The model circle is very lively in July.
On the early morning of the 25th, Anthropic officially launched its new model Claude Opus 5.
This time, Anthropic did not emphasize the “strongest model”, but instead provided a more pragmatic positioning: Opus 5 has fewer missing steps in complex tasks, is more inclined to actively verify intermediate results, balances performance and cost in multiple evaluations, and is suitable for daily high-frequency use.
The change in positioning is also directly reflected in the product. Currently, Opus 5 has become the default model for Claude Max and the strongest model that can be called by individual subscription users (Max/Pro); The Fable 5, which has higher capabilities, is mainly aimed at API and enterprise customers. Although it is available to individual subscribers, the amount is quite limited.
According to the official results, Opus 5 focuses on improving the efficiency of programming, knowledge work, and intelligent agent task completion, ranking among the top in multiple software engineering, business process, and computer operation evaluations; At the highest thinking level, some grades have approached or even surpassed Fable 5. However, in high-risk scenarios such as offensive network exploitation and long-term autonomous biological research, the restricted Mythos 5 is still stronger.
Opus 5 provides adjustable effort levels, allowing users to adjust their inference resource allocation based on task difficulty. The higher the gear level, the stronger the ability to handle complex problems, but the speed is slower and the token consumption is higher; Lowering it saves time and money.
At present, Opus 5 has been launched on all platforms, with an API price of $5 per million tokens input and $25 per million tokens output, which is on par with Opus 4.8 and about half of Fable 5; To pursue speed, you can activate the Fast mode and exchange it for 2.5 times the speed at approximately twice the price.
Anthropic has launched two beta features simultaneously: dynamic adjustment tool and automatic rollback, which are used to reduce the cost of Agent tasks due to cache failure or the risk of interruption due to security interception.
Can Opus 5 replace Fable 5 with similar performance and halved price?
01. Ranking the top spot, but not leading by a fault
Let’s first take a look at the official transcript provided.
In order to highlight the leading edge of the model, the comparison objects of Opus 5 are the strongest models currently available: the previous generation Opus 4.8, the publicly available flagship Fable 5, and OpenAI’s recently released flagship model GPT-5.6 Sol.
Software engineering is the most prominent aspect. In the Frontier Bench v0.1 test, Opus 5 outperformed all tested models, with a score more than double that of Opus 4.8, and a lower cost per task. In the CursorBench 3.2 test, when max effort was enabled, the difference in performance between the two was less than 0.5%, and the cost of each task was about half of the latter.
Similar advantages also appear in tasks that require continuous operation of the model. In the Zapier Automation Bench test, based on the same single task cost calculation, the pass rate of Opus 5 is approximately 1.5 times that of the second place (Fable 5). In the OSWorld 2.0 test, it only cost slightly more than one-third of the single task cost of Fable 5, surpassing the latter’s best performance.
The commonality of these projects is that they test not only whether the model can answer a question correctly, but also whether it can call tools, cross multiple steps, identify errors, and continue to move forward.
ARC-AGI 3 mainly tests the model’s ability to handle unfamiliar rules and new problems. Opus 5 received 30.2%, previously leading with 7.8% of GPT-5.6 Sol. However, the score is taken from the high level rather than the max level, and the results for the max level have not yet been announced. At least under this setting, Opus 5 demonstrates strong abilities in rule induction, observation, and trial and error.
However, Opus 5 is not leading in all projects. Among the more than ten comparisons published by Anthropic, four lagged behind other models.
In the Humanity’s Last Exam test, when no tools were used, Opus 5 scored 56.3%, slightly lower than Fable 5’s 56.5%; After opening the tool, Opus 5 rose to 64.7%, surpassing Fable 5’s 63.9%. This further illustrates the difference between the two: Fable 5 still has stronger pure modeling capabilities, while Opus 5 is better at searching, calling tools, checking results, and continuously advancing tasks.
The case disclosed by Anthropic also conforms to this characteristic. Opus 5 independently developed a computer vision program to extract data from the raw pixels of mechanical drawings without image viewing tools; In the absence of real-time market trends, proactively set up a testing environment to verify whether the exchange data interface is correct.
From these performances, it can be seen that the focus of the upgrade of Opus 5 is not so much on the leap of pure model capabilities, but rather on the maturity of execution, verification, and error correction capabilities.
Of course, official scoring needs to be approached with caution. Taking Frontier Bench testing as an example, the result is to take the average of five tests conducted under a specific intelligent agent framework. When triggering the security classifier, Opus 4.8 will also take over, indicating that changes in tool configuration, prompt words, and operating environment may also affect the results.
Let’s take a look at the third-party rankings again.
As of July 25th, on the Artificial Analysis Intelligence Index list, Opus 5 max ranked first with a score of 61, only one point higher than Fable 5 in second place. This list combines 9 evaluations, which better reflect comprehensive abilities than looking at a single advantageous item.
Two conclusions can be drawn here. Firstly, Opus 5 has entered the current top tier and is temporarily ranked first in the highest thinking level. Secondly, in various tests, Opus 5 did not show significant differences from other models and did not form a fault advantage.
So, Opus 5 is just the leader on the current chart. The benchmark provides the upper limit of ability, and the stability, token consumption, and final delivery quality in real use determine whether the cost-effectiveness is high or not.
02. Higher running scores do not necessarily mean better work
After the model was launched, evaluations quickly polarized.
Positive reviews focus on programming and interactive content generation. A netizen shared on X the interactive 3D black hole visualizer generated by Opus 5, which can also synthesize an electronic music in real-time. The code, visual effects, and sound are integrated into the same product, indicating its strong ability to quickly produce product prototypes.
Some netizens have also compared Opus 5 with Fable 5, GPT-5.6, and Kimi K3, believing that it performs better in 3D production.
But some developers also believe that Opus 5 overthinks in some tasks, has a long reasoning process, and may even deviate from its initial goal by going its own way later on. After testing, some netizens claimed that the model consumed a large amount of tokens and occupied a considerable amount of contextual space, but did not strictly follow the requirements. Higher effort does not necessarily lead to better results.
More systematic testing comes from the code review platform CodeRabbit. Its conclusion is that Opus 5 is more suitable for writing code rather than independently undertaking all code review work. It performs better when dealing with open, design oriented tasks, but may still miss out on issues such as logical errors, race conditions, and API misuse.
Independent developer Li Mo described Opus 5 as a ‘reliable colleague with a slight personality’. In actual experience, its debugging ability is very strong, overall much higher than Opus 4.8, and it is also a suitable substitute for Fable 5 after the quota is exhausted. However, the workflow and prompt words formed in the Opus 4 era cannot be directly transferred to Opus 5. To achieve significant improvements, it is necessary to understand the new model habits, adjust task splitting, effort levels, and validation methods.
Li Mo mentioned that two practical Beta features have directly improved the Agent development experience: in the past, changing the entire cache of the tool list once would render it invalid, but now changing tools during conversations will not render it invalid; The API can also enable automatic degradation, and requests intercepted by the security classifier will be automatically forwarded to other models.
Overall, Opus 5 is currently more like a colleague with strong execution ability and willingness to self check, but still needs to adapt. It is suitable for generating code, fixing problems, and completing tasks with clear boundaries, but Fable 5 still has advantages in complex planning, long-term autonomous tasks, and some textual expressions.
03. Anthopic needs a more everyday flagship model
Anthropic has been accelerating its model update speed recently.
On May 29th, Opus 4.8 was released; On June 10th, Fable 5 and Mythos 5 were released; On July 1st, Sonnet 5 was launched; On July 25th, Opus 5 was released, and in less than two months, Claude’s main product line was almost updated.
In the context of intensive release of models, the industry’s divergence on how models should be opened up is also expanding. On July 16th, the dark side of the moon released the open weight model Kimi K3, which is considered to be approaching the forefront of performance. A few days later, OpenAI’s strategic leader predicted on X that Washington would create “regulatory risks” around China’s open weight model, which was widely interpreted as pressure on the open source camp.
On the day of the release of Opus 5, Nvidia Meta、 Microsoft and 25 other companies and institutions jointly issued an open letter calling on the US government not to prematurely restrict open weight models. OpenAI, Anthropic, and Google were not included in the signing list.
Being absent does not necessarily mean explicitly opposing open weighting, but compared to open weighting, Anthropic does emphasize closed source, hierarchical openness, and risk control more. On June 12th, the US government implemented regulations on Fable 5 and Mythos 5, and both models were temporarily taken offline. After the restrictions were lifted, Fable 5 resumed public service, while Mythos 5, which has fewer security restrictions, is still only open to some approved institutions.
The positioning of Opus 5 is embedded in Anthropic’s strategy of “how to open up models”. Its ability to discover source code vulnerabilities is close to Mythos 5, but its ability to transform vulnerabilities into actual attack tools is significantly weaker. For Anthropic, it is not just a cheaper alternative to Fable 5, but also an option that is easier to open to ordinary users and businesses within a compliance framework.
At the same time, Anthropic is still expanding its computing power. Two days before the release of Opus 5, the company announced the deployment of AMD Instinct MI450 series GPUs and Helios rack level systems with a maximum scale of 2 gigawatts, with the first 1 gigawatt deployment scheduled to begin in the first half of 2027. AMD also pledged to make a strategic equity investment of up to $5 billion in Anthropic in the future.
From intensive release of models, to expanding chip supply, to streamlining system prompt words, Anthropic not only improves benchmark scores, but also breaks down model capabilities into different price, risk, and usage scenarios: Fable 5 is responsible for more complex and longer cycle tasks, Sonnet 5 undertakes cost sensitive universal needs, and Opus 5 attempts to become the daily default choice for developers and knowledge workers.
So, Opus 5 did not make Fable 5 meaningless. The latter still represents the upper limit of Anthropic’s publicly available model capabilities, while Opus 5 addresses how to deliver near flagship capabilities to more users at a lower cost. Its competitiveness lies in whether it can truly complete a task with more stable performance, less manual takeover, and controllable costs.
相关文章
