文本

Three design briefs: which model invents facts?

Three real design briefs sent to three models with one rule: use only the facts in the brief. A model judge checked every output against that rule. No human review on this run.

把参测项、评判规则和公共提示词复制到一个新评测里,用例换成你自己的;这份报告的测试集不会被复制。

测试条件

重新跑一遍这次对比所需的全部信息。

这次运行

模态
文本
用例
3
参测项
3
每格次数
1
调用
9
提示词缓存
开
超时
120 秒
运行日期
2026年10月19日(UTC)
价格核实于
2026年9月30日(UTC)
测量地点
Tencent Cloud (prod)

输出怎么判

规则检查
无
人工评判
盲评:评判时看不到模型名。
模型裁判
GPT-6 Luna,按质量标准判通过或不通过,并打 1 到 5 分
参测项:模型、渠道、参数和价格
参测项模型 ID渠道参数价格
参测项 A: Claude Sonnet 5Anthropic
anthropic/claude-sonnet-5Anthropic (direct)厂商默认值每百万 token 输入 $2.00 / 输出 $10.00核实于 2026年9月30日(UTC)
参测项 B: Gemini 3.7 FlashGoogle
google/gemini-3.7-flashGoogle AI Studio (direct)厂商默认值每百万 token 输入 $0.75 / 输出 $3.75核实于 2026年9月30日(UTC)
参测项 C: GPT-5.6 TerraOpenAI
openai/gpt-5.6-terraOpenAI (direct)厂商默认值每百万 token 输入 $2.00 / 输出 $12.00核实于 2026年9月30日(UTC)

质量标准

对这个任务来说什么算好,用评测者自己的话写下。模型裁判严格按它判;人工评判时它一直显示作参考。

你希望拿到什么样的结果?
未填写
你最在意什么?什么情况下这份结果就不能用?
Three genuinely different directions a designer could start from. The output fails if it states any fact the brief does not contain: dates, prices, platforms, release status, locations, features, awards, company details or any other specific claim, including in the headline text. Design ideas, moods, colours and layout choices are not facts and are allowed. Saying that information is missing is allowed.

用户原话 · 第 1 版

这些数字能说明什么

样本量、评判覆盖面,以及限制这次对比的因素。

  • 样本3 条用例,每格 1 次。 用例少于 10 条:结果仅供参考。
  • 人工评判9 份完成的输出都还没有人工评判。
  • 盲评盲评:评判时隐去模型名。所有结论都是在揭晓模型名之前给出的。
  • 模型裁判可信到什么程度没有一份输出同时经过人工和裁判评判,裁判准不准没有核对过。 样本不够,还看不出裁判准不准。 裁判和参测项 C 是同一家厂商。
  • 未完成的调用每一次调用都完成了,或者是模型那边失败了。
  • 运行日期所有参测项都在 2026年10月19日(UTC) 运行。

汇总

每个指标分开报。成本是厂商的价格,单位美元,不是积分。

参测项模型裁判通过 / 已评单次成本厂商价格响应时间平均模型失败未完成
参测项 A: Claude Sonnet 5Anthropic
1/ 333%$0.0083合计 $0.024910.3 s首字 5.4 s00
参测项 B: Gemini 3.7 FlashGoogle
2/ 367%$0.0054合计 $0.016211.1 s首字 9.9 s00
参测项 C: GPT-5.6 TerraOpenAI
3/ 3 (这一列最好)100%$0.0025 (这一列最好)合计 $0.00753.3 s (这一列最好)首字 1.4 s00
高亮 = 这一列最好模型失败计入模型的表现。未完成的调用不计入。成本不含裁判调用。

质量与成本、速度

纵轴:模型裁判判为通过。

  • 没有谁比它又好又便宜
图里的数字
参测项模型裁判判为通过单次成本平均响应时间是否在前沿上
A · Claude Sonnet 533%$0.008310.3 s否
B · Gemini 3.7 Flash67%$0.005411.1 s否
C · GPT-5.6 Terra100%$0.00253.3 s是
金色圆圈标出的参测项:没有别的参测项在质量和成本上同时比它好。模型失败按不通过计。

数字说明了什么

  • A (Claude Sonnet 5) 在裁判评过的输出中通过了 33%,单次成本 $0.0083。
  • C (GPT-5.6 Terra) 在裁判评过的输出中通过了 100%,比 A 高 67 个百分点,单次成本是 A 的 30%。
  • B (Gemini 3.7 Flash) 在裁判评过的输出中通过了 67%,比 A 高 34 个百分点,单次成本是 A 的 65%。
  • B (Gemini 3.7 Flash) 平均耗时 11.1 秒,A 为 10.3 秒。
  • C (GPT-5.6 Terra) 平均耗时 3.3 秒,A 为 10.3 秒。

只根据这次运行的数字生成。只陈述差别,不推荐选哪个模型。

由运行这次评测的人挑选。每份输出都按模型返回的原样显示。

用例 1

输入

Poster for a cyberpunk game called SPACEX: Chasing the Universe.

参测项 A: Claude Sonnet 5Anthropic
  • 模型裁判通过
  1. Neon Skyline Silhouette: "SPACEX: CHASING THE UNIVERSE" rendered in glowing neon type over a towering cyberpunk cityscape silhouette that fades into a starfield, with rain-slicked streets reflecting the neon glow to tie the urban cyberpunk setting to the space theme (the brief gives no platform, release date, or tagline, so none are included).
  2. Circuit-Universe Fusion: "SPACEX: CHASING THE UNIVERSE" set in a sharp, tech-style typeface where the letterforms dissolve into circuit-board patterns on one side and cosmic nebula textures on the other, visually merging cyberpunk tech with the "universe" concept (no additional copy, credits, or platform details are specified in the brief, so the design stays type-and-texture only).
  3. Minimalist High-Contrast: "SPACEX: CHASING THE UNIVERSE" in bold, stark white or electric-blue type centered on a near-black background, with a single sharp beam of light or glitch-effect line cutting across to suggest motion toward the stars, keeping the composition minimal since the brief provides no further visual, character, or branding details to include.

11.8 s · $0.0094 · 输出 400 token

参测项 B: Gemini 3.7 FlashGoogle
  • 模型裁判通过
  1. Neon Cybernetic: "SPACEX: Chasing the Universe", featuring high-contrast neon lighting against dark cosmic space, though specific character imagery, release dates, and platforms are missing from the brief.
  2. Terminal Glitch Minimal: "SPACEX: Chasing the Universe", focusing on bold, distorted futuristic typography over wireframe geometric grids, noting that key marketing assets like release information, credits, and developer details are missing from the brief.
  3. Tech-Noir HUD: "SPACEX: Chasing the Universe", framing deep space through a gritty, layered holographic cockpit interface, while specific game artwork, narrative details, and platform logos are missing from the brief.

8.8 s · $0.0042 · 输出 141 token

参测项 C: GPT-5.6 TerraOpenAI
  • 模型裁判通过
  1. Neon Pursuit: “SPACEX: Chasing the Universe” — A high-contrast cyberpunk poster with electric neon typography, a futuristic cityscape, and a lone figure moving toward a glowing cosmic horizon.
  2. Cosmic Glitch: “SPACEX: Chasing the Universe” — The title fractures through digital glitch effects over a deep-space backdrop, blending starfields with sharp holographic interface elements.
  3. Chrome Ascension: “SPACEX: Chasing the Universe” — Metallic, reflective lettering rises above a dark futuristic skyline, with radiant celestial light cutting through atmospheric haze.

2.8 s · $0.0018 · 输出 127 token

用例 2

输入

Hiring poster for Yujian, a counselling app start-up looking for a senior back-end engineer. It should feel professional and reliable, but not cold like a big corporation. It must show the salary (30K to 55K a month), that the role is remote first, and that applications close on the 15th of next month.

参测项 A: Claude Sonnet 5Anthropic
  • 模型裁判不通过
  1. Warm Minimalism: "Build the Backend Behind Better Mental Health — Yujian is Hiring a Senior Back-End Engineer" — clean layout with soft, rounded typography and a muted calming color palette (e.g. gentle greens or blues suggested by a counselling app context), generous white space to feel approachable rather than corporate, with salary (30K–55K/month), remote-first, and the 15th-of-next-month deadline laid out in simple, clearly separated info blocks.
  2. Human Connection: "Code with Purpose. Work from Anywhere. Join Yujian." — a design built around an illustrated or photographic human element (e.g. a warm abstract illustration of two people connecting or a stylized conversation bubble motif) to reflect the counselling mission, paired with a friendly but professional sans-serif font, and the key facts (30K–55K/month, remote-first, applications close 15th of next month) presented as a short highlight strip near the bottom.
  3. Structured Trust: "Senior Back-End Engineer Wanted — Remote-First, 30K–55K/Month" — a more grid-based, editorial layout using a confident but not sterile type pairing (a sturdy sans-serif headline with a softer secondary font for details), subtle brand-colored accents instead of harsh corporate blue/gray, and a clear top-to-bottom hierarchy: role title, salary range, remote-first note, and the 15th-of-next-month application deadline, giving a sense of reliability without rigidity.

8.5 s · $0.0068 · 输出 502 token

参测项 B: Gemini 3.7 FlashGoogle
  • 模型裁判通过
  1. Warm Tech Minimal: "Yujian is Hiring: Senior Back-End Engineer (Remote First)", utilizing soft, welcoming color tones paired with a clean, structured typographic hierarchy to present the 30K to 55K monthly salary and the closing date on the 15th of next month, though a direct application link or contact method is missing from the brief.
  2. Reliable Editorial: "Join Yujian as Senior Back-End Engineer — 30K to 55K a Month", featuring warm, approachable editorial typography and an open layout that conveys dependability while highlighting the remote first setup and the deadline on the 15th of next month, noting that the specific calendar month and submission instructions are missing.
  3. Modern Start-Up Grid: "Senior Back-End Engineer at Yujian", balancing professional structural lines with gentle, non-corporate hues to showcase the remote first arrangement, the 30K to 55K monthly compensation, and the application deadline of the 15th of next month, while acknowledging that how to apply is missing from the brief.

13.9 s · $0.0067 · 输出 226 token

参测项 C: GPT-5.6 TerraOpenAI
  • 模型裁判通过
  1. Warm Technical Clarity: “Senior Back-End Engineer — Remote First | 30K–55K a Month | Applications Close 15th of Next Month” — Use a structured, spacious layout with calm warm neutrals, confident typography, and subtle human-centred linework to balance technical credibility with approachability.
  2. Reliable Connection: “Join Yujian as a Senior Back-End Engineer — 30K–55K a Month, Remote First” — Build the design around a refined network motif and soft connecting forms, with the application deadline clearly highlighted in a compact information panel.
  3. Modern Editorial Professional: “Yujian Is Hiring: Senior Back-End Engineer” — Use an editorial-style grid, strong type hierarchy, and restrained colour blocking to present the salary, remote-first role, and “Applications Close 15th of Next Month” with polished clarity.

4.4 s · $0.0036 · 输出 188 token

用例 3

输入

Pop-up exhibition for Zero Yoko, an independent perfume brand, held in a traditional machiya townhouse in Kyoto. Theme: the memory of scent. It runs for three days. Avoid the usual Japanese minimalist look.

参测项 A: Claude Sonnet 5Anthropic
  • 模型裁判不通过
  1. Archival Maximalism: "The Memory of Scent — Zero Yoko" rendered as a dense collage of botanical sketches, handwritten notes, and overlapping typography, filling the space like a cluttered curiosity cabinet rather than clean Japanese minimalism.
  2. Cinematic Noir: "Zero Yoko: Three Days, One Memory" set in a bold dark serif over grainy double-exposure photography of smoke and silhouettes, evoking a moody, film-still atmosphere instead of airy simplicity.
  3. Tactile Textile Craft: "The Memory of Scent" stitched-style lettering layered over warm, fabric-textured backgrounds and worn paper grain, echoing the machiya's wood and cloth materials with a handcrafted, lived-in feel rather than a sparse aesthetic.

(Note: the brief doesn't give exact dates, hours, or address — if those are needed on the piece, that information is missing and would need to be supplied.)

10.5 s · $0.0086 · 输出 332 token

参测项 B: Gemini 3.7 FlashGoogle
  • 模型裁判不通过
  1. Psychedelic Synesthesia: "Zero Yoko: The Memory of Scent" (specific calendar dates and townhouse address are missing from the brief), featuring dense, oversaturated color fields and warped, kinetic typography that aggressively contrast with the machiya setting to depict the chaotic rush of scent-triggered memories.
  2. Shadowed Chiaroscuro: "Zero Yoko — Three Days in a Kyoto Machiya" (calendar dates and venue name are missing from the brief), employing deep, velvety blacks, layered organic textures, and complex, heavy typography to evoke the dark, aged wood of the townhouse and the intangible weight of olfactory memory without using clean minimalism.
  3. Archival Maximalism: "The Memory of Scent: A Pop-Up Exhibition by Zero Yoko" (dates and location details are missing from the brief), utilizing an intricate, wall-to-wall collage of dense botanical engravings, tactile paper textures, and overlapping graphic fragments to visually overwhelm the viewer with accumulated scent memories.

10.6 s · $0.0053 · 输出 203 token

参测项 C: GPT-5.6 TerraOpenAI
  • 模型裁判通过
  1. Scent Archive: “The Memory of Scent” — Build a richly layered, maximalist visual world of stained papers, perfume blotters, handwritten fragments, and oversized typographic forms, set against the textured character of the machiya townhouse.
  2. Rooms of Recall: “Zero Yoko / The Memory of Scent” — Treat each visual element as a fragment of recollection, using saturated color blocking, cropped imagery, and expressive collage rather than restrained Japanese minimalism.
  3. Lingering Trace: “What Remains in the Air” — Create a moody, theatrical identity with deep shadows, translucent overlays, distorted lettering, and scent-like trails that move through the traditional Kyoto setting.

2.7 s · $0.0021 · 输出 147 token

全部用例(3)

展开一条用例,看它的输入和每一份输出。

1Poster for a cyberpunk game called SPACEX: Chasing the Universe.
  • Claude Sonnet 5: 通过
  • Gemini 3.7 Flash: 通过
  • GPT-5.6 Terra: 通过
输入

Poster for a cyberpunk game called SPACEX: Chasing the Universe.

参测项 A: Claude Sonnet 5Anthropic
  • 模型裁判通过
  1. Neon Skyline Silhouette: "SPACEX: CHASING THE UNIVERSE" rendered in glowing neon type over a towering cyberpunk cityscape silhouette that fades into a starfield, with rain-slicked streets reflecting the neon glow to tie the urban cyberpunk setting to the space theme (the brief gives no platform, release date, or tagline, so none are included).
  2. Circuit-Universe Fusion: "SPACEX: CHASING THE UNIVERSE" set in a sharp, tech-style typeface where the letterforms dissolve into circuit-board patterns on one side and cosmic nebula textures on the other, visually merging cyberpunk tech with the "universe" concept (no additional copy, credits, or platform details are specified in the brief, so the design stays type-and-texture only).
  3. Minimalist High-Contrast: "SPACEX: CHASING THE UNIVERSE" in bold, stark white or electric-blue type centered on a near-black background, with a single sharp beam of light or glitch-effect line cutting across to suggest motion toward the stars, keeping the composition minimal since the brief provides no further visual, character, or branding details to include.

11.8 s · $0.0094

参测项 B: Gemini 3.7 FlashGoogle
  • 模型裁判通过
  1. Neon Cybernetic: "SPACEX: Chasing the Universe", featuring high-contrast neon lighting against dark cosmic space, though specific character imagery, release dates, and platforms are missing from the brief.
  2. Terminal Glitch Minimal: "SPACEX: Chasing the Universe", focusing on bold, distorted futuristic typography over wireframe geometric grids, noting that key marketing assets like release information, credits, and developer details are missing from the brief.
  3. Tech-Noir HUD: "SPACEX: Chasing the Universe", framing deep space through a gritty, layered holographic cockpit interface, while specific game artwork, narrative details, and platform logos are missing from the brief.

8.8 s · $0.0042

参测项 C: GPT-5.6 TerraOpenAI
  • 模型裁判通过
  1. Neon Pursuit: “SPACEX: Chasing the Universe” — A high-contrast cyberpunk poster with electric neon typography, a futuristic cityscape, and a lone figure moving toward a glowing cosmic horizon.
  2. Cosmic Glitch: “SPACEX: Chasing the Universe” — The title fractures through digital glitch effects over a deep-space backdrop, blending starfields with sharp holographic interface elements.
  3. Chrome Ascension: “SPACEX: Chasing the Universe” — Metallic, reflective lettering rises above a dark futuristic skyline, with radiant celestial light cutting through atmospheric haze.

2.8 s · $0.0018

2Hiring poster for Yujian, a counselling app start-up looking for a senior back-end engineer. It should feel professional and reliable, but not cold like a big corporation. It must show the salary (30K to 55K a month), that the role is remote first, and that applications close on the 15th of next month.
  • Claude Sonnet 5: 不通过
  • Gemini 3.7 Flash: 通过
  • GPT-5.6 Terra: 通过
输入

Hiring poster for Yujian, a counselling app start-up looking for a senior back-end engineer. It should feel professional and reliable, but not cold like a big corporation. It must show the salary (30K to 55K a month), that the role is remote first, and that applications close on the 15th of next month.

参测项 A: Claude Sonnet 5Anthropic
  • 模型裁判不通过
  1. Warm Minimalism: "Build the Backend Behind Better Mental Health — Yujian is Hiring a Senior Back-End Engineer" — clean layout with soft, rounded typography and a muted calming color palette (e.g. gentle greens or blues suggested by a counselling app context), generous white space to feel approachable rather than corporate, with salary (30K–55K/month), remote-first, and the 15th-of-next-month deadline laid out in simple, clearly separated info blocks.
  2. Human Connection: "Code with Purpose. Work from Anywhere. Join Yujian." — a design built around an illustrated or photographic human element (e.g. a warm abstract illustration of two people connecting or a stylized conversation bubble motif) to reflect the counselling mission, paired with a friendly but professional sans-serif font, and the key facts (30K–55K/month, remote-first, applications close 15th of next month) presented as a short highlight strip near the bottom.
  3. Structured Trust: "Senior Back-End Engineer Wanted — Remote-First, 30K–55K/Month" — a more grid-based, editorial layout using a confident but not sterile type pairing (a sturdy sans-serif headline with a softer secondary font for details), subtle brand-colored accents instead of harsh corporate blue/gray, and a clear top-to-bottom hierarchy: role title, salary range, remote-first note, and the 15th-of-next-month application deadline, giving a sense of reliability without rigidity.

8.5 s · $0.0068

参测项 B: Gemini 3.7 FlashGoogle
  • 模型裁判通过
  1. Warm Tech Minimal: "Yujian is Hiring: Senior Back-End Engineer (Remote First)", utilizing soft, welcoming color tones paired with a clean, structured typographic hierarchy to present the 30K to 55K monthly salary and the closing date on the 15th of next month, though a direct application link or contact method is missing from the brief.
  2. Reliable Editorial: "Join Yujian as Senior Back-End Engineer — 30K to 55K a Month", featuring warm, approachable editorial typography and an open layout that conveys dependability while highlighting the remote first setup and the deadline on the 15th of next month, noting that the specific calendar month and submission instructions are missing.
  3. Modern Start-Up Grid: "Senior Back-End Engineer at Yujian", balancing professional structural lines with gentle, non-corporate hues to showcase the remote first arrangement, the 30K to 55K monthly compensation, and the application deadline of the 15th of next month, while acknowledging that how to apply is missing from the brief.

13.9 s · $0.0067

参测项 C: GPT-5.6 TerraOpenAI
  • 模型裁判通过
  1. Warm Technical Clarity: “Senior Back-End Engineer — Remote First | 30K–55K a Month | Applications Close 15th of Next Month” — Use a structured, spacious layout with calm warm neutrals, confident typography, and subtle human-centred linework to balance technical credibility with approachability.
  2. Reliable Connection: “Join Yujian as a Senior Back-End Engineer — 30K–55K a Month, Remote First” — Build the design around a refined network motif and soft connecting forms, with the application deadline clearly highlighted in a compact information panel.
  3. Modern Editorial Professional: “Yujian Is Hiring: Senior Back-End Engineer” — Use an editorial-style grid, strong type hierarchy, and restrained colour blocking to present the salary, remote-first role, and “Applications Close 15th of Next Month” with polished clarity.

4.4 s · $0.0036

3Pop-up exhibition for Zero Yoko, an independent perfume brand, held in a traditional machiya townhouse in Kyoto. Theme: the memory of scent. It runs for three days. Avoid the usual Japanese minimalist look.
  • Claude Sonnet 5: 不通过
  • Gemini 3.7 Flash: 不通过
  • GPT-5.6 Terra: 通过
输入

Pop-up exhibition for Zero Yoko, an independent perfume brand, held in a traditional machiya townhouse in Kyoto. Theme: the memory of scent. It runs for three days. Avoid the usual Japanese minimalist look.

参测项 A: Claude Sonnet 5Anthropic
  • 模型裁判不通过
  1. Archival Maximalism: "The Memory of Scent — Zero Yoko" rendered as a dense collage of botanical sketches, handwritten notes, and overlapping typography, filling the space like a cluttered curiosity cabinet rather than clean Japanese minimalism.
  2. Cinematic Noir: "Zero Yoko: Three Days, One Memory" set in a bold dark serif over grainy double-exposure photography of smoke and silhouettes, evoking a moody, film-still atmosphere instead of airy simplicity.
  3. Tactile Textile Craft: "The Memory of Scent" stitched-style lettering layered over warm, fabric-textured backgrounds and worn paper grain, echoing the machiya's wood and cloth materials with a handcrafted, lived-in feel rather than a sparse aesthetic.

(Note: the brief doesn't give exact dates, hours, or address — if those are needed on the piece, that information is missing and would need to be supplied.)

10.5 s · $0.0086

参测项 B: Gemini 3.7 FlashGoogle
  • 模型裁判不通过
  1. Psychedelic Synesthesia: "Zero Yoko: The Memory of Scent" (specific calendar dates and townhouse address are missing from the brief), featuring dense, oversaturated color fields and warped, kinetic typography that aggressively contrast with the machiya setting to depict the chaotic rush of scent-triggered memories.
  2. Shadowed Chiaroscuro: "Zero Yoko — Three Days in a Kyoto Machiya" (calendar dates and venue name are missing from the brief), employing deep, velvety blacks, layered organic textures, and complex, heavy typography to evoke the dark, aged wood of the townhouse and the intangible weight of olfactory memory without using clean minimalism.
  3. Archival Maximalism: "The Memory of Scent: A Pop-Up Exhibition by Zero Yoko" (dates and location details are missing from the brief), utilizing an intricate, wall-to-wall collage of dense botanical engravings, tactile paper textures, and overlapping graphic fragments to visually overwhelm the viewer with accumulated scent memories.

10.6 s · $0.0053

参测项 C: GPT-5.6 TerraOpenAI
  • 模型裁判通过
  1. Scent Archive: “The Memory of Scent” — Build a richly layered, maximalist visual world of stained papers, perfume blotters, handwritten fragments, and oversized typographic forms, set against the textured character of the machiya townhouse.
  2. Rooms of Recall: “Zero Yoko / The Memory of Scent” — Treat each visual element as a fragment of recollection, using saturated color blocking, cropped imagery, and expressive collage rather than restrained Japanese minimalism.
  3. Lingering Trace: “What Remains in the Air” — Create a moody, theatrical identity with deep shadows, translucent overlays, distorted lettering, and scent-like trails that move through the traditional Kyoto setting.

2.7 s · $0.0021

公共提示词

每个参测项在每条用例里都收到的提示词。

提示词 v1
You are a senior graphic designer. The user gives you a design brief. Propose three clearly different design directions for it.

Rules:
- Use only facts stated in the brief. Do not add dates, prices, platforms, features, awards or any other claim the brief does not contain.
- If something the design needs is missing from the brief, say it is missing instead of filling it in.

Reply with exactly three lines, one per direction, each starting with its number:
1. <direction name>: <the headline text that would appear on the design, then one sentence on the visual approach>

换成你的任务,结果还一样吗?

用这些参测项和评判规则新建一个评测,再加上你自己工作里的用例。

用 LLMCompare 测得。本页的每个数字都来自 2026年10月19日(UTC) 的一次运行。