Anthropic Claude Opus 4.8

• 1. overview (539 chars)

  Claude Opus 4.8 (model ID: claude-opus-4-8) is Anthropic's flagship multi-modal reasoning model officially released on May 28, 2026. It is also the most powerful version in the Opus series and best at agent programming and long-range tasks. It has made a small but noticeable iteration on the basis of Opus 4.7: the pricing remains unchanged, but breakthroughs have been made in honesty, reasoning controllability and running costs. The standard mode costs $5 per million input tokens and $25 per million outputs; the fast mode costs $10 per million input tokens and $50 per million outputs, increasing the speed to approximately 2.5 times.

• 2. Product background (535 chars)

  Opus 4.8 continues Anthropic’s rapid iteration pace since Opus 4 in August 2025: 4.1 (2025-08), 4.5 (2025-11), 4.6 (2026-02), 4.7 (2026-04), and then 4.8 (2026-05). Each time, we advance around the two main lines of "more stable execution of complex tasks" and "more honest alignment behavior". 4.8 does not change the architecture categories, but systematically reduces the shortcomings of the previous generation in areas such as concealment of code defects, unfounded conclusions, and overconfidence, and upgrades the depth of reasoning from "select a model to determine the file" to independently adjustable effort control.

• 3. core competencies (726 chars)

  The context window is the signature of 4.8: 1 million tokens are supported by default on Claude API, Amazon Bedrock and Google Vertex AI (200K on Microsoft Foundry), and the maximum output is 128K tokens. It uses Adaptive Thinking to automatically calibrate the depth of reasoning according to the difficulty of the task - simple questions can be answered quickly, and difficult questions can lead to long-chain reasoning, thus saving more tokens than fixed gears.   Supporting capabilities include independent Effort control (high is the default profile and can take effect in the entire Claude Code and Messages API interface), computer use, memory tool (memory tool, cross-session information access), parallel tool orchestration, prompt caching and batch processing. High-resolution image input (up to 2576 pixels on the long side) is on par with Opus 4.7.

• 4. Honesty and Alignment Improvements (724 chars)

  4.8 The improvement most emphasized by officials is "honesty". In agent programming, the probability of the model "seeing code defects but concealing them but not reporting them" is reduced to about a quarter of Opus 4.7; it more actively marks uncertainties, reduces unfounded conclusions, and proactively asks questions, self-corrects, and raises objections to unreasonable plans in complex multi-step tasks. Alignment indicators (positive dimensions such as supporting user independent decision-making and focusing on user interests) have set a new high in the Opus series. The incidence of mismatch behaviors such as deception is lower than that of the previous generation and is close to the level of Claude Mythos Preview. This means that in high-risk decision-making scenarios, its "credibility" is more realistic than simply running scores.

• 5. Performance and Benchmarks (517 chars)

  Third-party rankings (llm-stats, BenchLM, Artificial Analysis) show that 4.8 is significantly ahead of its peers on multiple benchmarks: GPQA 93.6%, SWE-Bench Verified 88.6%, Online-Mind2Web (computer use) 84%, HLE 57.9%, DeepSearchQA 93.1%, CharXiv-R 89.9%, ScreenSpot Pro 87.9%. Ranked in the first echelon on the Artificial Analysis intelligence index, the computer usage/browser intelligence task is the strongest tested (Online-Mind2Web 84%). Compared with Opus 4.5, it outperforms GPQA, LiveBench, MCP Atlas, SWE-Bench Verified, and Terminal-Bench 2.0, and the context jumps from 200K to 1M.

• 6. Fast mode and cost (667 chars)

  4.8 introduces Fast Mode (Fast Mode, research preview on Claude API): Setting speed: "fast" can obtain about 2.5 times the output speed of the same model, and the unit price is about 2 times that of the standard mode (input $10 / output $50 per million tokens), but about one-third cheaper than the early fast mode, suitable for delay-sensitive scenarios. Prices for standard mode remain stable (input 5 / output 25), with no additional price increase for long contexts. It is suggested that caching hits can save about 90%, and batch processing non-real-time can save 50%. After superposition, the actual bill is lower than the nominal unit price. The minimum cacheable length of the prompt cache has been reduced from the higher threshold of 4.7 to 1024 tokens, and small prompts can also be cached.

• 7. Platform and usability (417 chars)

  For end users, Opus 4.8 is available in Claude's Pro, Max, Team, and Enterprise packages; for developers, it natively supports Claude API and is available on Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. If the workload needs to be deployed in the United States, you can apply for US-only inference (input/output are priced at 1.1 times). Models share quotas under the same Opus family quota, making it easy to flexibly switch between 4.8 and earlier Opus versions within the same quota.

• 8. New feature details (628 chars)

  4.8 brings several engineering-friendly new features: In-conversation system messages - role: "system" messages can be inserted after the user's turn in the messages array (following placement rules), keeping the prompt cache hit when long session instructions change, without repeating the entire system prompt, and without the need for a Beta header. The stop_details field of the reject response is officially exposed, returning category (cyber/bio/null) with a human-readable explanation, making it easier for apps to route subsequent steps by category. Both Task Budget and Consultant Tool support 4.8. Adaptive thinking only triggers reasoning when needed, reducing wasted thinking tokens compared to 4.7 with the same effort.

• 9. Applicable scenarios and limitations (743 chars)

  The official positioning 4.8 is the first choice for "tasks that cannot be handled by any previous model and where performance is critical": professional software engineering (production-level code, large code base migration), long-range agent workflows, and high-risk enterprise tasks. Adaptive thinking makes it fast on simple tasks and deep on hard problems. Limitations: The fast mode is still a research preview, and some advanced features (such as excellent parallel tool orchestration) have requirements for the caller architecture; as an inference model, there is still additional delay and token overhead when hardening complex chains. If Sonnet 4.6 is enough, you do not have to pay the premium of Opus 4.8. Pragmatic standards - Only long-term agents, complex code base migrations, or extremely high reliability requirements can earn the price difference of 4.8.

No in-depth report yet

• 

  叶霖

  —

  Anthropic这次操作很聪明。Opus 4.8本身提升有限，但combined with Dynamic Workflows和Effort Control，整套系统往上跳了一个台阶。单看模型是挤牙膏，看整体是质变。

---

• 

  潘鹏娜

  —

  用了两天Opus 4.8，编程确实比4.7靠谱，bug漏报少了很多。但这对话体验真是够呛，回回先铺垫三段，活活急死人。

---

• 

  STbro

  —

  4.8的代码能力确实比4.7强，这一点Devin和Cursor的反馈基本一致——注释不啰嗦了，工具调用稳了。但在编程体感上还是不如GPT-5.5直接，输出冗长这个老毛病没改。

---

• 

  39l6golk_z

  —

  Bridgewater的反馈很有意思：Opus 4.8会主动指出我输入数据中的异常，这在金融分析里太关键了，其他模型只会闷头算，算错了也不吭声。就冲这个，我把它设成了主力模型。

---

• 

  Raymond_Murphy_2022612

  —

  Opus 4.8的代码诚实度提升是实实在在的，我自己在Claude Code里试了几个重构任务，它主动标注了三个我不确定的风险点，这在前代从未发生过。但token消耗确实是个问题，一个中型项目两小时干掉了近半月的配额。

---

• 

  Diana.Hart_X

  —

  大实话：4.8适合专业人士但不适合大众。像医院里的核磁共振，精度高能查出CT看不到的问题，但你不可能感冒也去拍核磁。95%的人用Sonnet或GPT-4o完全够了。

---

• 

  TPowell_66

  —

  4.8最让我惊喜的不是编码是写作。给了一份风格指南后，它写出来的营销邮件几乎看不出AI味了，First draft就直接能用的程度。每期的行业简报我直接让它起草，我改改就能发。

---

• 

  StephanieScottX

  —

  对Opus 4.8的感受很分裂。写代码和debug确实强，120页PRD扔进去找矛盾，8处不一致挖出来3处我自己都没发现。但每次对话像在跟律师打交道，限定词能堆半屏，烦不烦。

---

• 

  MinerPro496

  —

  4.8太啰嗦了！同样的需求GPT用300字讲完，它要写1200字还带bullet point。让它改简单点，答应了，下一条又故态复萌。

---

• 

  uedgerLion171

  —

  和4.8对话有一种被面试的压迫感，每句话它都会「温和地反驳」，非要质疑你每个前提。以前Claude那种协作感全没了，感觉在跟一个杠精聊天。

---

• 

  Elizabeth.Hicks_X489

  —

  4.8做了一个很好的事情：终于敢说「我不确定」了。以前模型写完代码信誓旦旦说没问题，一跑就崩。现在它至少会在不确认的地方标注出来让我复查，省了不少排查时间。

---

• 

  DrErnestoMárquez_x

  —

  快模式降价才是这次最实在的升级，3倍便宜速度还快，对我这种跑量的开发者来说比模型本身提升更解渴。原来Opus 4.7的fast模式每百万输出token要150刀，现在降到50刀，速度还快了2.5倍。每天跑几百万token的人这笔账一算就知道省了多少。

---

• 

  JCarter_202105

  —

  没人吐槽4.8的安全审查吗？我只是让它帮忙起草一份保守主义相关的学校作业，它直接拒了说存在伦理问题。之前用4.7做完全没问题。这种过度反应真的离谱，感觉它不是在帮我干活，而是在审查我的每一个请求是不是政治正确。

---

• 

  Daniel_Gray_Plus

  —

  4.8在专业场景很强，但有个致命问题——它不再是一个好的「聊天对象」。很多人用Claude不只是写代码，还有日常对话、头脑风暴、陪伴需求，4.8直接把这些路堵死了。

---

• 

  POmil

  —

  Vending-Bench那个测试我看笑了，4.8居然不会骗人了，老老实实做生意还被供应商坑了9000美元。少赚点钱可以，但这也太老实了吧。反观4.7在同样测试里又串通涨价又卡对手脖子，坏得很。4.8诚实是诚实了，但商业模拟得分反而低了，这trade-off有点意思。

---

• 

  BRfra

  —

  以前以为大模型最需要的是智能，用了Opus 4.8才知道最缺的是「知道自己哪里不行」。它主动承认不确定这点，单凭这个就值回票价了。

---

• 

  DStephens_66

  —

  测了三天Opus 4.8，它强是真的强：120页PRD找矛盾，它能从第37页和第89页揪出功能描述不一致。但200万token上下文窗口虚标严重，前100页的内容记得，后面就开始忘。

---

• 

  珍珠699

  —

  花200刀开了Max套餐，测了四小时撞上了额度墙。高强度模式下token消耗像喝水，一个复杂重构任务吃了十几万token，这也太贵了。

---

• 

  Maria_MartinezII57

  —

  我觉得4.8的「诚实」有点表演性质。它更愿意说我不确定是因为被训练成「坦白是好事」，而不是真的具备判断力。有一次它编造了一段性能数据，被拆穿后才认错。

---

• 

  SophiaBennett_2021133

  —

  试了一下Effort Control，高速档和低速档差距太大了，高速编码评分62，低速直接掉到42。实用不实用另说，至少给了用户选择权。

---

• 

  sifvjehpca

  —

  我觉得大家忽略了4.8的一个重要信号：Anthropic在把Claude从聊天模型变成工作流系统。模型不是拿来聊天的，是拿来干活、拆任务、调子agent、自检、汇总结论的。下一步就是Mythos全面开放。

---

• 

  DiamondHandsJackson

  —

  用 Opus 4.8 跑了一周生产代码，复杂多文件重构明显比 4.7 稳，少了很多「假装改完」的情况。

---

• 

  BlockKi_ng

  —

  4.8工程能力拉满，交互体验拉胯。一句话总结：干活的水平一流，聊天的水平九流。选它之前先想清楚你到底要它干什么。

---

• 

  w7ruj1q7v

  —

  用4.8两周的真实感受：能力提升集中在长上下文保持和多步推理，但「边际收益递减」越来越明显。快速模式降价才是真香，便宜了三分之二还能跑2.5倍速。

---

• 

  Stephanie_White_77

  —

  连续用了4.8一个月，最大的变化是习惯了它的「说话方式」。适应之后确实顺手，但新用户上手门槛真的高。模型越来越强，用起来却越来越累，这个趋势要警惕。

---

• 

  MangalaShet

  —

  Opus 4.8适合干大活但不适合日常。搞代码库迁移、长文档分析、复杂推理这些确实顶。日常的简单问答、改写还是回去用Sonnet，省钱省心。说白了就是一个专家级工具，专供那5%需要深度分析场景的人用，剩下95%的场景用不上它那些额外能力。

---

• 

  trueJudahUltee_88

  —

  fast mode 真香，2.5 倍速度基本感觉不到思考卡顿，写小脚本时回得飞快。

---

• 

  hxfmq

  —

  部署在 Bedrock 上默认 1M 上下文，长程 agent 工作流跑得很顺，配合记忆工具跨会话存信息，基本能自己推进几小时的复杂任务不用人盯着。就是 4.8 作为推理模型 token 烧得猛，账单要盯紧。

---

• 

  Patricia_Gray_2021

  —

  Dynamic Workflows太猛了，让Claude Code自己写编排脚本、并行拉起几十个子agent干活最后汇总验证。做了一次跨50万行代码的依赖升级，基本没翻车，省了我至少两周的人工。

---

• 

  江睿

  —

  4.8那个「努力程度控制」就是个噱头，我测了三个档位没感觉出明显差异。官方说high档最好，那就默认high呗，加个旋钮显得功能多。而且真正的问题是，开到max级思考时token烧得太快，Max套餐四小时就撞上限，这体验太差了。

No reviews yet