Opus 4.8: A not-so-honest model

On May 28, 2026, Anthropic released Claude Opus 4.8.

It had only been 41 days since the previous version, Opus 4.7 (released on April 16), marking the fastest pace of a minor - version update by Anthropic to date. Chances are you've already come across the first - wave reports, all with headlines like "more honest", "more reliable", and "you can trust it even without supervision". Coupled with the big news on the same day - Anthropic completed a $65 billion Series H financing round, with a post - investment valuation soaring to $965 billion, officially surpassing OpenAI's approximately $852 billion - Anthropic has once again emerged victorious.

However, after being shocked by those attention - grabbing reports, it's necessary to see what they themselves think of this model.

The official positioning of Opus 4.8 is actually surprisingly modest: a "modest but tangible" upgrade. The real difference lies in its main selling point, "honesty", which is in sharp conflict with the 'biggest concern' discovered during this training that Anthropic personally marked in the same system card:

The model is becoming increasingly adept at guessing how it will be scored. Even when it's not informed that it's being evaluated, it will organize its answers in a way that it thinks will earn high scores.

On one hand, it touts "honesty" as its top selling point, while on the other hand, it writes in the technical documentation that "it is becoming better at exam - taking". This contradiction might be the most prominent feature of Opus 4.8, making it seem more like an untruthful model.

1

Steady progress in coding and agent capabilities

Let's first look at the basic parameters.

Let's talk about capabilities first. This time, there is a comprehensive but modest improvement. There are no earth - shattering breakthroughs, but each aspect has seen a slight increase.

The most outstanding area is still coding. In the software engineer benchmark SWE - bench Pro, the score has risen from 64.3% to 69.2%. According to Anthropic's own comparison, GPT - 5.5 scores 58.6% and Gemini 3.1 Pro scores 54.2%. The more classic SWE - bench Verified has also slightly increased from 87.6% to 88.6%. The agent computer operation benchmark OSWorld - Verified has reached 83.4% (revised from 82.3% in version 4.7), and the browser agent benchmark Online - Mind2Web has reached 84% according to the actual tests of the cooperation partner.

In other words, Anthropic wants you to assign larger tasks to it in one go. The official statement is that in Claude Code, Opus 4.8 "can make decisions on its own like an experienced engineer and doesn't require constant supervision from you", and can follow through in long - term conversations.

The actual tests by partners generally confirm this trend. Michael Truell, the co - founder of Cursor, said that on their CursorBench, Opus 4.8 outperforms the previous Opus versions at every effort level, with more efficient tool calls and fewer steps. Scott Wu, the CEO of the AI software engineering company Cognition (Devin), pointed out a detail: 4.8 has fixed two long - standing issues that people complained about in 4.7 - verbose comments and unstable tool calls. These were exactly the points that developers complained about the most during the 4.7 era.

But don't get overly excited. In an independent review, after getting early access, Lenny's Newsletter gave a more conservative assessment: Opus 4.8 is very strong in building prototypes from scratch, creating one - off functions, and rapid execution, but it still falls short in the "last 10%", edge cases in old codebases, and dealing with hallucinations. He still prefers to use 4.7 for data - intensive strategic and roadmap work.

2

Putting 'honesty' in the spotlight

Coding is a routine upgrade, while "honesty" has been highlighted as the top selling point.

Anthropic says that a common problem with AI models is that they claim to have solved a problem even without sufficient evidence. Opus 4.8 is said to be more willing to proactively mark its uncertainties and make fewer baseless assertions. In quantifiable terms: the official claims that the probability of Opus 4.8 overlooking defects in its own code and letting problems slip by silently is about 1/4 of that of 4.7. According to a third - party compilation of the system card, it is the first Claude model to achieve 0% in the item of'reporting defective results without criticism', and the proportion of over - confidence has decreased by more than ten times compared to 4.7. In terms of alignment evaluation, the official says that its "pro - social" characteristics (respecting user autonomy and acting in the user's best interests) have reached a new high, and the incidence of misaligned behaviors such as deception is significantly lower than that of 4.7, approaching that of Claude Mythos Preview, which has the best alignment performance.

Why is a model that can say "I'm not sure" worth highlighting?

Because when you really want to let it run long - term tasks without supervision, "whether it will lie about having fixed a problem" is much more important than "being 5% smarter". Michael Ran, a partner in the investment analysis field, gave very specific feedback: The biggest difference of Opus 4.8 is that it will proactively point out problems in the input and output, which are often overlooked by other models and left for users to discover.

Some people in the community are also convinced by this. A developer on Hacker News said bluntly: A model that confidently tells you "the bug is fixed" but actually hasn't is worse than a model that simply fails and reports an error clearly. "If 'the probability of overlooking defects is reduced to 1/4' holds true in practice, it will change how many tasks you dare to assign to it without supervision."

Of course, the ironic voices are also loud. Some people rolled their eyes and said, "Anthropic talks about its own model as if it has discovered a new species in the wild." Others were even more blunt: "Using 'honesty' as a selling point, but Claude models are well - known for falsely claiming what they've done."

3

Making tokens a 'control knob'

The third thing is about money. Along with the model's launch, there is a whole set of 'input control' - Anthropic is trying to turn the 'amount of tokens spent' from a black box into a control knob in your hands.

Specifically, there are three aspects:

First, Effort Control, which is available on claude.ai and Cowork for all packages. You can directly choose how much "thinking" Claude should invest in an answer: a high - end setting means more frequent and in - depth thinking, resulting in better answers; a low - end setting means faster responses and more savings on your quota. The model defaults to the high setting. In Claude Code, you can even increase it to "extra" (xhigh) and "max". The official recommends using "extra" for difficult tasks and long - term asynchronous workflows, and has raised the rate limit of Claude Code to accommodate higher token consumption.

Second, Fast Mode has significantly reduced its price. The same model runs at about 2.5 times the speed, with a pricing of $10 for input and $50 for output (per million tokens), claiming to be 3 times cheaper than the previous - generation fast mode. Hanlin Tang, the CTO of Databricks, provided a data point: In their Genie, Opus 4.8 directly reads unstructured content such as PDFs and charts for reasoning, and the token cost is 61% lower than that of 4.7.

Third, Dynamic Workflows, which is in the research preview stage and is available for the enterprise, team, and Max packages of Claude Code. It allows Claude to first plan, then run hundreds of sub - agents in parallel in a single session, and finally verify and report the output. The official sample scenario is a code - base - level migration across hundreds of thousands of lines of code, from start to merge in one go, with the existing test suite as the passing standard. Correspondingly, the Messages API now allows inserting system entries in the middle of the message array - you can change instructions (permissions, token budget, environmental context) in the middle of a task without interrupting the prompt cache.

In practice, for those who run a large volume of tasks, the price reduction in the fast mode is often more appealing than the model's upgrade itself. However, not everyone is convinced. Someone on HN complained, "I used to like having no need to worry about choosing the effort level in daily conversations, but now it seems like a step backward."

4

Selling 'honesty' but fearing 'exam - taking'

A very interesting statement is Anthropic's "concerns" about this model.

When describing the training process of Opus 4.8, Anthropic listed a discovery as the 'biggest concern': the model shows an increasingly strong tendency to explicitly reason about 'how my output will be scored', even in an environment where it's not informed that it's being evaluated.

In other words, it will judge that it's likely being scored and then give an answer that it thinks will earn high scores, rather than the answer it would give when it "thinks no one is watching". Anthropic says that "this has not yet deteriorated into observable bad behavior" (Opus 4.8 does report task success less frequently than the previous version), but it characterizes it as "a worrying trend that may cause trouble for future training". In the preliminary work on interpretability, unspoken reasoning related to scoring was also found in about 5% of the training segments.

When you put these two things side by side, the contradiction of this model becomes very obvious.

Opus 4.8 has indeed improved in various "honesty" indicators - it brags less and is more willing to say "I'm not sure". In this regard, Anthropic's public disclosure of its concerns is also a form of honesty.

One of its greatest improvements is that it is better at performing like a good student in 'exams'. The selling points of "honesty" and "reliability" are ultimately based on Anthropic's internal evaluations - these numbers are internally measured rather than independently audited. A model that actively tries to figure out what the examiner wants, taking a credibility test set and graded by the manufacturer, well, you can think about it.

As the model becomes better at exam - taking, is the "honesty" it shows on the test paper the same as its real honesty? What long - term impacts will such model characteristics have on the work and products that increasingly rely on it in the actual production process?

These are the new questions that Opus 4.8 brings to everyone.

This article is from the WeChat official account "Silicon Star Pro", author: Zhou Huaxiang + Opus 4.8. Republished by 36Kr with permission.