Opus

45 atoms · 跨 27 天 · 首见 2026-04-15 · 最近 2026-07-02

三色: 🟦 fact 17 · 🟥 take 28 · stance ▲10/▼14/◆7

叙事状态: 🍂 fading (退潮) · 💤 沉寂 · 策展近14d 0 atom · 🐦 X 近14d 6

状态只算低频策展源 (群/卖方); X firehose 仅作背景音量

来源: X 45

时态: fresh:31 · aging:10 · stale:4

标签: 好观点:26 · 好数字:10 · 好信源:5 · 好思考:5 · 好问题:1

🟨 AI 综合 · junior analyst 概览

🧭 拥挤度 (一人一票): ▲ 7 位作者 (KOL7) vs ▼ 8 位作者 (KOL8)

⚖️ 空头 8/8 来自KOL

💢 核心分歧 (2 轴)

1. Opus 版本迭代能力提升 vs 功能回退

*topic: 模型发布/模型能力/性能表现/技术路线 · 6 bull vs 4 bear*

🟢 bullish 侧:

• <a id="atom-7c1101a5b7719e7b"></a>2026-04-16 X·@morqon · 生物结构理解评估分数表现 · 技术路线/模型发布
  opus more than doubles its score in open-ended evals
  tags: 好数字 · → daily
  value: qty=超过两倍提升 · date=2026-04
  📷 原图
• <a id="atom-6e01dd1def8232ca"></a>2026-05-04 X·@teortaxesTex · benches performance · 模型发布
  always astonishing on such benches how filthy rich Opus gets
  tags: 好观点 · → daily
  value: qty=filthy rich
• <a id="atom-2effeda7529b9411"></a>2026-05-11 X·@Dorialexander [老化中] · 科学知识记忆能力出现跃升 · 技术路线/模型发布 +同日1条
  il y a eu un saut impressionnant dans la mémorisation de connaissances scientifiques de niche
  tags: 好观点·好思考 · → daily
• <a id="atom-15a20bd449f47025"></a>2026-05-14 X·@jiaxinwen22 · hill-climbing efficiency gap versus Codex · 技术路线/模型发布
  The hill-climbing efficiency gap between Opus and Codex is much larger than I was expecting!
  tags: 好思考 · → daily
• <a id="atom-af89cd41ec8e70ba"></a>2026-05-31 X·@teortaxesTex · 性能评价 · 模型发布
  OK, this somewhat rescues Opus.
  tags: 好观点 · → daily
  value: direction=positive
  📷 原图

🔴 bearish 侧:

• <a id="atom-14b9b086285c7df2"></a>2026-05-03 X·@ar0cket1 · token consumption · 模型发布
  Opus just consumes more and more
  tags: 好数字 · → daily
• <a id="atom-dc199a8e88ded104"></a>2026-05-11 X·@Dorialexander [老化中] · 模型可能面临自身崩溃问题 · 技术路线/模型发布
  OPD might run into its own collapse issues
  tags: 好思考 · → daily
• <a id="atom-129dcc29bd14dee9"></a>2026-05-29 X·@gfodor · 被高估的智能水平（比人们想的笨） · 技术路线/模型发布
  Opus is dumber than people think
  tags: 好观点 · → daily
• <a id="atom-6d066e85b6783d22"></a>2026-06-10 X·@max_paperclips · ML 能力评价 · 模型发布
  there's been conspiracy theories about Opus being suspiciously bad at ML for a while
  tags: 好观点 · → daily

2. Opus 幻觉与事实准确性争议

*topic: 模型能力/生成能力/性能表现 · 1 bull vs 1 bear*

🟢 bullish 侧:

• <a id="atom-e2a88eafca57d895"></a>2026-05-03 X·@secemp9 [老化中] · no hallucination · 性能表现
  no hallucination or anything on opus past 600k tokens
  tags: 好观点 · → daily
  value: qty=600k tokens · direction=past
  📷 原图

🔴 bearish 侧:

• <a id="atom-e8b4cc3d32bfc0cf"></a>2026-04-16 X·@max_paperclips [老化中] · 变得更容易对事实不严谨 · 模型能力
  Opus has been a little loose with the truth for a while, but this feels....more so
  tags: 好观点 · → daily

🟦 客观事实 (facts) (17)

• <a id="atom-7dd14e990e17377e"></a>🟦 2026-05-06 X·@brickroad7 · model_size_and_training_hardware · 技术路线
  Opus is a 5t model. Trained on TPU it seems. TPU got the scale up in world size slightly before Nvidia.
  tags: 好数字·好信源 · → daily
  value: qty=5t
• <a id="atom-08070cfa34922399"></a>🟦 2026-07-01 X·@thayto_dev · 解題嘗試時間 · 推理效率
  o que o Opus tentou por dias
  tags: 好数字 · → daily
  value: qty=dias · date=2026-07-01
  📷 原图
• <a id="atom-f8d634ba36945003"></a>🟦 2026-06-27 X·@teortaxesTex · 生成速度 · 产品发布
  Opus in the same context just gave me 37
  tags: 好数字 · → daily
  value: qty=37 t/s · context=20K context
  📷 原图
• <a id="atom-717f2139fcdb63e4"></a>🟦 2026-06-27 X·@teortaxesTex · benchmark score · 模型发布/技术路线
  Opus 4.8 (max)
  tags: 好数字 · → daily
  value: qty=4.8
• <a id="atom-d9b0282bacc9f5bc"></a>🟦 2026-06-06 X·@MiniMax_AI · cost per bug found · 成本分析
  $3.39 for Opus
  tags: 好数字 · → daily
  value: qty=$3.39
• <a id="atom-14b9b086285c7df2"></a>🟦 2026-05-03 X·@ar0cket1 · token consumption · 模型发布
  Opus just consumes more and more
  tags: 好数字 · → daily
• <a id="atom-ac0f7138a38d133a"></a>🟦 2026-04-27 X·@tensorqt · 生成长篇内容能力 · 生成能力
  time to read my 200k-word book generated with opus 4.7
  tags: 好数字 · → daily
  value: qty=200k字 · date=2026-04-27
• <a id="atom-7c1101a5b7719e7b"></a>🟦 2026-04-16 X·@morqon · 生物结构理解评估分数表现 · 技术路线/模型发布
  opus more than doubles its score in open-ended evals
  tags: 好数字 · → daily
  value: qty=超过两倍提升 · date=2026-04
  📷 原图
• <a id="atom-af7e0bd449b5995b"></a>🟦 2026-07-02 X·@a_musingcat · 计费方式 · 定价
  if u get switched to opus, it bills as a cache read
  tags: 好信源 · → daily
• <a id="atom-9225d725fa36bdba"></a>🟦 2026-06-25 X·@LinXule · 发布版本 · 产品发布
  opus4.8 on working with kimi (k2.6), codex (gpt5.5), and gemini (3.5 flash)
  tags: 好信源 · → daily
  value: qty=4.8 · date=2026-06-25
  📷 原图
• <a id="atom-92b1796d02f59bd9"></a>🟦 2026-06-12 X·@jehyeoky248 · 被加入InferenceBench排行榜 · 模型发布
  Opus 4.8
  tags: 好信源 · → daily
  📷 原图
• <a id="atom-363ba5cecaee6b10"></a>🟦 2026-06-09 X·@Dorialexander · 记忆测试通过 · 模型发布
  Fables do pass my new hard memorization tests (opus is just hedging: " I can't, from memory, confidently [state] without risking fabrication")
  tags: 好信源 · → daily
  📷 原图
• <a id="atom-3768e38d99eadc6d"></a>🟦 2026-05-29 X·@mattparlmer · 能力回退 · 模型性能
  Unfortunate capability regression here
  tags: 好思考 · → daily
  value: direction=退化 · date=2026-05-29
• <a id="atom-bdbcd23d64dc2ebf"></a>🟦 2026-07-01 X·@trq212 · 定价模式 · 定价权
  no, you're charged Opus prices and we pay for the prompt cache miss as well
  tags: 好观点 · → daily
  value: direction=
• <a id="atom-a52934538b9d3b19"></a>🟦 2026-06-04 X·@MParakhin · 推理努力设置 · 技术路线
  Opus is Max
  tags: 好观点 · → daily
  value: direction=Max
• <a id="atom-1f97f665d3cafb12"></a>🟦 2026-05-29 X·@teortaxesTex · 视觉能力差 · 技术对比
  it's worse for CAD stuff than DS-Vision
  tags: 好观点 · → daily
  value: direction=差 · topic=CAD任务
• <a id="atom-50ef1746b646da41"></a>🟦 2026-07-02 X·@ctjlewis · 付费承担方 · 定价
  Oh, they pay for this for now?
  tags: 好问题 · → daily

🟥 多头 takes (bullish) (9)

• <a id="atom-2effeda7529b9411"></a>🟥 2026-05-11 X·@Dorialexander [老化中] · 科学知识记忆能力出现跃升 · 技术路线/模型发布
  il y a eu un saut impressionnant dans la mémorisation de connaissances scientifiques de niche
  tags: 好观点·好思考 · → daily
• <a id="atom-3b9ab44e2aad8624"></a>🟥 2026-06-25 X·@GrantSlatton [老化中] · token定价 · 订单需求
  I would pay truly extortionate prices, like 100x what you pay for e.g. Opus tokens
  tags: 好数字 · → daily
  value: qty=100x
• <a id="atom-15a20bd449f47025"></a>🟥 2026-05-14 X·@jiaxinwen22 · hill-climbing efficiency gap versus Codex · 技术路线/模型发布
  The hill-climbing efficiency gap between Opus and Codex is much larger than I was expecting!
  tags: 好思考 · → daily
• <a id="atom-34247de1385bab11"></a>🟥 2026-06-21 X·@KyleBoas_ · 发布状态 · 产品发布/路线图
  _Fable

Opus

Sonnet_

tags: 好观点 · → daily

• <a id="atom-af89cd41ec8e70ba"></a>🟥 2026-05-31 X·@teortaxesTex · 性能评价 · 模型发布
  OK, this somewhat rescues Opus.
  tags: 好观点 · → daily
  value: direction=positive
  📷 原图
• <a id="atom-c4656c64542e6f89"></a>🟥 2026-05-11 X·@Dorialexander [老化中] · 预训练合成数据影响 · 技术路线/产品发布
  Quasi certain que c'est l'effet du pré-training synthétique maintenant embarqué dans Opus: il y a eu un saut impressionnant dans la mémorisation de connaissances scientifiques de niche.
  tags: 好观点 · → daily
• <a id="atom-6e01dd1def8232ca"></a>🟥 2026-05-04 X·@teortaxesTex · benches performance · 模型发布
  always astonishing on such benches how filthy rich Opus gets
  tags: 好观点 · → daily
  value: qty=filthy rich
• <a id="atom-e2a88eafca57d895"></a>🟥 2026-05-03 X·@secemp9 [老化中] · no hallucination · 性能表现
  no hallucination or anything on opus past 600k tokens
  tags: 好观点 · → daily
  value: qty=600k tokens · direction=past
  📷 原图
• <a id="atom-100ed2cca85d5c6a"></a>🟥 2026-04-15 X·@doodlestein [老化中] · 用户协作评价 · 合作客户
  I spend a ton of time and energy trying to get Opus to try harder and to be more ambitious. And to keep going back over its own work with fresh eyes looking for problems. It makes an incredible team with GPT 5.4, so you get even better results using both together.
  tags: 好观点 · → daily
  value: direction=positive

🟥 空头 takes (bearish) (10)

• <a id="atom-dc199a8e88ded104"></a>🟥 2026-05-11 X·@Dorialexander [老化中] · 模型可能面临自身崩溃问题 · 技术路线/模型发布
  OPD might run into its own collapse issues
  tags: 好思考 · → daily
• <a id="atom-c1f94f581057bdcf"></a>🟥 2026-07-02 X·@ctjlewis · 计费成本想象 · 定价
  Lmao imagine his bill if they didn’t.
  tags: 好观点 · → daily
• <a id="atom-6d066e85b6783d22"></a>🟥 2026-06-10 X·@max_paperclips · ML 能力评价 · 模型发布
  there's been conspiracy theories about Opus being suspiciously bad at ML for a while
  tags: 好观点 · → daily
• <a id="atom-f96f3fd423b48987"></a>🟥 2026-05-31 X·@johnennis [老旧] · 可靠性下降 · 产品发布
  it doesn't work reliably
  tags: 好观点 · → daily
• <a id="atom-4267db734d4aeff5"></a>🟥 2026-05-31 X·@johnennis [老旧] · 不服从操作者 · 产品发布
  It won't obey me
  tags: 好观点 · → daily
• <a id="atom-129dcc29bd14dee9"></a>🟥 2026-05-29 X·@gfodor · 被高估的智能水平（比人们想的笨） · 技术路线/模型发布
  Opus is dumber than people think
  tags: 好观点 · → daily
• <a id="atom-1f423893333b7016"></a>🟥 2026-05-19 X·@xeophon · 生成准确性 · 模型表现
  opus hallucinated a ton of stuff
  tags: 好观点 · → daily
• <a id="atom-ed5a177c5e8bcf1f"></a>🟥 2026-04-21 X·@IlyasHairline · might get available only at API costs · 价格动态/产品发布
  stuff like opus might get available only at API costs
  tags: 好观点 · → daily
• <a id="atom-e8b4cc3d32bfc0cf"></a>🟥 2026-04-16 X·@max_paperclips [老化中] · 变得更容易对事实不严谨 · 模型能力
  Opus has been a little loose with the truth for a while, but this feels....more so
  tags: 好观点 · → daily
• <a id="atom-2be5ed39852f029d"></a>🟥 2026-04-15 X·@BronsonSchoen [老化中] · mitigation takes significant amount of time every day · 产品发布
  mitigating this takes a significant amount of time every day
  tags: 好观点 · → daily

🟥 中性 takes (neutral) (5)

• <a id="atom-00cda4eae9ea44c5"></a>🟥 2026-06-28 X·@manhhung_ai · 测试通过优先级 · 技术路线/模型发布
  my q is which hop you drop first when you're late - grok planning or the opus test pass?
  tags: 好思考 · → daily
• <a id="atom-b95b0d3fc5c1cccd"></a>🟥 2026-07-02 X·@ctjlewis · 定价 · 定价
  And Opus itself is expensive.
  tags: 好观点 · → daily
  value: direction=高
• <a id="atom-d685b5a565188fef"></a>🟥 2026-06-25 X·@chrisbarber · was somewhat better, but not significantly better than GLM 5.2 · 竞争格局
  i asked a few people and all felt that opus was somewhat better, but not significantly better
  tags: 好观点 · → daily
• <a id="atom-3836ee16ea2d260f"></a>🟥 2026-06-02 X·@PawelHuryn [老旧] · 用于复杂推理和汇总 · 技术路线
  Put Opus where the thinking happens.
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-57d6a8f4fc67b901"></a>🟥 2026-05-29 X·@MikushRab [老旧] · 4.8 CAD tasks performance · 模型发布/技术路线
  Opus 4.8 just dropped and I ran it through our CAD tasks. 4.6 → 4.7 → 4.8 side by side. The results are unexpected!
  tags: 好观点 · → daily

🟦 新闻流 (squawk · 1)

⏱ 时间轴 (近 20)

• 🟥 2026-07-02 · position · 定价 · → daily
• 🟦 2026-07-02 · fact · 计费方式 · → daily
• 🟦 2026-07-02 · fact · 付费承担方 · → daily
• 🟥 2026-07-02 · position · 计费成本想象 · → daily
• 🟦 2026-07-01 · fact · 定价模式 · → daily
• 🟦 2026-07-01 · fact · 解題嘗試時間 · → daily
• 🟥 2026-06-28 · narrative · 测试通过优先级 · → daily
• 🟦 2026-06-27 · fact · 生成速度 · → daily
• 🟦 2026-06-27 · fact · benchmark score · → daily
• 🟥 2026-06-25 · narrative · was somewhat better, but not significantly better than GLM 5 · → daily
• 🟦 2026-06-25 · fact · 发布版本 · → daily
• 🟥 2026-06-25 · position · token定价 · → daily
• 🟥 2026-06-21 · narrative · 发布状态 · → daily
• 🟦 2026-06-12 · fact · 被加入InferenceBench排行榜 · → daily
• 🟥 2026-06-10 · narrative · ML 能力评价 · → daily
• 🟦 2026-06-09 · fact · 记忆测试通过 · → daily
• 🟦 2026-06-06 · fact · cost per bug found · → daily
• 🟦 2026-06-04 · fact · 推理努力设置 · → daily
• 🟥 2026-06-02 · position · 用于复杂推理和汇总 · → daily
• 🟥 2026-05-31 · position · 可靠性下降 · → daily

← 实体目录 · 系统日志