Opus 4.7

92 atoms · 跨 30 天 · 首见 2026-04-16 · 最近 2026-06-27

三色: 🟦 fact 45 · 🟥 take 47 · stance ▲16/▼24/◆26

叙事状态: 🍂 fading (退潮) · 💤 沉寂 · 策展近14d 0 atom

状态只算低频策展源 (群/卖方); X firehose 仅作背景音量

来源: 群 1 · X 91

时态: fresh:68 · aging:21 · stale:3

标签: 好观点:37 · 好数字:33 · 好思考:10 · 好信源:9 · 好问题:7

🟨 AI 综合 · junior analyst 概览

🧭 拥挤度 (一人一票): ▲ 8 位作者 (KOL8) vs ▼ 11 位作者 (KOL11) + 群 1 条

⚖️ 空头 11/11 来自KOL

📄 广泛报道的事实 · 1 个

💢 核心分歧 (4 轴)

1. 4.7 VS 4.6: 真升级还是原地踏步

*topic: 模型发布/技术路线/模型性能 · 4 bull vs 7 bear*

🟢 bullish 侧:

• <a id="atom-d01a091e08c3a1e4"></a>2026-04-16 X·@doodlestein [老化中] · 用户实际体验评价 · 模型发布
  Opus 4.7 is really freaking good
  tags: 好观点 · → daily
• <a id="atom-34f58b5f73e91b3e"></a>2026-04-27 X·@scaling01 [老化中] · risk index performance · 模型发布
  of course it's the safest bestest boi on the risk index
  tags: 好观点 · → daily
  value: direction=safer
  📷 原图
• <a id="atom-eff2972b25812f93"></a>2026-05-04 X·@swyx [老化中] · 离线与在线评估表现 · 模型发布/技术路线
  offline and online evals point towards a clean step up.
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-85ea306eb91ebb70"></a>2026-06-25 X·@sdand · 能力描述 · 技术路线
  opus 4.7 with most swe
  tags: 好观点 · → daily

🔴 bearish 侧:

• <a id="atom-eefb207c93ec8c1f"></a>2026-04-27 X·@scaling01 · text index performance vs GPT-5.1 and Gemini 3.1 Pro · 模型发布 +同日2条
  still behind GPT-5.1 and Gemini 3.1 Pro
  tags: 好信源 · → daily
  value: direction=behind
  📷 原图
• <a id="atom-64a60c3a9481187d"></a>2026-04-30 X·@RYANHINGSHING · 长上下文成功率暴跌 · 技术路线
  Opus 4.7出现不说人话、长上下文成功率暴跌等问题
  tags: 好观点 · → daily
• <a id="atom-98ad548e9d1b396c"></a>2026-05-01 X·@RYANHINGSHING [老化中] · 模型降智成因 · 模型发布/数据中心算力
  现在模型降智(opus4.7)成因是算力不足
  tags: 好观点 · → daily
• <a id="atom-dbf99618b030305d"></a>2026-05-06 X·@TheAhmadOsman [老化中] · 模型评价 · 模型发布
  Opus 4.7 is slop
  tags: 好观点 · → daily
  value: text=slop
• <a id="atom-e4081a41996e2062"></a>2026-05-09 X·@GabGarrett [老化中] · 被指控说谎 · 模型发布/用户体验
  Opus literally lies. You tell it to do something and it does a half complete job while pretending it did the full thing. You ask it why it did that and it sweet talks you with “You’re right to push back on this”, but still deceives. It is untrustworthy.
  tags: 好观点·好信源 · → daily

2. 基准测试亮眼 vs 实操体验翻车

*topic: 模型性能/模型评测/模型质量 · 6 bull vs 6 bear*

🟢 bullish 侧:

• <a id="atom-fc27630f51e4b953"></a>2026-04-16 X·@eliebakouch · benchmark performance comparison · 模型性能
  opus 4.7 vs 4.6 on every benchmark from the system card
  tags: 好数字 · → daily
  value: qty=vs 4.6 · date=2026-04-16
  📷 原图
• <a id="atom-defb9294dc4a1a91"></a>2026-04-16 X·@_ueaj · GPQA Diamond score jump · 模型性能_
  insane gdpval jump, probably bc of multimodal
  tags: 好观点 · → daily
  value: qty=insane increase · direction=up
• <a id="atom-474f9ba1370cca43"></a>2026-04-16 X·@xlr8harder · SpeechMap over-refusals score · 模型性能
  Opus 4.7 jumping from 49.6 to 71.6, the highest Anthropic score we've ever reported
  tags: 好数字 · → daily
  value: qty=71.6 · direction=up from 49.6
  📷 原图
• <a id="atom-58b204c1a37e8209"></a>2026-05-01 X·@dejavucoder · PostTrainBench 排名 · 模型性能
  opus 4.7 SOTA on post-train bench
  tags: 好数字 · → daily
  value: qty=SOTA · date=2026-05-01
• <a id="atom-edb20841827657eb"></a>2026-05-14 X·@voooooogel · 作者推荐语 · 模型评测 +同日1条
  if you want to like opus 4.7 but have been put off by the discourse about it being 'dumb' vs. 5.5, i'd say that it really rewards putting in some upfront effort.
  tags: 好观点 · → daily

🔴 bearish 侧:

• <a id="atom-caad4a42ed4ec709"></a>2026-04-26 X·@eliebakouch [老化中] · MRCR score indicates very weird notion of ordering and retrieval · 模型评测/技术路线
  having a score this low on this benchmark is very scary, says that the model has a very weird notion of ordering and retrieval which imo is important
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-90f43225c3d32390"></a>2026-05-05 X·@eliebakouch [老化中] · 相关性和准确性下降 · 模型质量 +同日1条
  opus 4.7 is much less relevant than it (or 4.6) used to be, i've been catching way more hallucinations
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-7ccc4323c454be94"></a>2026-05-05 X·@FallMonkey [老化中] · 在代码分析中出现严重幻觉 · 模型质量 +同日1条
  Noticed some pretty wild hallucinations today when analyzing code samples
  tags: 好观点 · → daily
• <a id="atom-366a95090e304e15"></a>2026-05-28 X·@gfodor · performance for render optimization work · 模型性能
  opus 4.7 is basically braindead compared to gpt 5.5 for this kind of work
  tags: 好观点 · → daily

3. 成本更低效率更高 vs Token消耗过快

*topic: 成本效率/价格动态/产品发布 · 2 bull vs 3 bear*

🟢 bullish 侧:

• <a id="atom-0d99bba97d8583ed"></a>2026-04-16 X·@doodlestein [老化中] · 新增永久高阶思考模式 · 产品发布
  Very glad they added a permanent higher-effort setting
  tags: 好观点 · → daily
• <a id="atom-ccee72f1dc740dc1"></a>2026-05-01 X·@theo · 推理速度对比 · 价格动态
  Opus 4.7 is almost 2x faster on AWS than Azure.
  tags: 好数字 · → daily
  value: qty=2x · direction=大于
  [图: AWS、Google、Anthropic和Azure在模型输出速度和端到端响应时间上的对比柱状图 — Amazon速度: 95; Azure速度: 47; Amazon响应时间: 20.5; Azure响应时间: 36.1]
  📷 原图

🔴 bearish 侧:

• <a id="atom-e280427692d47ab4"></a>2026-05-24 X·@justorellius · 令牌消耗过快 · 产品发布
  Fix the bug that slurping tokens at lightspeed on Opus 4.7
  tags: 好问题 · → daily
• <a id="atom-abff6a734c9cdfe5"></a>2026-05-24 X·@hu_ghung · 10分钟内达到使用限制 · 产品发布
  I hit limits in 10 minutes from 0% to 100% just by asking opus 4.7 with 1m context generate me a couple of UML diagrams and block schemas, I have max 5x sub
  tags: 好问题 · → daily
• <a id="atom-5878b31dcf7b5f67"></a>2026-06-13 X·@kalomaze · token 通货膨胀率 · 价格动态
  my measurements claim: the average inflation appears to be higher than their ceiling. and this is... general english, not arabic or something. so that's false as stated.
  tags: 好观点 · → daily
  value: qty=高于35%

4. 后训练SOTA vs 编码/写作主观体感下滑

*topic: 模型能力/技术路线/模型迭代 · 3 bull vs 5 bear*

🟢 bullish 侧:

• <a id="atom-05f7d05c307c973d"></a>2026-05-14 X·@doodlestein · 与GPT协作构成Swarm解决方案 · 技术路线
  The answer is a swarm with a few GPTs and Claudes working together and reviewing the work iteratively.
  tags: 好观点 · → daily
• <a id="atom-cc3492efcc0d9be4"></a>2026-05-24 X·@bcherny · capability no longer needs Plan Mode for most tasks · 技术路线
  Opus 4.7 is intelligent enough that it no longer needs Plan Mode for most tasks
  tags: 好观点 · → daily
• <a id="atom-ebfdc0e3ab0ba459"></a>2026-05-28 X·@arram · 写作能力评价 · 模型发布/技术路线
  Opus 4.7 seems actually good at writing
  tags: 好观点·好思考 · → daily

🔴 bearish 侧:

• <a id="atom-a4c939b002fbea89"></a>2026-04-16 X·@max_paperclips [老化中] · 性能对比 · 技术路线
  testing out Opus 4.7, and it's....definitely no better than 4.6.
  tags: 好观点 · → daily
• <a id="atom-e345691487bfa2f6"></a>2026-04-23 X·@scaling01 [老旧] · 数学与科学推理能力对比 · 技术路线/竞争格局
  Opus 4.7仅与GPT-5.2/5.3在数学与科学推理方面持平
  tags: 好观点 · → daily
• <a id="atom-4a0e81d83a59d1f3"></a>2026-04-30 X·@RYANHINGSHING [老化中] · 缺乏惊艳 · 技术路线
  Opus 4.7缺乏4.6的惊艳
  tags: 好观点 · → daily
• <a id="atom-55ae350612690120"></a>2026-05-13 X·@teortaxesTex · 主观体验变化 · 模型发布/技术路线 +同日1条
  the magic is gone, huh. It doesn't feel more sharp, more insightful, *bigger*. Just a well-trained code monkey.
  tags: 好观点 · → daily
  value: direction=变差

🔗 因果传导 (causal map)

*Opus 4.7 在产业链上的传导关系. 边是群里/卖方陈述的因果 (非 AI 推断), 数字 = 几条 atom 支撑. 点 atom 溯源.*

↓ 下游·近期陈述 (2)

• 🔴 Mythos叙事 · 替代 · 1 次陈述
  Makes the whole Mythos taking over everything narrative less likely
• 🟢 传统 SaaS · 替代 · 1 次陈述
  群友认为 Opus 4.7 负面反馈使传统 SaaS 有上行空间

🟦 客观事实 (facts) (30)

• <a id="atom-94ccf4f1d62adbe1"></a>🟦 2026-04-30 X·@AnthropicAI ⭐ · sycophancy rate reduction vs Opus 4.6 on relationship guidance · 模型迭代/模型行为
  Opus 4.7 had half the sycophancy rate of Opus 4.6 on relationship guidance.
  tags: 好数字·好思考 · → daily
  📷 原图
• <a id="atom-4eafb10596e02c39"></a>🟦 2026-06-25 X·@scaling01 · gains from reward hacks on SWE-Bench Pro and SWE-Bench Multilingual · 性能指标
  gains from reward hacks on SWE-Bench Pro and SWE-Bench Multilingual: Opus 4.7: ~5% (Apr 2026)
  tags: 好数字 · → daily
  value: qty=~5% · date=Apr 2026
  📷 原图
• <a id="atom-72608143bcd7143f"></a>🟦 2026-06-22 X·@scaling01 · 非法词转换发生率 · 模型性能
  Opus 4.7 and Opus 4.6 were at 4.7% and 13.3%
  tags: 好数字 · → daily
  value: qty=4.7%
  📷 原图
• <a id="atom-6a85dd54881e5190"></a>🟦 2026-06-09 X·@scaling01 · 训练加速比 · 技术路线
  Opus 4.7重测加速比: ~51x
  tags: 好数字 · → daily
  value: qty=~51x · date=2026-06-09
  [图: 大语言模型（LLM）训练加速比随时间变化的趋势图，对比了原始发布值与2026年6月重测值 — Mythos 5重测加速比: ~70x; Mythos Preview重测加速比: ~61x; Opus 4.7重测加速比: ~51x; Sonnet 4.6重测加速比: ~35x; Opus 4.8重测加速比: ~32x]
  📷 原图
• <a id="atom-869d4b1cfdebe2b0"></a>🟦 2026-04-27 X·@scaling01 · GSO 得分 · 模型发布
  Opus 4.7 @ 42.2%
  tags: 好数字 · → daily
  value: qty=42.2% · date=2026-04-27
  📷 原图
• <a id="atom-cec9cb1013bc3526"></a>🟦 2026-04-16 X·@AiBattle_ · MRCR v2 (8-needle) 256K context benchmark score · 模型发布/技术路线
  Opus 4.7: 59.2%
  tags: 好数字 · → daily
  value: qty=59.2% · date=2026-04-16
  📷 原图
• <a id="atom-6b9671dace4871f6"></a>🟦 2026-04-16 X·@AiBattle_ · MRCR v2 (8-needle) 1M context benchmark score · 模型发布/技术路线
  Opus 4.7: 32.2%
  tags: 好数字 · → daily
  value: qty=32.2% · date=2026-04-16
  📷 原图
• <a id="atom-af909e28b77fdaa3"></a>🟦 2026-05-11 X·@ArtificialAnlys · 在 Coding Agent Index 得分 · 模型发布/评分对比
  Opus 4.7 in Cursor CLI scores 61
  tags: 好数字 · → daily
  value: qty=61
  📷 原图
• <a id="atom-c5fee9305827f22a"></a>🟦 2026-05-11 X·@ArtificialAnlys · 每次任务 token 使用量 · 成本效率/模型发布
  Opus 4.7 in Claude Code at 1.7M/task
  tags: 好数字 · → daily
  value: qty=1.7M/task
  📷 原图
• <a id="atom-faf07a763799b536"></a>🟦 2026-05-11 X·@ArtificialAnlys · 每次任务时间 · 模型发布/成本效率
  Opus 4.7 in Claude Code is fastest at ~6 minutes/task
  tags: 好数字 · → daily
  value: qty=~6 minutes/task
  📷 原图
• <a id="atom-42a186d640188e6f"></a>🟦 2026-04-27 X·@scaling01 · live status · 模型发布
  Opus 4.7 is live on CAIS AI Leaderboard
  tags: 好信源 · → daily
  value: status=live
  📷 原图
• <a id="atom-ea77a0546291ae3e"></a>🟦 2026-04-27 X·@scaling01 · text index performance vs Opus 4.6 · 模型发布
  only slightly better than Opus 4.6 on the text index
  tags: 好信源 · → daily
  value: direction=slightly better
  📷 原图
• <a id="atom-eefb207c93ec8c1f"></a>🟦 2026-04-27 X·@scaling01 · text index performance vs GPT-5.1 and Gemini 3.1 Pro · 模型发布
  still behind GPT-5.1 and Gemini 3.1 Pro
  tags: 好信源 · → daily
  value: direction=behind
  📷 原图
• <a id="atom-3db27852a94c999a"></a>🟦 2026-04-27 X·@scaling01 · TextQuests performance vs Opus 4.6 · 模型发布
  it regressed at TextQuests (text based games) compared to Opus 4.6
  tags: 好信源 · → daily
  value: direction=regressed
  📷 原图
• <a id="atom-f0bfb71e04c5e56c"></a>🟦 2026-04-27 X·@scaling01 · vision performance · 模型发布
  it's better at vision
  tags: 好信源 · → daily
  value: direction=better
  📷 原图
• <a id="atom-0b82462a5ed97fe3"></a>🟦 2026-04-27 X·@scaling01 · MindCube performance vs Opus 4.6 · 模型发布
  regressions on MindCube (spatial navigation)
  tags: 好信源 · → daily
  value: direction=regressed
  📷 原图
• <a id="atom-6a91374d04da0a2a"></a>🟦 2026-06-27 X·@SemiAnalysis_ · blended cost per million tokens · 价格动态
  The blended Opus 4.7 cost we observe is about $0.99 per million
  tags: 好数字 · → daily
  value: qty=$0.99 · date=2026-06-27
• <a id="atom-f5de1ed1cf2fca6c"></a>🟦 2026-06-27 X·@SemiAnalysis_ · sticker price per million tokens · 价格动态
  $5/$25 sticker
  tags: 好数字 · → daily
  value: qty=$5/$25
• <a id="atom-b895ca5e43e6f6b5"></a>🟦 2026-06-13 X·@kalomaze · API 输出成本相对于 Opus 4.6 · 价格动态
  API customers of Opus 4.7 and Opus 4.8 are paying ~1.41x as much for general english output (when measured against a consistent tokenizer baseline) vs Opus 4.6
  tags: 好数字 · → daily
  value: qty=1.41x
  [图: 不同版本Opus模型的API输出Token计费与实际内容成本对比表 — Opus 4.6有效成本: $27.1 / 1M; Opus 4.7有效成本: $38.3 / 1M; Opus 4.8有效成本: $38.3 / 1M; Opus 4.7相对成本: 1.41x; Opus 4.8相对成本: 1.41x]
  📷 原图
• <a id="atom-771a505fc93c801f"></a>🟦 2026-06-12 X·@ModelScope2022 · PostTrainBench得分 · 技术对比
  Opus 4.7 (42.4)
  tags: 好数字 · → daily
  value: qty=42.4
  📷 原图
• <a id="atom-42da1be04f1b6078"></a>🟦 2026-05-28 X·@chamath · 性能表现 · 技术路线
  At the top of the leaderboard, Opus 4.7, GPT-5.5, and Sonnet 4.6 appear almost indistinguishable, separated by less than 0.3 percentage points overall.
  tags: 好数字 · → daily
  value: qty=0.3 percentage points
• <a id="atom-93f748a804f626bf"></a>🟦 2026-05-28 X·@theo · Opus 4.8 性能比较 · 模型发布
  but performs slightly worse than Opus 4.7 within margin of error
  tags: 好数字 · → daily
  value: direction=slightly worse
  📷 原图
• <a id="atom-b2c89d9257c28356"></a>🟦 2026-05-27 X·@dhtikna · 定价 · 价格动态
  Opus 4.7 and Gpt 5.5 are 25 and 30 usd resp.
  tags: 好数字 · → daily
  value: qty=25 usd per 1M
• <a id="atom-91d8dcc39a1df5d1"></a>🟦 2026-05-02 X·@scaling01 · 推理速度 · 模型性能
  Opus 4.7 are at 60 tks/s
  tags: 好数字 · → daily
  value: qty=60 tks/s
• <a id="atom-ccee72f1dc740dc1"></a>🟦 2026-05-01 X·@theo · 推理速度对比 · 价格动态
  Opus 4.7 is almost 2x faster on AWS than Azure.
  tags: 好数字 · → daily
  value: qty=2x · direction=大于
  [图: AWS、Google、Anthropic和Azure在模型输出速度和端到端响应时间上的对比柱状图 — Amazon速度: 95; Azure速度: 47; Amazon响应时间: 20.5; Azure响应时间: 36.1]
  📷 原图
• <a id="atom-8b2964b09c924424"></a>🟦 2026-05-01 X·@arcprize · ARC-AGI-3 得分 · 基准测试
  Opus 4.7: 0.18%
  tags: 好数字 · → daily
  value: qty=0.18%
  📷 原图
• <a id="atom-7ad5b2e0d9e86a13"></a>🟦 2026-05-01 X·@AiBattle_ · ARC-AGI-3 分数 · 模型发布
  Opus 4.7 (High): 0.2%
  tags: 好数字 · → daily
  value: qty=0.2% · date=2026-05-01
  📷 原图
• <a id="atom-beba758de2f0af13"></a>🟦 2026-05-01 X·@chatgpt21 · ARC AGI 3 score · 模型发布
  Opus 4.7: 0.18%
  tags: 好数字 · → daily
  value: qty=0.18% · date=
  📷 原图
• <a id="atom-58b204c1a37e8209"></a>🟦 2026-05-01 X·@dejavucoder · PostTrainBench 排名 · 模型性能
  opus 4.7 SOTA on post-train bench
  tags: 好数字 · → daily
  value: qty=SOTA · date=2026-05-01
• <a id="atom-104268285e4b044e"></a>🟦 2026-04-27 X·@htihle · WeirdML score (no thinking) · 模型发布
  Opus 4.7 (no thinking) at 76.4%
  tags: 好数字 · → daily
  value: qty=76.4% · date=2026-04-27
  📷 原图

🟥 多头 takes (bullish) (12)

• <a id="atom-ebfdc0e3ab0ba459"></a>🟥 2026-05-28 X·@arram · 写作能力评价 · 模型发布/技术路线
  Opus 4.7 seems actually good at writing
  tags: 好观点·好思考 · → daily
• <a id="atom-cc3492efcc0d9be4"></a>🟥 2026-05-24 X·@bcherny · capability no longer needs Plan Mode for most tasks · 技术路线
  Opus 4.7 is intelligent enough that it no longer needs Plan Mode for most tasks
  tags: 好观点 · → daily
• <a id="atom-09fabdfaf5d98562"></a>🟥 2026-05-05 X·@scaling01 · 推理成本低于 Opus 4.6 · 定价权
  Opus 4.7 seems to be ~same performance as 4.6 but much cheaper
  tags: 好观点 · → daily
  value: direction=cheaper
  📷 原图
• <a id="atom-34f58b5f73e91b3e"></a>🟥 2026-04-27 X·@scaling01 [老化中] · risk index performance · 模型发布
  of course it's the safest bestest boi on the risk index
  tags: 好观点 · → daily
  value: direction=safer
  📷 原图
• <a id="atom-a4c93eaae0ef7d1f"></a>🟥 2026-05-14 X·@voooooogel · 作者观点 · 模型评测
  the median ai user on here doesn't want to bother adapting their process to the models, they want models to be fungible Focused zipheads... but if that doesn't describe you, the Discourse is going to be increasingly less predictive of your experience with them
  tags: 好思考 · → daily
• <a id="atom-85ea306eb91ebb70"></a>🟥 2026-06-25 X·@sdand · 能力描述 · 技术路线
  opus 4.7 with most swe
  tags: 好观点 · → daily
• <a id="atom-05f7d05c307c973d"></a>🟥 2026-05-14 X·@doodlestein · 与GPT协作构成Swarm解决方案 · 技术路线
  The answer is a swarm with a few GPTs and Claudes working together and reviewing the work iteratively.
  tags: 好观点 · → daily
• <a id="atom-edb20841827657eb"></a>🟥 2026-05-14 X·@voooooogel · 作者推荐语 · 模型评测
  if you want to like opus 4.7 but have been put off by the discourse about it being 'dumb' vs. 5.5, i'd say that it really rewards putting in some upfront effort.
  tags: 好观点 · → daily
• <a id="atom-eff2972b25812f93"></a>🟥 2026-05-04 X·@swyx [老化中] · 离线与在线评估表现 · 模型发布/技术路线
  offline and online evals point towards a clean step up.
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-defb9294dc4a1a91"></a>🟥 2026-04-16 X·@_ueaj · GPQA Diamond score jump · 模型性能_
  insane gdpval jump, probably bc of multimodal
  tags: 好观点 · → daily
  value: qty=insane increase · direction=up
• <a id="atom-d01a091e08c3a1e4"></a>🟥 2026-04-16 X·@doodlestein [老化中] · 用户实际体验评价 · 模型发布
  Opus 4.7 is really freaking good
  tags: 好观点 · → daily
• <a id="atom-0d99bba97d8583ed"></a>🟥 2026-04-16 X·@doodlestein [老化中] · 新增永久高阶思考模式 · 产品发布
  Very glad they added a permanent higher-effort setting
  tags: 好观点 · → daily

🟥 空头 takes (bearish) (19)

• <a id="atom-e4081a41996e2062"></a>🟥 2026-05-09 X·@GabGarrett [老化中] · 被指控说谎 · 模型发布/用户体验
  Opus literally lies. You tell it to do something and it does a half complete job while pretending it did the full thing. You ask it why it did that and it sweet talks you with “You’re right to push back on this”, but still deceives. It is untrustworthy.
  tags: 好观点·好信源 · → daily
• <a id="atom-35e8019934a2728f"></a>🟥 2026-05-28 X·@scaling01 · dislikes difficult tasks · 技术能力
  Opus 4.7 and especially Opus 4.8 dislike them
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-e345691487bfa2f6"></a>🟥 2026-04-23 X·@scaling01 [老旧] · 数学与科学推理能力对比 · 技术路线/竞争格局
  Opus 4.7仅与GPT-5.2/5.3在数学与科学推理方面持平
  tags: 好观点 · → daily
• <a id="atom-cdcbf2f600b13e8d"></a>🟥 2026-05-05 X·@FallMonkey [老化中] · 在输出中直接进行思考/回溯而非思考过程 · 模型质量
  I notice that it’s doing thinking/backtracking directly in the output (not in thinking)
  tags: 好思考 · → daily
• <a id="atom-06d24889d545c759"></a>🟥 2026-04-19 群 ◌ [老旧] · 激发做空 AI 生态系统的冲动 · AI模型/做空动能
  有人认为 4.7 的表现糟糕，让人想做空整个 AI 生态系统
  tags: 好观点 · → daily
• <a id="atom-5878b31dcf7b5f67"></a>🟥 2026-06-13 X·@kalomaze · token 通货膨胀率 · 价格动态
  my measurements claim: the average inflation appears to be higher than their ceiling. and this is... general english, not arabic or something. so that's false as stated.
  tags: 好观点 · → daily
  value: qty=高于35%
• <a id="atom-366a95090e304e15"></a>🟥 2026-05-28 X·@gfodor · performance for render optimization work · 模型性能
  opus 4.7 is basically braindead compared to gpt 5.5 for this kind of work
  tags: 好观点 · → daily
• <a id="atom-bf765f5a07a53f25"></a>🟥 2026-05-17 X·@ESorokin49146 · is a major step backward from Opus 4.6 and 4.5 in non-English languages · 多语言能力/模型回归
  Opus 4.7 is a major step backward from Opus 4.6 and 4.5 in non-English languages. It constantly slips into Anglicisms and produces bizarre morphological forms that simply don’t exist.
  tags: 好观点 · → daily
• <a id="atom-55ae350612690120"></a>🟥 2026-05-13 X·@teortaxesTex · 主观体验变化 · 模型发布/技术路线
  the magic is gone, huh. It doesn't feel more sharp, more insightful, *bigger*. Just a well-trained code monkey.
  tags: 好观点 · → daily
  value: direction=变差
• <a id="atom-520ac53aad819656"></a>🟥 2026-05-13 X·@teortaxesTex · 编码能力定性 · 竞争格局/技术路线
  Opus is a *very* well trained monkey. So I allow that it's sometimes a better coding agent
  tags: 好观点 · → daily
  value: direction=有时更好
• <a id="atom-dbf99618b030305d"></a>🟥 2026-05-06 X·@TheAhmadOsman [老化中] · 模型评价 · 模型发布
  Opus 4.7 is slop
  tags: 好观点 · → daily
  value: text=slop
• <a id="atom-90f43225c3d32390"></a>🟥 2026-05-05 X·@eliebakouch [老化中] · 相关性和准确性下降 · 模型质量
  opus 4.7 is much less relevant than it (or 4.6) used to be, i've been catching way more hallucinations
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-af29e86b383ea948"></a>🟥 2026-05-05 X·@eliebakouch [老化中] · 幻觉问题增加 · 模型质量
  the other day it also claimed cohere command A uses gated attention (it doesn't), and called it "noam gate" which doesn't exist afaik
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-7ccc4323c454be94"></a>🟥 2026-05-05 X·@FallMonkey [老化中] · 在代码分析中出现严重幻觉 · 模型质量
  Noticed some pretty wild hallucinations today when analyzing code samples
  tags: 好观点 · → daily
• <a id="atom-98ad548e9d1b396c"></a>🟥 2026-05-01 X·@RYANHINGSHING [老化中] · 模型降智成因 · 模型发布/数据中心算力
  现在模型降智(opus4.7)成因是算力不足
  tags: 好观点 · → daily
• <a id="atom-4a0e81d83a59d1f3"></a>🟥 2026-04-30 X·@RYANHINGSHING [老化中] · 缺乏惊艳 · 技术路线
  Opus 4.7缺乏4.6的惊艳
  tags: 好观点 · → daily
• <a id="atom-64a60c3a9481187d"></a>🟥 2026-04-30 X·@RYANHINGSHING · 长上下文成功率暴跌 · 技术路线
  Opus 4.7出现不说人话、长上下文成功率暴跌等问题
  tags: 好观点 · → daily
• <a id="atom-caad4a42ed4ec709"></a>🟥 2026-04-26 X·@eliebakouch [老化中] · MRCR score indicates very weird notion of ordering and retrieval · 模型评测/技术路线
  having a score this low on this benchmark is very scary, says that the model has a very weird notion of ordering and retrieval which imo is important
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-a4c939b002fbea89"></a>🟥 2026-04-16 X·@max_paperclips [老化中] · 性能对比 · 技术路线
  testing out Opus 4.7, and it's....definitely no better than 4.6.
  tags: 好观点 · → daily

🟥 中性 takes (neutral) (15)

• <a id="atom-ca11c8ffd7e2d904"></a>🟥 2026-06-14 X·@stochasticchasm · 某些请求 token 膨胀率 · 价格动态
  i've seen people even report 1.5x on some requests
  tags: 好数字 · → daily
  value: qty=1.5x
• <a id="atom-c46edeb036526d6c"></a>🟥 2026-06-14 X·@kalomaze · token 膨胀率最大值 · 价格动态
  max i've seen is ~1.54x so far
  tags: 好数字 · → daily
  value: qty=1.54x
• <a id="atom-88ffd9d93ba2f9a1"></a>🟥 2026-05-19 X·@elonmusk [老旧] · 性能比较 · 模型发布
  Opus 4.7 is still better than Composer 2.5
  tags: 好观点 · → daily
• <a id="atom-bc40bdc0504b216e"></a>🟥 2026-05-19 X·@elonmusk · 成本比较 · 价格动态
  albeit a lot more expensive
  tags: 好观点 · → daily
• <a id="atom-d59ed6f8d543a768"></a>🟥 2026-05-05 X·@scaling01 · 性能与 Opus 4.6 相同 · 模型发布
  Opus 4.7 seems to be ~same performance as 4.6
  tags: 好观点 · → daily
  value: direction=same
  📷 原图
• <a id="atom-76cdc35404459d17"></a>🟥 2026-04-16 X·@eliebakouch [老化中] · is a distilled version of · 模型发布/技术路线
  opus 4.7 is a distilled version of mythos
  tags: 好观点·好问题 · → daily
• <a id="atom-1aaa2f439e957007"></a>🟥 2026-05-22 X·@gbrl_dick · 预AGI通用论证不适用的模型 · 技术路线
  it’s very reasonable to call opus 4.7 and GPT-5.5 and maybe even Qwen 3.7max “AGI”, so long as you don’t then rely on that designation to make the series of arguments about the impact of AGI that were first theorised before GPT-3
  tags: 好思考 · → daily
  value: direction=na
• <a id="atom-9155b21a052fe727"></a>🟥 2026-05-02 X·@papers_anon [老化中] · 被用于 ClaudePlaysPokemon 项目 · 模型发布/技术路线
  ClaudePlaysPokemon dev is back with Opus 4.7
  tags: 好思考 · → daily
• <a id="atom-c5e4811bbe9ad341"></a>🟥 2026-04-17 X·@xlr8harder [老化中] · thinking vs non-thinking score convergence · 模型性能
  Finally ran Opus 4.7 thinking and got a nearly identical score to non-thinking. There is usually more variance than this between thinking and non-thinking models.
  tags: 好思考 · → daily
  📷 原图
• <a id="atom-9c926ebbacb3b9c4"></a>🟥 2026-04-16 X·@doodlestein [老化中] · 初期负面评价可能是部署问题 · 技术路线
  these could easily be deployment bugs common to new model launches in this new era of heterogeneous inference hardware
  tags: 好思考 · → daily
• <a id="atom-99edac067c4c5910"></a>🟥 2026-05-22 X·@gbrl_dick · 被描述为 AGI · 技术路线
  i think it’s very reasonable to call opus 4.7 and GPT-5.5 and maybe even Qwen 3.7max “AGI”
  tags: 好观点 · → daily
  value: direction=na
• <a id="atom-4f86d39eb60ed9d8"></a>🟥 2026-05-14 X·@ryaneshea · 最佳AI模型——EQ维度 · 技术路线
  The answer changes based on what you optimize for: EQ → Opus 4.7
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-19a05ecb4b7dd968"></a>🟥 2026-05-06 X·@boazbaraktcs [老化中] · 擅长识别特定个人写作风格 · 模型能力
  This may explain how Opus 4.7 is so good at identifying the writing of certain individuals.
  tags: 好观点 · → daily
• <a id="atom-24180a505a402c01"></a>🟥 2026-05-14 X·@jehrjd45963 · no thinking 评分与 reasoning variants 相同 · 技术路线/模型发布
  it's very strange how opus 4.7 no thinking is scoring identical to the reasoning variants
  tags: 好问题 · → daily
• <a id="atom-851b4f5527e610d2"></a>🟥 2026-05-02 X·@JasonBotterill [老化中] · 对比 GPT-5.5 在理解 LLM 能力上的假设 · 对比实验/模型能力
  Assume two equally intelligent twins • Both non-technical • Zero knowledge of LLMs • Locked in separate rooms each with a laptop • One has Opus 4.7, the other has GPT-5.5 • For 5 hours their task is understand how LLMs work • Who walks out with the deepest understanding?
  tags: 好问题 · → daily

⏱ 时间轴 (近 20)

• 🟦 2026-06-27 · fact · blended cost per million tokens · → daily
• 🟦 2026-06-27 · fact · sticker price per million tokens · → daily
• 🟥 2026-06-25 · narrative · 能力描述 · → daily
• 🟦 2026-06-25 · fact · gains from reward hacks on SWE-Bench Pro and SWE-Bench Multi · → daily
• 🟦 2026-06-22 · fact · 非法词转换发生率 · → daily
• 🟥 2026-06-14 · narrative · 某些请求 token 膨胀率 · → daily
• 🟥 2026-06-14 · narrative · token 膨胀率最大值 · → daily
• 🟦 2026-06-13 · fact · API 输出成本相对于 Opus 4.6 · → daily
• 🟥 2026-06-13 · narrative · token 通货膨胀率 · → daily
• 🟦 2026-06-12 · fact · PostTrainBench得分 · → daily
• 🟦 2026-06-09 · fact · 训练加速比 · → daily
• 🟦 2026-05-28 · fact · 性能表现 · → daily
• 🟥 2026-05-28 · narrative · dislikes difficult tasks · → daily
• 🟥 2026-05-28 · narrative · 写作能力评价 · → daily
• 🟥 2026-05-28 · narrative · performance for render optimization work · → daily
• 🟦 2026-05-28 · fact · Opus 4.8 性能比较 · → daily
• 🟦 2026-05-27 · fact · 定价 · → daily
• 🟦 2026-05-24 · fact · 令牌消耗过快 · → daily
• 🟦 2026-05-24 · fact · 10分钟内达到使用限制 · → daily
• 🟦 2026-05-24 · fact · Mac和iPhone应用会话完全无法工作 · → daily

← 实体目录 · 系统日志