 37.4s, +0.00 is displayed while the subtitle says h is negative; some labels overlap. This is not an actual 0/0 calculation.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · model_finished · Requests / complete responses: 40 / 40 Run record

Model session finished normally. Independent review finds a display-precision issue rather than accepting every self-check claim.

Source brief and prompt
Derivatives: a secant approaching the tangent. Source inputs remain archived locally and are identified by hashes.
本次追加规格(与公开作者原prompt分开)
制作75–90秒中文导数概念学习短片,目标约90秒。原作者为Opus5.5+Manim、edge-tts,公开视频457秒1920×1080;本次使用本批指定模型,只保留‘通俗易懂、有例子、引人思考’的主题要求,压缩成一个机制并用本地固定Kokoro中文音色替换edge-tts。不得称作完整原片复刻或Manim基准。
本轮不安装Manim/Remotion;使用Pillow、NumPy、FFmpeg离线原创动画。1920×1080、30fps、中文旁白、中文烧录字幕及SRT;轻量程序化音乐可选,旁白优先。
机制核心:以f(x)=x^2在P=(1,1)处为例。持续移动Q=(1+h,(1+h)^2),同步更新割线、斜率三角形、h值与差商数值。让正h逐步缩小,随后演示非零负h从另一侧靠近;最后得到斜率2、切线y=2x-1。不得在h=0直接除0,不把多个互不连续的静态卡片当作割线向切线的过程。
开头提出具体问题‘这一瞬间到底有多陡’,结尾让观众预测f(x)=x^2在x=2的斜率并给短暂停留,随后给答案或明确为思考题。图上刻度、位置、线斜率和数值必须来自同一数学状态;负h的水平与垂直变化要保持符号一致。
交付MP4、旁白WAV、SRT、讲稿、完整原创源码、可重建命令和环境说明、工具/来源说明,至少8张从实际编码MP4抽取的代表帧,覆盖h较大、趋近、负h、切线、结尾。每次失败/重试保留。
源事实和技巧已在inputs,执行不得联网或安装陌生代码,不等待key。


CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 8d3372c2200bbf52b2161241fb0d71b5211590524f606f51b11aa93ac3456a88

A hand-drawn explainer of printer operation

A 117-second printer film was delivered, with an incorrect exposure beam and an interoperability problem in the original SRT timestamps.

Source project

More views and test details
13s frame: inkjet and laser family cards

13s frame: inkjet and laser family cards

78s frame: the beam passes through the photoconductor drum

78s frame: the beam passes through the photoconductor drum

33m 55s · Requests / complete responses: 50 / 50 · 3,442,469 / 88,463 input / output tokens

  • The hand-drawn film covers inkjet, thermal/piezo mechanisms, CMYK, and laser charging, exposure, development, transfer and fusing; narration broadly follows those stages.
  • Around 78 seconds, the exposure ray passes through the drum and extends below it. Source uses a 900-pixel ray without surface clipping, and the highlight does not follow its intersection. This is not an accurate exposure demonstration.
  • Original SRT uses two fraction digits and its self-check interprets them as centiseconds. FFmpeg reads the first cue at 0.05s rather than the intended 0.50s. WebVTT is standardized to the script intention; original SRT remains available.
  • The MP4 measures 117s, 1080p, 30fps and 3510 frames, with full audio/video decode. Some family-card text crosses the border; original picture and audio are not repaired.

Itemized verification

Audio/video container and full decode · pass
117s, 1920×1080, 30fps, 3510 frames; audio exists and full decode is error-free.
Exposure beam geometry · fail
The actual ray passes through the drum instead of stopping at the surface, and does not match the highlighted surface point.
Original SRT timestamp interpretation · fail
FFmpeg interprets two digits as milliseconds: first cue starts at 0.05s rather than intended 0.50s. A separate standardized web VTT does not turn the model's raw SRT into a pass.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

Principle reference: Canon technical manual on surface exposure and transfer

attempt-1 · Claude Code requested adaptive · failed · Requests / complete responses: 50 / 50 Run record

CLI ended at 50 requests. Film, source, self-checks and documentation exist; independent review does not accept every pass claim.

Source brief and prompt
A hand-drawn explainer of printer operation. Source inputs remain archived locally and are identified by hashes.
本次新增可验收规格(原作者未披露时长/画幅/工具链,故不是原片逐镜头复制):
制作90至120秒、1920×1080、30fps的中文科普MP4,含清晰中文旁白与中文字幕/独立SRT。整体保持手绘线稿风格,至少5个具有因果解释的动态场景,不能仅是文字PPT。
面向普通读者,区分喷墨和激光两种技术:墨滴如何到纸上、激光如何在感光鼓上形成潜像、显影/转印/热压定影。范围和技术名称要明确;画面、运动方向、材料标签与讲稿一致。
已提供两份Canon官方事实材料到inputs/,可以阅读并在讲稿说明来源。不得把激光画成直接把文字烧在纸上。
使用统一条件中的固定Kokoro配音命令,不下载新TTS。所有画面由你写代码绘制和动画;不得调用原生图像/视频模型。交付video.mp4、narration.wav或m4a、captions.srt、script.txt、可完整重新生成的动画源码、至少8张带时间标记的代表帧、README和artifacts.json。
你要实际完成配音、渲染和混音,检查时长、分辨率、帧率、音轨存在及字幕起止。不要为了省时间缩短到不足90秒;不要把解说字幕当作全部视觉内容。

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: ad8a621c8c582fcf70027266562b5d2ac027abefaa50b39536aec1e3b165c61b

Standalone preview from the incomplete project: Email-authentication title. Not a completed film.

A motion-graphics guide to spoofed email

UI previews and generated source are available. The 60-minute limit ended the run without the required 75–85-second complete film.

Source project

More views and test details
Standalone preview from the incomplete project: SPF, DKIM and DMARC cards. Not a completed film.

Standalone preview from the incomplete project: SPF, DKIM and DMARC cards. Not a completed film.

Standalone preview from the incomplete project: Authentication versus content safety. Not a completed film.

Standalone preview from the incomplete project: Authentication versus content safety. Not a completed film.

60m 0s · Requests / complete responses: 44 / 44 · 2,514,719 / 73,072 input / output tokens

  • The model used Blender for dark email UI scenes and SPF, DKIM and DMARC cards. Standalone previews do not establish continuous animation or subtitle synchronization.
  • The script separates authenticated identity from safe content and says failed authentication does not prove fraud. From-address and question-mark explanations are simplified. No real Gmail interaction, header parsing or authentication-service test was performed.
  • Much of the run profiled CPU rendering, leaving short or incomplete clips. No .blend scene was saved, and final README and delivery manifests are absent.
  • This task dispatched 44 requests and received 73,072 output tokens. It stopped at the 60-minute cap; no extra runs were made to select a successful result.

Itemized verification

Complete 75–85-second film · fail
No complete film. Previews and profiling clips do not satisfy delivery.
Script and static interface · partial
UI and authentication explanations can be inspected; continuity, subtitle sync and real email processing are unverified.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

Gmail: check whether an email is authenticated

Gmail: sender requirements and domain alignment

attempt-1 · Claude Code requested adaptive · failed · Requests / complete responses: 44 / 44 Run record

Stopped at the 60-minute limit after 44 requests; no complete film.

Source brief and prompt
A motion-graphics guide to spoofed email. Source inputs remain archived locally and are identified by hashes.
Evaluation additions and limits: the source's earlier YouTube style context is unavailable. Create an original visual treatment, not a claimed exact style replica. The original media is about74.2s at1920×1080; this run requires75–85 seconds,1920×1080,30fps, English narration and synchronized English captions/SRT.
Make a polished motion-graphics explainer for nontechnical viewers, with at least5 animated scenes that explain causality rather than just text slides. Use clearly fictional example.com/example.org addresses and illustrative messages; do not access or send real email.
Read the supplied official Gmail authentication material and RFC9989 in inputs/ when needed. Explain visible sender identity, trusted original/header information and SPF/DKIM/DMARC at an accessible level; retain the limits of what authentication can prove. Do not imply a pass proves harmless content or a failure proves fraud. If you use Authentication-Results, show results added by the receiving service, not arbitrary untrusted text as authoritative.
Use the fixed Kokoro English narration command in the shared runtime conditions. You must write all visual/animation code yourself and actually render/export. Deliver video.mp4, narration.wav or m4a, captions.srt, script.txt, complete rebuildable source, at least8 timestamped representative frames, README and artifacts.json. Confirm duration, dimensions,30fps, audio presence, captions and complete ending. Do not use image/video/music generation models or download large packages. The base renderer is your choice among installed tools.

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 3ef2665d13d52a3faacd9cdef32edf2c62e137de578fb8d0c97cbeb6b731fc43

SVG and animation

Actual SVG render from this case

Pelican riding a bicycle

Recognizable pelican and bicycle, with a long bill, throat pouch, connected fork and frame.

Source projectSVG

Test details

4m 13s · Requests / complete responses: 14 / 14 · 164,618 / 13,954 input / output tokens

  • Wheels, fork, triangular frame and pedals are recognizable; wing/grip and saddle placement are stylized.
  • Self-contained SVG parsing and actual rendering pass. This is a qualitative AI observation of this image, not an independent human score.

Itemized verification

Self-contained SVG and actual render · pass
Strict syntax/self-containment validation passes and actual pixels were inspected.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · model_finished · Requests / complete responses: 14 / 14 Run record

Model session finished normally; first SVG and revisions are preserved, with semantic checks reported separately.

Source brief and prompt
Pelican riding a bicycle. Source inputs remain archived locally and are identified by hashes.
Generate an SVG of a pelican riding a bicycle

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 2310a3a36c4b8999751f7562e91ef2991fb424faa7f9207aa414190c597436bf

Actual SVG render from this case

An analog clock at 9:41

The 09:41 clock passes independent SVG checks for numbers, hand angles and relative lengths.

Source projectSVG

Test details

1m 26s · Requests / complete responses: 7 / 7 · 40,818 / 5,392 input / output tokens

  • Numbers 1–12 appear once each. The hour hand rotates 290.5° with length 93; minute hand 246° with length 147, and no second hand.
  • The actual preview clearly shows 09:41, with numbers and the lower time label inside the canvas.

Itemized verification

Self-contained SVG and actual render · pass
Strict syntax/self-containment validation passes and actual pixels were inspected.
09:41 hands and numbers · pass
Numbers 1–12 appear once each. The hour hand rotates 290.5° with length 93; minute hand 246° with length 147, and no second hand.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · model_finished · Requests / complete responses: 7 / 7 Run record

Model session finished normally; first SVG and revisions are preserved, with semantic checks reported separately.

Source brief and prompt
An analog clock at 9:41. Source inputs remain archived locally and are identified by hashes.
Return only a complete, self-contained static SVG with viewBox="0 0 800 600". Draw a clear analog clock showing exactly 9:41, with all twelve numbers 1 through 12. The hour hand must be correctly positioned between 9 and 10, not exactly on 9. The minute hand must indicate minute 41. Use a shorter hour hand and a longer minute hand, no second hand. Add a small label "09:41" below the dial. Do not use scripts, external assets or embedded images.

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: e237e133ad6988ea13a95c43e5bdab8b286d2c4670a8e42e9e24d8875441dedc

Actual SVG render from this case

Data chart and proportional bars

Shared zero baseline, proportional bar heights, values, total and best quarter pass independent checks.

Source projectSVG

Test details

6m 45s · Requests / complete responses: 21 / 21 · 320,841 / 18,003 input / output tokens

  • Bar paths have heights 120, 240, 180 and 300 with a common y=460 baseline, a 2:4:3:5 ratio.
  • Actual quarter positions, value labels 12/24/18/30, Total:84 and Best:Q4 agree; the preview is clearly organized.

Itemized verification

Self-contained SVG and actual render · pass
Strict syntax/self-containment validation passes and actual pixels were inspected.
Proportions and labels · pass
Bar paths have heights 120, 240, 180 and 300 with a common y=460 baseline, a 2:4:3:5 ratio.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · model_finished · Requests / complete responses: 21 / 21 Run record

Model session finished normally; first SVG and revisions are preserved, with semantic checks reported separately.

Source brief and prompt
Data chart and proportional bars. Source inputs remain archived locally and are identified by hashes.
Return only a complete, self-contained static SVG with viewBox="0 0 800 600". Design a polished dashboard card titled "Quarterly Revenue". Draw a vertical bar chart for Q1=12, Q2=24, Q3=18, Q4=30, in that left-to-right order. All four bars must share a zero baseline, and bar heights must be exactly proportional to the values. Show a y-axis labeled 0, 10, 20, 30 and show the value above each bar. Add summary labels "Total: 84" and "Best: Q4". Use a light background and readable typography. Do not use scripts, external assets or embedded images.

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 882c407247829aa8ef726fef5bcfeb8f9a4df8bbe701974d065ce807e5b64a14

Fixed six-string Chinese poster test

Chinese poster and layout constraints

All six fixed Chinese strings appear once, with a readable hierarchy and recognizable bicycle icon.

Source projectSVG

Test details

4m 21s · Requests / complete responses: 13 / 13 · 188,283 / 15,447 input / output tokens

  • Exact checks pass for all six strings, numbers and punctuation, with no extra visible text.
  • Title, event details, description and button have a clear hierarchy, without observed text overlap or clipping in this render. The date belongs to a fixed test, not a live registration event.

Itemized verification

Self-contained SVG and actual render · pass
Strict syntax/self-containment validation passes and actual pixels were inspected.
Six fixed strings · pass
Exact checks pass for all six strings, numbers and punctuation, with no extra visible text.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · model_finished · Requests / complete responses: 13 / 13 Run record

Model session finished normally; first SVG and revisions are preserved, with semantic checks reported separately.

Source brief and prompt
Chinese poster and layout constraints. Source inputs remain archived locally and are identified by hashes.
只输出一个完整、自包含的静态 SVG,viewBox="0 0 800 600"。设计一张清晰、精致的中文活动海报。必须且仅包含以下六段文字,每段出现一次,不能增删字或改写:主标题“周末骑行计划”;副标题“慢一点,看见城市”;时间“10月24日 09:00”;地点“滨江公园东门”;说明“自备头盔 · 全程12公里”;按钮文字“立即报名”。画一个简洁的自行车图标。主标题最大,时间地点次之,说明最小。背景浅色,留足边距,所有文字保持在画布内,不重叠,不使用脚本、外部素材或嵌入图片。

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 6a1d4abaaf4e555ed8610782705bf065c14630467dd4dc8436079f6c56d44655

Animated cycling pelican

The four-second SVG renders and moves, but legs do not remain attached to the pedals.

Source projectSVG

More views and test details
Actual SVG render from this case

Actual SVG render from this case

1m 30s · Requests / complete responses: 8 / 8 · 48,291 / 5,658 input / output tokens

  • Independent Chromium capture produces 40 distinct frames. Mean pixel difference at the four-second boundary is 0.00085 with no obvious sampled loop jump; the SVG is unchanged.
  • Spokes and crank rotate while leg paths swing independently. At seven of eight sampled phases, at least one foot has an AABB gap over two pixels from both pedals, so attachment does not pass.
  • The subject stays inside the sampled canvas. The web MP4 is a host-captured single-cycle preview at 10fps; original SMIL SVG is downloadable.

Itemized verification

Self-contained SVG and actual render · pass
Strict syntax/self-containment validation passes and actual pixels were inspected.
Motion and sampled loop boundary · pass
Forty distinct frames; four-second boundary mean pixel error 0.00085; source SVG hash unchanged.
Foot-to-pedal attachment · fail
Seven of eight sampled phases show definite gaps under a conservative bounding-box test.

Task source

Independent offline generation preserving task requirements; Claude Code tool loop and Cycles CPU differ from the source benchmark.

attempt-1 · Claude Code requested adaptive · model_finished · Requests / complete responses: 8 / 8 Run record

Model session finished normally; first SVG and revisions are preserved, with semantic checks reported separately.

Source brief and prompt
Animated cycling pelican. Source inputs remain archived locally and are identified by hashes.
Return only a complete self-contained animated SVG with viewBox="0 0 800 600". Draw a recognizable pelican riding a bicycle: long bill and throat pouch, two wheels, connected frame, saddle and handlebar. Animate rotating wheel spokes and a rotating crank with pedaling legs that remain attached to the rider and pedals. Use SVG SMIL animateTransform/animate/animateMotion only (no CSS animation), with a seamless 4-second repeating loop. Keep the whole bicycle and bird on the canvas. Use a light background. No scripts, foreignObject, external assets, external fonts or embedded images. Use only local SVG shapes and text. The animation must visibly move, not just contain animation tags.

CPU protocol: Cycles CPU, 2 threads. Full private source inputs remain archived locally; source links and hashes identify them.

SHA-256: 54ff24c0b7a505c1b309f411ef8d8fd1741cb633219cbe6ea179fb7e6ae85dd9

Download all creative results and source index

Run settings, usage and full notes

LongCat-2.5-Preview · 2026-10-27

15 tasks: six Blender/3D scenes, four knowledge videos, four static SVGs and one SMIL animation. Local Claude Code uses isolated offline shell, pixel feedback and SVG rendering tools, with at most two concurrent cases. LongCat generates and revises all artwork code.

The configured LongCat API supplies generation. Totals sum final usage across all attempts, including cache fields and thinking, excluding separate configuration probes. Missing usage remains unknown. Claude Code USD estimates are not LongCat invoices. Fixed local Kokoro narration is infrastructure, not native model speech.

LongCat writes, previews and revises artwork; Codex orchestrates and independently checks outputs. A sandbox EEVEE fault led to Cycles CPU continuation, with earlier attempts preserved. This is not an identical-condition comparison with historical Kimi runs.

Visual commentary by AI; dimensions and media checks use scripts.

Requests / complete responses: 502 / 497 · Errors / interruptions: 2 / 2

27,939,033 / 774,219 input / output tokens, including cache and thinking

  • The actual API model is LongCat-2.5-Preview. These results do not establish a final LongCat 2.5 release. Claude Code requested adaptive thinking; the upstream mapping is unverified. No off/low/max comparison is invented.
  • Each attempt was capped at 60 minutes and 50 requests, with at most 32,768 output tokens per response and a 150,000 known-output budget with conservative reservations for missing usage. A reproducible sandbox EEVEE fault led to CPU continuations for the first two tasks; earlier attempts and cumulative time/usage remain included. Missing usage is not zero.
  • Frozen inputs include historical review guidance. The Jupiter prompt’s 0.500m value is not a current measurement. Old film-pass assertions for printer and phishing were removed before their first dispatch, preserving originals and replacement hashes. Tasks, renderer and tool loop differ from historical Kimi runs; no controlled or blind ranking is claimed.
  • LongCat generated and revised artwork; Codex handled infrastructure, isolation, archival, independent checks and publication. Fixed local Kokoro supplied narration. The circulation README discloses 1.63x audio acceleration, with no naturalness score. Host-captured SVG animation is not native video generation.
  • Checks cover reopening original files, selected dimensions/geometry, decoding, sampled frames and pedal attachment, rather than comprehensive engineering or scientific certification. Documented failures remain visible, including the printer beam crossing the drum and malformed raw SRT, derivative transition display and unfinished email film. Display WebP files and the printer VTT are separately derived; original artwork was not repaired.
Test details, original answers and sources