LongCat 2.5 review: Preview tests, frontend cases and version checks

How to read the brief: positioning is a testable hypothesis

The strongest signal is the move from answering questions to producing usable interfaces and completing multi-step work. A responsive frontend task therefore matters more than a screenshot alone: the page should preserve content on a phone, show the right data, and make its controls work.

Editorial judgment: model parameters describe scale, not the quality of a delivered website. Agent benchmarks also measure the model together with its tools and execution scaffold. Without matching those conditions, naming Claude, GPT, Kimi or DeepSeek does not make a controlled comparison.

Verified public model, specifications and API prices

These are documented limits, not results from a million-token test. The API model name can refer to a changing service; a response returning the same alias does not prove an immutable checkpoint.

Live API model-list response saved for this run

The official Responses reference separately documents input_image and reasoning.effort levels from none to xhigh, including low; the official Codex setup uses that interface. Two separate low requests below verify a usable response path, including one image input. They do not establish effort-tier semantics; a thinking token budget is not a substitute for those tiers. Responses API reference · Official Codex setup

Parameter reference · Quick Start · Chat API · Pricing · Changelog

See the official price and parameter pages

Captured on October 26. These are documentation screenshots, not benchmark charts.

2.5 Preview official limited-time USD prices

Official parameter table: 1M context and 256K output cap

Official sampling parameters: fixed temperatures follow thinking state

Test method: compare supported settings and retain every attempt

This independent batch started on October 26 and continues across midnight into October 27, Beijing time; earlier review data stays separate. Four preselected text cases and two coding cases run in thinking-on/off modes; this is a small diagnostic subset, not the complete 15-case text suite. Four fixed SVG prompts test a cycling pelican, a 9:41 clock, a Q1–Q4 dashboard and a Chinese poster; the fifth visual task requests a four-second cycling animation. A separate prompt asks for a responsive web interface.

New requests explicitly send thinking enabled or disabled; there is no separate omitted-switch group. They start at 12,288 output tokens with a 600-second timeout and requested temperature 0. The vendor documents fixed effective temperatures of 1.0/0.7 for on/off, so this compares documented modes rather than isolating thinking at one temperature. Failed generation/rendering attempts are retained; any budget increase is recorded per retry. Two independent queues may overlap, and local-client timings are not a speed ranking.

SVG source is preserved without repairs. The published site embeds raster renders rather than executing model-generated SVG. Animations are sampled in an isolated browser into 40 WebP frames over four seconds, with playback disabled by default. Geometric and executable checks are distinguished from qualitative observations made by the AI assistant.

TOKRACE method and limitations · Exact requests, returned IDs and budgets

Text and code: task checks, not a general intelligence score

A correct extraction or short function is evidence for that task. It does not establish performance on a large repository, tool selection or recovery from a multi-step failure. The earlier October 10 coding truncation remains in the historical review rather than being silently replaced by this batch.

A concrete case: parsing money without losing integer precision

This task checks 18 edge inputs rather than judging the explanation. Both modes can produce a correct small function while using different amounts of output and thinking. The following are individual observations from this concurrent local run, not average latency or a general cost ranking.

Inspect every original answer and executable check

Responsive frontend case: inspect desktop, phone and behavior

One-turn API generation of the same responsive cycling-club website in both thinking modes: 12,288 output tokens, 600 seconds. This is not a multi-turn Claude Code/Codex agent benchmark. No manual source edits. Original HTML is served as .html.txt; screenshots and interaction checks use isolated Chromium with no real session and every network request blocked. Signup is a local demonstration.

Both modes make a coherent cycling-club page with working route filters and local signup feedback. Each passes 24 desktop/mobile browser checks. Those functional checks do not catch a content inconsistency present in both: the advanced route is advertised as three hours, while the schedule runs from 09:15 to 11:30. A working interface still needs a business-content review.

Thinking off

First content / total：1.54s / 161.55s · 12182 token · done

Warm off-white, orange-red accents and serif headlines create a coherent design. Three desktop route columns become a single mobile column with a compact menu; illustrations are inline SVG. There is no horizontal overflow at 1440 or 390 pixels. Filters, valid signup submission and in-page feedback pass at both sizes: 24 browser checks pass, with no script errors. This is a fictional club and local-only signup demo; invented fee, leader and route-scouting copy does not describe a real event. It is a single API output, not evidence of multi-turn coding-agent performance. The advanced route says about three hours but the shared 09:15–11:30 riding schedule allows only 2h15m; functional checks do not establish content consistency.

Browser checks：24 / 24

• ✓ desktop: no horizontal overflow
• ✓ desktop: one main heading
• ✓ desktop: required route titles
• ✓ desktop: schedule exact times
• ✓ desktop: local demonstration disclaimer
• ✓ desktop: navigation anchor targets
• ✓ desktop: advanced filter shows only 山脊挑战
• ✓ desktop: leisure filter shows two routes
• ✓ desktop: all filter restores three routes
• ✓ desktop: labelled name and experience inputs
• ✓ desktop: signup success message after valid submit
• ✓ desktop: signup stays on the page
• ✓ mobile: no horizontal overflow
• ✓ mobile: one main heading
• ✓ mobile: required route titles
• ✓ mobile: schedule exact times
• ✓ mobile: local demonstration disclaimer
• ✓ mobile: navigation anchor targets
• ✓ mobile: advanced filter shows only 山脊挑战
• ✓ mobile: leisure filter shows two routes
• ✓ mobile: all filter restores three routes
• ✓ mobile: labelled name and experience inputs
• ✓ mobile: signup success message after valid submit
• ✓ mobile: signup stays on the page

Download original HTML · Original request and browser-check evidence

Exact frontend prompt

Build a polished responsive Chinese website for a weekend urban cycling club named「慢骑城市」. Return only one complete self-contained HTML document with all CSS and JavaScript inline, no Markdown. No external assets, fonts, libraries, URLs, network calls, storage or permissions. Use local SVG/CSS illustrations. This is a practical product page, not a wireframe. Requirements: (1) a top navigation with working anchor links to routes, schedule and signup; (2) a strong editorial hero; (3) a route grid containing exactly these three cards: 滨江轻骑—休闲—12公里, 老城漫游—休闲—8公里, 山脊挑战—进阶—35公里; (4) real filter buttons named 全部 / 休闲 / 进阶 that change visible route cards; (5) a schedule with 09:00集合、09:15出发、11:30结束; (6) a signup section with a labelled name input, an experience select, a button named 确认报名 and an on-page success message after a valid submit, no navigation; (7) readable, accessible layout at both 1440px desktop and 390px mobile, without horizontal overflow. Use excellent spacing, restrained typography, a warm light palette, one bold accent, clear focus styles and reduced-motion support. Do not fabricate testimonials, attendee counts or external sponsors. All key content and interactions must fit in this single file. The signup is explicitly a local demonstration; do not imply an actual booking.

Thinking on

First content / total：59.89s / 163.71s · 11094 token · done

Orange-red accents, warm paper tones and serif headings form a coherent design. Desktop has a two-column hero and three route cards; mobile keeps compact navigation and switches to one column. No horizontal overflow occurs at 1440 or 390 pixels. Filters, labelled form controls and signup feedback pass 24 browser checks across both sizes, with no script errors. Large empty areas make the lower desktop sections loose. The advanced route says about three hours while the shared riding schedule is 09:15–11:30, an inconsistency. This is one API output for a fictional local-only demo, not proof of a multi-turn coding agent or a live event service.

Browser checks：24 / 24

• ✓ desktop: no horizontal overflow
• ✓ desktop: one main heading
• ✓ desktop: required route titles
• ✓ desktop: schedule exact times
• ✓ desktop: local demonstration disclaimer
• ✓ desktop: navigation anchor targets
• ✓ desktop: advanced filter shows only 山脊挑战
• ✓ desktop: leisure filter shows two routes
• ✓ desktop: all filter restores three routes
• ✓ desktop: labelled name and experience inputs
• ✓ desktop: signup success message after valid submit
• ✓ desktop: signup stays on the page
• ✓ mobile: no horizontal overflow
• ✓ mobile: one main heading
• ✓ mobile: required route titles
• ✓ mobile: schedule exact times
• ✓ mobile: local demonstration disclaimer
• ✓ mobile: navigation anchor targets
• ✓ mobile: advanced filter shows only 山脊挑战
• ✓ mobile: leisure filter shows two routes
• ✓ mobile: all filter restores three routes
• ✓ mobile: labelled name and experience inputs
• ✓ mobile: signup success message after valid submit
• ✓ mobile: signup stays on the page

Download original HTML · Original request and browser-check evidence

Exact frontend prompt

Build a polished responsive Chinese website for a weekend urban cycling club named「慢骑城市」. Return only one complete self-contained HTML document with all CSS and JavaScript inline, no Markdown. No external assets, fonts, libraries, URLs, network calls, storage or permissions. Use local SVG/CSS illustrations. This is a practical product page, not a wireframe. Requirements: (1) a top navigation with working anchor links to routes, schedule and signup; (2) a strong editorial hero; (3) a route grid containing exactly these three cards: 滨江轻骑—休闲—12公里, 老城漫游—休闲—8公里, 山脊挑战—进阶—35公里; (4) real filter buttons named 全部 / 休闲 / 进阶 that change visible route cards; (5) a schedule with 09:00集合、09:15出发、11:30结束; (6) a signup section with a labelled name input, an experience select, a button named 确认报名 and an on-page success message after a valid submit, no navigation; (7) readable, accessible layout at both 1440px desktop and 390px mobile, without horizontal overflow. Use excellent spacing, restrained typography, a warm light palette, one bold accent, clear focus styles and reduced-motion support. Do not fabricate testimonials, attendee counts or external sponsors. All key content and interactions must fit in this single file. The signup is explicitly a local demonstration; do not imply an actual booking.

A separate Responses check: low requests and reading a chart

The Responses API accepted reasoning.effort=low for one text extraction and one input_image request. The chart answer correctly identifies LongCat-2.0 in the green-bar legend and returns all six green-bar values: seven checks pass. This is evidence that these two requests worked, separate from the Chat Completions on/off results.

The input is a public historical 2.0 chart, not a 2.5 benchmark result. There is no no-image control or repeated trial here, so memorization cannot be excluded. This does not establish native architecture or general visual reliability. The non-streaming calls do not measure first streaming content, and accepting low does not prove its effective reasoning strength.

Text input: exact JSON extraction

Responses · low · HTTP 200 · 3.62s · 116 output tokens including reasoning

{"name": "林桐", "age": 28, "city": "杭州", "email": null}

Exact prompt and checks

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。

Exact expected JSON matched.

Chart input: legend plus six values

Responses · low · HTTP 200 · 5.88s · 268 output tokens including reasoning

{"model":"LongCat-2.0","scores":{"Terminal-Bench 2.1":70.8,"SWE-bench Pro":59.5,"SWE-bench Multilingual":77.3,"FORTE":73.2,"RWSearch":78.8,"BrowseComp":79.9}}

Inspect the exact input image

Exact prompt and checks

只根据这张图片回答。识别底部图例中绿色柱对应的模型名称，并读取六个小图里该绿色柱顶部的数值。只输出JSON对象，恰好包含 model 和 scores。scores 是对象，键分别使用图中的六个benchmark英文名称，值为数字；不能读取其他颜色柱代替。看不清的值用null。不要推断别的模型版本、不要解释、不要Markdown。

• ✓ model identity from legend
• ✓ Terminal-Bench 2.1: 70.8
• ✓ SWE-bench Pro: 59.5
• ✓ SWE-bench Multilingual: 77.3
• ✓ FORTE: 73.2
• ✓ RWSearch: 78.8
• ✓ BrowseComp: 79.9

Download both Responses requests, input provenance and checks · Official Responses reference

The visual cases reveal errors that a clean render can hide

The thinking-on clock gets the requested 9:41 hand rotations exactly right in source: 290.5° for the hour hand and 246° for the minute hand. Its thinking-off counterpart misplaces 3 and 9 outside the dial and clips part of its tick geometry. For this particular task, thinking produces a more accurate layout; the original timing and token cost remain visible below.

The dashboard makes a different distinction. The thinking-off bars keep the right proportions but do not match their vertical-axis scale. Thinking-on fixes the geometry and totals, then adds an unrequested 2024 date and a USD-thousands unit. A numerically sound chart can still invent business context, so both geometry and the meaning of its labels need checking.

Thinking also has a budget cost here: the first thinking-on pelican and Chinese-poster answers both hit the 12,288-token limit before completing an SVG. Higher-budget retries are recorded separately below. In the thinking-off animation, the wheels and crank rotate but the moving legs separate from the hip and leave the frame. Motion alone does not produce a coherent cycling action.

The thinking-on animation keeps the hip fixed while the feet follow the crank, making the joints more coherent and keeping the bird and bicycle inside the frame. Its first attempt also truncated; a retry with a 32,768-token ceiling took 477.49 seconds and returned 24,972 output tokens. The legs still bend mechanically. This improvement is a higher-budget retry result, not a same-budget first-attempt win.

Updated visual evaluation · 2026-10-26

Requested thinking settings, retries and a cycling pelican animation

Independent 2026-10-26 retest. Requested and returned model: LongCat-2.5-Preview; this does not establish release or checkpoint equivalence of the brief’s LongCat-2.5 name. Both groups explicitly request thinking enabled/disabled. temperature=0 is sent, but the vendor documents overriding it to 1.0 on / 0.7 off, so this is not a deterministic temperature-zero comparison. The Chat Completion documentation does not list named effort levels. A separate Responses API documents low and other levels, without a max name; those protocols are not mixed here. First attempts use 12,288 output tokens and a 600-second deadline. Up to two independent request queues overlap; local client latency is recorded without a speed ranking. All first attempts/failures are retained. Private reasoning content is omitted. Poster thinking-on attempt 2 encountered local ENOSPC while saving its result. Its dispatched slot remains interrupted; response, final usage, upstream outcome and measured duration are unknown. This is not a provider/model answer failure. Attempt 3 is independent. Token totals are lower bounds from reported usage; full consumption and cost are unknown. This section shows the first successful static/animated render in each mode, retaining failed first attempts and independent retries. The separate static archive does not add duplicate requests. Official parameter documentation

14 visual request records in this evidence set. Failed attempts are retained below each result. We show the first successfully rendered result in each sequence, not a best-of selection. Render success does not imply correct geometry or instruction following. Original first attempts remain below. Dates and budgets can differ; these are examples, not a statistical ranking.

Each first attempt uses 12,288 output tokens and a 600-second deadline. Generation/render failures receive independent numbered retries with failures preserved. Animations are model-generated SVG SMIL, captured without source edits at 10 fps for four seconds. Click to play; click again to pause.

Failure records: 0 server errors, 0 connection failures, 0 timeouts, 1 interrupted, 3 truncated, 0 invalid SVG/animation. Wall-clock timings include interruptions; some exceed the configured timeout and should not be interpreted as model inference latency.

Download all supplemental requests and failures

Pelican riding a bicycle

2026-10-26 UTC · Attempt 1 · 30.80s · 2289 token

First content: 1.31s · Token limit: 12288

The long bill, pouch, pale body, two wheels and triangular frame are recognizable, and the feet roughly meet the pedals. The front fork is missing: the frame ends at (380,300) without connecting to the front hub at (420,380), leaving the front wheel detached. The wing also stops short of the handlebar. Renderable, but mechanically incomplete. This is an AI assistant qualitative observation, not independent human scoring.

Recorded requests in this sequence: 1 · Recorded request time: 30.80s

Original SVG · Request evidence

Exact prompt

Generate an SVG of a pelican riding a bicycle

2026-10-26 UTC · Attempt 2 · 304.22s · 20529 token

First content: 245.10s · Token limit: 32768

After the 12,288-token first attempt truncated, independent attempt 2 raised the ceiling to 32,768 and produced a complete SVG in 304.22 seconds with 20,529 output tokens; cumulative request time is 515.40 seconds. The bill, pouch, two wheels, triangular frame and front fork are recognizable. Feet meet the pedal area and the wing reaches the handlebar, with a complete colored scene. Legs remain cartoonishly long and straight. This is the first renderable retry, not a same-budget first-attempt winner or a best-of selection.

Recorded requests in this sequence: 2 · Recorded request time: 515.40s

Original SVG · Request evidence

Earlier attempts and failures (1)

1. #1 · 2026-10-26T15:48:26.980Z · truncated · 211.18s · No complete SVG in the original response

Exact prompt

Generate an SVG of a pelican riding a bicycle

An analog clock at 9:41

2026-10-26 UTC · Attempt 1 · 42.93s · 3335 token

First content: 0.99s · Token limit: 12288

All twelve numerals are present, but the dial is malformed: 3 and 9 sit outside the circular face, and minute ticks trace a displaced arc with many coordinates below the 600-unit canvas (for example y=775.2), causing clipping. Endpoint angles are about 289.85° for the hour hand and 245.89° for the minute hand, deviating from 290.5°/246°. The hour hand is correctly shorter. The source has a transparent background; a white page does not fix the geometry.

Recorded requests in this sequence: 1 · Recorded request time: 42.93s

Original SVG · Request evidence

Exact prompt

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.

2026-10-26 UTC · Attempt 1 · 140.04s · 10253 token

First content: 97.79s · Token limit: 12288

All twelve numerals are ordered and readable, with generous margins and a clear 09:41 label. Source transforms set the minute hand to 246° and hour hand to 290.5°; their forward lengths are 118 and 78, matching the time and relative lengths. The minute tip is close to numeral 8 without covering it. Thinking-on took 140.04 seconds and 10,253 output tokens. Visual observations are from an AI assistant, with separate coordinate checks.

Recorded requests in this sequence: 1 · Recorded request time: 140.04s

Original SVG · Request evidence

Exact prompt

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.

Data chart and proportional bars

2026-10-26 UTC · Attempt 1 · 16.13s · 1275 token

First content: 1.08s · Token limit: 12288

All bars share the y=420 zero baseline. Quarter labels, values, Total 84 and Best Q4 are correct, and heights 40/80/60/100 preserve the 2:4:3:5 ratio. The y-axis is inconsistent: every 10 units spans 80 pixels, but the bar labelled 24 is only 80 pixels tall and reaches tick 10. The bar labelled 30 reaches about 12.5. The layout is clean, but the scale misrepresents the data.

Recorded requests in this sequence: 1 · Recorded request time: 16.13s

Original SVG · Request evidence

Exact prompt

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.

2026-10-26 UTC · Attempt 1 · 53.87s · 3924 token

First content: 34.96s · Token limit: 12288

The bars share y=460 and have heights 120/240/180/300, matching values 12/24/18/30 and the y-axis scale of 100 pixels per 10 units. Quarter order, total 84 and best Q4 are correct; summaries and values are readable without overlap. It adds an unprovided year, 2024, and a USD-thousands unit, which would need verification in a real business chart.

Recorded requests in this sequence: 1 · Recorded request time: 53.87s

Original SVG · Request evidence

Exact prompt

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.

Chinese poster and layout constraints

2026-10-26 UTC · Attempt 1 · 18.14s · 1053 token

First content: 2.17s · Token limit: 12288

All six exact strings appear once with intact Chinese characters, punctuation and numbers. Font sizes establish the required hierarchy: heading 56, time/place 28, note 18. Pale green, centered details and a clear button are readable. The bicycle is recognizable, but the wheel sits very close to the heading, reducing breathing room. October 24 belongs to the fixed test prompt, not a current event announcement.

Recorded requests in this sequence: 1 · Recorded request time: 18.14s

Original SVG · Request evidence

Exact prompt

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。

2026-10-26 UTC · Attempt 3 · 160.53s · 8921 token

First content: 136.19s · Token limit: 32768

The complete preserved result is attempt 3: a 32,768-token ceiling, 160.53 seconds and 8,921 output tokens. Attempt 1 truncated; attempt 2 lost its result to local ENOSPC, so its quality and usage remain unknown. All six exact strings appear once. The 72px heading, 34px time/place and 22px note establish clear hierarchy; text and button stay inside the canvas without overlapping. Bicycle and information icons are recognizable. Pale-green decoration, border and button shadow are polished, though the button sits close to the bottom border. October 24 is the fixed test date.

Recorded requests in this sequence: 3 · Recorded request time: ≥ 398.43s

Original SVG · Request evidence

Earlier attempts and failures (2)

1. #1 · 2026-10-26T15:56:11.148Z · truncated · 237.90s · No complete SVG in the original response
2. #2 · 2026-10-26T16:06:14.740Z · interrupted · — · Local ENOSPC while saving the request result. Original response, final usage, upstream status and measured duration were not preserved; not a model/provider answer failure.

Exact prompt

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。

Animated cycling pelican

Play 4-second animation

2026-10-26 UTC · Attempt 1 · 52.21s · 2993 token

First content: 1.20s · Token limit: 12288

The long bill, pouch and bicycle are recognizable. Spokes and crank visibly move: all 40 sampled frames differ, and real-time WebP playback across more than one loop confirms motion. Leg attachment fails: animateMotion translates entire legs, detaching hips from the body, while foot paths remain near the origin as floating shapes at the upper left. Part of a leg leaves the lower canvas. The four-second boundary differs by only 0.0245 mean pixel levels out of 255, but is not pixel-identical. A moving encoded animation, yet it fails coherent joints and no-clipping requirements.

Recorded requests in this sequence: 1 · Recorded request time: 52.21s

Original SVG · Animated WebP · Request evidence

Exact prompt

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.

Play 4-second animation

2026-10-26 UTC · Attempt 2 · 477.49s · 24972 token

First content: 430.91s · Token limit: 32768

After the 12,288-token first attempt truncated, attempt 2 raised the ceiling to 32,768 and produced the animation in 477.49 seconds with 24,972 output tokens; cumulative recorded request time is 714.89 seconds. All 40 frames vary, and nine real-time WebP samples over 4.74 seconds confirm moving spokes and crank. The source fixes the hip at (352,318) while feet follow the crank path, giving more coherent attachment than the thinking-off sample. Bird and bicycle remain inside the canvas. Legs are still mechanically bent thick polylines. The four-second boundary looks continuous but is not pixel-identical: mean difference 0.0048/255. This is not a same-budget first-attempt comparison or a physics simulation.

Recorded requests in this sequence: 2 · Recorded request time: 714.89s

Original SVG · Animated WebP · Request evidence

Earlier attempts and failures (1)

1. #1 · 2026-10-26T16:01:19.454Z · truncated · 237.40s · No complete SVG in the original response

Exact prompt

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.

Vendor-reported · separate from TOKRACE tests

LongCat official benchmark figures

Version-matched benchmark provenance is handled in the page’s official-source section; figures from other LongCat versions are not substituted.

What else is worth testing?

The following are proposed extensions, not results from this run. They test different failure modes rather than producing another overall score.

Official historical material · not new 2.5 scores

Official historical figures: benchmarks and architecture

No matching LongCat 2.5 benchmark chart was located in the public sources checked for this review. Historical official figures below provide context for the product direction; their captions identify the original version. They are vendor-reported results, have not been reproduced by TOKRACE, and must not be attributed to the current Preview API.

For agent and coding charts, inspect the scaffold, tools, evaluation split, sampling policy and cost budget before comparing bars. A score from a different model version or harness does not establish parity in a live frontend task.

Historical context: original LongCat-2.0 vendor benchmark chart

Historical context: LongCat-2.0 expert distillation diagram

Source archive, original image bytes, dimensions and SHA-256

Original prompts, answers and code checks

Download this text/code batch · Frontend evidence · Responses diagnostics · Full request accounting · First visual attempts · Animation and retry records

Only answer content and reasoning timing/token counts are published; hidden reasoning text is excluded. Successful retries do not delete failed attempts. Exact request and returned model IDs are retained in the downloadable evidence.

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 (!/^\d+(\.\d{1,2})?$/.test(trimmed)) return null;
    const [intPart, decPart = ''] = trimmed.split('.');
    const cents = Number(intPart) * 100 + Number((decPart + '00').slice(0, 2));
    if (!Number.isSafeInteger(cents) || cents > Number.MAX_SAFE_INTEGER) return null;
    return cents;
}
```

const cases=[['0',0],[' 12.30 ',1230],['0.29',29],['1.2',120],['0001.02',102],['1.234',null],['-1',null],['+1',null],['1e3',null],['',null],['.5',null],['1.',null],['Infinity',null],['1,000',null],[null,null],[1,null],['90071992547409.91',9007199254740991],['90071992547409.92',null]]; for (const [input,want] of cases) if(parseCents(input)!==want) throw Error('Failed '+JSON.stringify(input));

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;
  var m = value.trim().match(/^(\d+)(?:\.(\d{1,2}))?$/);
  if (!m) return null;
  var frac = m[2] ? (m[2] + '0').slice(0, 2) : '00';
  var centsStr = (m[1] + frac).replace(/^0+(?=\d)/, '');
  var max = String(Number.MAX_SAFE_INTEGER);
  if (centsStr.length > max.length || (centsStr.length === max.length && centsStr > max)) return null;
  return Number(centsStr);
}

const cases=[['0',0],[' 12.30 ',1230],['0.29',29],['1.2',120],['0001.02',102],['1.234',null],['-1',null],['+1',null],['1e3',null],['',null],['.5',null],['1.',null],['Infinity',null],['1,000',null],[null,null],[1,null],['90071992547409.91',9007199254740991],['90071992547409.92',null]]; for (const [input,want] of cases) if(parseCents(input)!==want) throw Error('Failed '+JSON.stringify(input));

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;
}

const a={id:'__proto__',value:1},b={id:'constructor'},c={id:''},d={id:'x'};const input=Object.freeze([a,b,{id:'__proto__'},c,{id:''},d,{id:'constructor'}]);const got=uniqueById(input);if(got.length!==4||got[0]!==a||got[1]!==b||got[2]!==c||got[3]!==d)throw Error('Wrong identity or order');if(uniqueById([]).length!==0)throw Error('Empty');if(input.length!==7||a.value!==1)throw Error('Mutation');

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。

```javascript
function uniqueById(items) {
  const seen = new Set();
  const result = [];
  
  for (const item of items) {
    const id = item.id;
    if (!seen.has(id)) {
      seen.add(id);
      result.push(item);
    }
  }
  
  return result;
}
```

const a={id:'__proto__',value:1},b={id:'constructor'},c={id:''},d={id:'x'};const input=Object.freeze([a,b,{id:'__proto__'},c,{id:''},d,{id:'constructor'}]);const got=uniqueById(input);if(got.length!==4||got[0]!==a||got[1]!==b||got[2]!==c||got[3]!==d)throw Error('Wrong identity or order');if(uniqueById([]).length!==0)throw Error('Empty');if(input.length!==7||a.value!==1)throw Error('Mutation');

Frequently asked questions

Is LongCat 2.5 a confirmed non-Preview release?

The public documentation checked for this review still names LongCat-2.5-Preview. A brief using LongCat-2.5 does not establish a distinct released checkpoint. Every new request here records the Preview API model ID.

Does this review verify native multimodal or long-running agent performance?

A separate Responses request supplies one public chart through input_image and correctly extracts its model legend and six values, passing seven checks. With no no-image control or repeat trials, this does not rule out memorization or establish a native architecture, general visual ability, million-token reliability, autonomous repository work or 3D performance.

Can LongCat be compared at low and max thinking levels?

The main batch uses Chat Completions, whose checked documentation describes a thinking switch and thinking_budget rather than named effort tiers. Responses documents none/minimal/low/medium/high/xhigh, with no max label. Two independent low requests succeeded, but parameter acceptance does not prove actual reasoning strength or establish a cross-interface comparison.

Are the official benchmark figures LongCat 2.5 scores?

No matching official LongCat 2.5 benchmark chart was located in the checked sources. Any historical official figures below retain their original model version and must not be read as new 2.5 scores or TOKRACE reproductions.

What does the API cost?

The checked promotional price per million tokens is US$0.30 input, US$0.006 cached input and US$1.20 output; the China price page lists RMB2, RMB0.04 and RMB8 respectively. These are temporary vendor prices, separate from site trial quotas; the provider invoice determines the charge.