Claude Opus 5

• 1. overview (690 chars)

  Claude Opus 5 is the new generation Opus flagship model launched by Anthropic in July 2026. The official core statement is so straightforward that it is almost provocative - "the world's most powerful programming model." It appears at the end of Anthropic’s intensive model offensive this year: Opus 4.8 (May 28), the mythical Fable 5 / Mythos 5 (June 9), and Sonnet 5 (June 30) have been rolled out all the way. Opus 5 has made up the most critical link in the flagship sequence. Its positioning is not to simply stack capabilities, but to bring Fable 5's "mythical" programming and long-term autonomy capabilities back to a more reasonable price and less security degradation. It focuses on million-level context, super agentic coding and better Token economics, directly benchmarking the GPT-5.6 just released by OpenAI.

• 2. Product background (1051 chars)

  Anthropic is the leading AI company currently competing head-on with OpenAI, backed by two major investors, Amazon and Google. In the first half of 2026, it entered the rhythm of "monthly release model". With its enterprise-level reliability and security reputation, the commercialization growth rate once pushed the valuation caliber above OpenAI. The Opus series has always been Anthropic's most powerful one, specializing in complex reasoning, long context, advanced coding and autonomous agent workflow; from Opus 4.6 to 4.8, the main line is "more honesty, fewer missed code defects, dynamic workflow and effort gear control".   Opus 5’s debut was quite dramatic. On July 9, a model codenamed "Honeycomb" briefly appeared in Cursor and was interpreted by the community as an outpost of Opus 5. In the early morning of July 15, the "Claude-Opus-5" entry appeared directly in the model directory of Google Vertex AI and was intercepted by the developer, confirming its existence. Many media position it as a "sub-flagship" between Sonnet 5 and Fable 5 - the performance benchmark even slightly surpasses Opus 4.8 and is close to Fable 5, but only uses about half the cost, accurately filling the largest price-performance gap in the product line.

• 3. Product functions and usage experience (1802 chars)

  Programming is Opus 5’s strongest draw. Anthropic directly calls it "the best programming model in the world", focusing on the reliability of real software engineering: fewer overlooked code defects, stronger multi-step planning, computer use and browser agents, and long-term tasks that can run continuously and autonomously in the background for a long time. As a frame of reference, Fable 5 with the same base soared to 80.3% on SWE-bench Pro, far exceeding GPT-5.5's 58.6%. In a 50 million-line Ruby code library in Stripe, it compressed the full library migration that took the engineering team more than two months to one day. Opus 5 follows the same technical route to make this capability more inclusive and stable.   Context is another big jump. Rumors and leaks consistently point to a million-Token-level standard context that, with adaptive thinking (enabled by default) and multi-level effort levels (including xhigh), can swallow the entire code base, extremely long legal or research documents in a single session. It continues the dynamic workflow and parallel sub-agent capabilities of Opus 4.8, and can break down a vague goal into a complete chain of research, outline writing, coding, running tests, self-proofreading, and error correction. Wharton School professor Ethan Mollick made the widely circulated judgment when testing contemporaneous models: the paradigm of collaboration between people and models is reversing, and users have changed from "wizards" who need to "recite spells" sentence by sentence to "party A" who only sign on the final product - you throw in a design document of more than ten pages and come back a few hours later to receive a high-quality finished product.   Vision and multimodality are also online. The model with the same base, without external scaffolding, can independently deduce and pass "Pokémon" using only original game screenshots, and is also significantly ahead of GPT-5.5 and Gemini 3.1 Pro on the visual document reasoning benchmark. Opus 5 continues its visual input capabilities and is expected to further improve in high resolution and continuous visual reasoning.

• 4. Pricing and business model (646 chars)

  This is Opus 5’s smartest move. At $10 per million inputs and $50 per million outputs, Fable 5 is fully twice as expensive as Opus 4.8 ($5 inputs and $25 outputs), making it expensive enough for only a handful of high-value enterprise customers. Opus 5 is generally expected to be in the Opus price range or slightly higher (industry estimates are about US$5-8 for input and US$25-40 for output), focusing on "half the price, 90% of the experience." It is open to the Claude API and enterprise subscription layer, and works with prompt word caching and batch processing interfaces to further reduce the cost of high-frequency calls. In other words, Anthropic wants to make near-Fable 5 capabilities affordable to more people without having to pay for the most expensive tier.

• 5. User reviews (1023 chars)

  The developer community speaks highly of the overall capabilities of this product line. The phrase "the best programming model in the world" first came from Every CEO Dan Shipper's long post on actual measurements of models of the same generation. Positive feedback focuses on: its ability to handle ultra-long autonomous tasks that span day and night, its ability to generate a complete and deliverable application with just one prompt word, its contextual integration, and its amazing "taste of problem solving."   But the complaints in actual use are also very concentrated, and almost all revolve around the old problems of the Fable 5 generation: expensive, token burning, and slow. Someone posted a $100 bill for using a heavy-duty model to make a web version of "The Sims" project, and someone found in the Max 20x plan that heavy tasks lost about 2% of their credits every minute. Dan Shipper's saying, "Using it for daily writing is like shooting ants with a rocket launcher" is widely circulated. The half-price positioning of Opus 5 is aimed at this pain point - but whether it can truly reduce the token consumption, rather than just halving the unit price, still needs to be verified by large-scale actual testing.

• 6. Industry evaluation (493 chars)

  The media generally read Opus 5 as Anthropic’s head-on riposte to OpenAI GPT-5.6. GPT-5.6 has previously been said to surpass Fable 5 in terms of coding and computing power efficiency. If Opus 5 is really implemented with millions of contexts and better economics, it will bring "the most watched benchmark comparison" back to Anthropic. From the product line, it completes the three echelons of "Sonnet 5 volume, Opus 5 main force, and Fable 5 top configuration", allowing enterprises to fine-tune routing models based on task value, rather than the most expensive tier without brains.

• 7. Controversies and Risks (1054 chars)

  The first is that "published" itself is still vague. As of mid-July, sources such as CometAPI and Crypto Briefing have emphasized that Anthropic has not yet made an official announcement. There is an information gap between the Vertex catalog exposure and some "already online" reports. Some English reports also misplaced the pricing and security details of Fable 5 on Opus 5, and the official system card needs to prevail.   The second is the experience cost of security downgrade. Fable 5's mechanism for automatically routing high-risk requests (network security, biochemistry, model distillation) back to Opus 4.8 has been criticized for being overly defensive - some Chinese users reported that saying "Hello" can trigger high-risk warnings. Whether the safety guardrails of Opus 5 are lighter or continue this set is directly related to the accidental injury rate of normal tasks.   The third is the structural problem of uncontrollable costs. Agentic workflow naturally stretches a task into a series of sub-tasks, and halving the unit price does not mean halving the total cost; Opus 5 may not be automatically immune to the lessons learned in the Fable 5 era when "medium-sized tasks quietly burned 500,000 to 1 million Tokens in the background."

• 8. Applicable groups and usage suggestions (596 chars)

  It is suitable for: architects who need to overcome difficult projects that require the entire team to develop for several months, engineering teams that run long-cycle independent agents, and legal/financial/research scenarios that need to process extremely long documents in a single session. Not suitable for: Only doing daily tasks such as light Q&A, changing titles, and writing short copy that require quick back and forth - those scenarios are where Sonnet 5 is more economical. A pragmatic approach is to build a model routing ladder: use Sonnet 5 for light tasks, use Opus 5 for the main complex tasks, and only use Fable 5 for extreme cutting-edge tasks, with usage monitoring and manual acceptance.

• 9. Summarize (511 chars)

  The value of Claude Opus 5 does not lie in refreshing the ability ceiling again, but in its attempt to turn the "mythical" ability into a profitable business - delivering programming and autonomous capabilities close to Fable 5 at about half the price. If the official benchmarks and token economic performance fulfill the rumors, it is likely to become the new default option for most complex workloads in the second half of 2026; but until the official system card and large-scale actual testing are implemented, the phrase "the world's most powerful programming model" should remain calm for the time being.

No in-depth report yet

• 

  DFoster_2023

  —

  半价对标 Fable 5，这波是真的会算账。

---

• 

  JAlin

  —

  等官方系统卡，现在都是 Vertex 目录截图和小道消息，先别急着封神。

---

• 

  ACruz_2023

  —

  昨晚拿 Opus 5 重构了一个中型 Node 服务，它自己读代码、拆任务、跑测试、连着修了三轮 CI 才收工，我全程就盯着屏幕喝咖啡。说它是世界最强编程模型我信一半，剩下那一半真得等这个月的 token 账单出来再下结论，能力是真强，就是不知道这钱花得值不值。

---

• 

  Jack_ClarkJr

  —

  Sonnet 5 走量，Opus 5 主力，Fable 5 顶配，这个三档梯队终于齐了，选型清晰多了。

---

• 

  Angela.Murphy_X

  —

  百万上下文是真香，整个 repo 塞进去它还能记住细节。

---

• 

  Melissa_Lopez_75

  —

  冲着挑战 GPT-5.6 来的，坐等两家跑分对轰。

---

• 

  流光_8

  —

  就问一句，Opus 5 还会不会像 Fable 5 那样说个「你好」就弹高危警告？

---

• 

  LindaMurphy

  —

  价格减半我承认，但 agentic 那套一个任务天然会拆成一长串子任务，读文件、改代码、跑测试、修故障、再自我验证，token 照样哗哗地烧。单价降了不代表总价降了，Fable 5 时代中型任务后台悄悄吃掉几十万 token 的教训还历历在目，别高兴太早。

---

• 

  MarkHughes

  —

  太强了，回不去了。

---

• 

  Elizabeth.Walker_66

  —

  Honeycomb 那会在 Cursor 里露头我就猜到了，果然是 Opus 5 的前哨。

---

• 

  JoanHernandez

  —

  我们团队之前把整个生产环境的 bug backlog 直接丢给同底座那个模型，然后就下班回家了，第二天早上回来一看，调用栈分析完了、覆盖率跑通了、PR 都提好了，整个缺陷库被扫得干干净净。这种「放手到天亮」的托管体验是真的爽，就是月底看账单的时候有点想哭。

---

• 

  沈玉然

  —

  拿它写日常文案属实是火箭发射器打蚊子，Dan Shipper 那句吐槽我天天在工位复述。

---

• 

  Logan.Diaz_Max

  —

  次旗舰这个定位其实最实用，大多数活儿用不上满血 Fable，Opus 5 刚刚好。

---

• 

  VEtay

  —

  请问 Opus 5 现在国内怎么调用啊，走 API 还是得挂 Vertex？

---

• 

  sadrabbit530

  —

  从 Opus 4.8 迁到 Opus 5，同一套 agent 脚本明显少了很多「被忽略的坑」，多步规划稳了不少，长任务跑到一半崩掉、上下文漂移的次数肉眼可见地降下来了。对我来说这点比那些漂亮的跑分更打动人，毕竟真实工程里，稳定和可复现比峰值分数重要太多了。

---

• 

  Evelyn_Collins_X3

  —

  凌晨刷到 Vertex 目录里蹦出 Claude-Opus-5，直接精神了。

---

• 

  P1JFWFQ0

  —

  绝了，这更新节奏，Anthropic 是真按月发模型啊。

---

• 

  Heather.Lee520761

  —

  重度用户表示：能力我一点不担心，我担心的是 Max 套餐里重任务一分钟掉 2% 额度那档子事会不会在 Opus 5 上原样重演。半价固然好，可要是任务照样膨胀，省下来的钱分分钟又烧回去，希望官方这次真能把 token 效率做扎实，而不只是把标价砍一半糊弄人。

---

• 

  wiipreh

  —

  别被英文媒体带偏，好几篇把 Fable 5 的 $10/$50 和安全路由细节硬安到 Opus 5 头上，信息差有点大。

---

• 

  Kelly_Collins_202264

  —

  等实测，先观望。

---

• 

  redelephant244

  —

  同底座那个模型无脚手架盲打通关宝可梦的 demo 我看了三遍，视觉推理是真的进化了。

No reviews yet