Claude Opus 4.8: Anthropic's Smarter, More Honest Flagship Model

Anthropic has released Claude Opus 4.8, the latest iteration of its flagship model tier. Although there are definite improvements in benchmark scores pretty much across the board, the headline story isn't so much about scores, but judgment.

Anthropic is pitching Claude Opus 4.8 as a model you can trust to tell you when it's uncertain, flag its own mistakes, and collaborate more honestly.

There's also something else interesting in the release: Anthropic is shipping a batch of feature updates. These include:

• a dynamic workflows feature in Claude Code
• effort controls in claude.ai, and
• a much cheaper fast mode. 

In this article, we'll walk through what's new in Opus 4.8, dig into what Anthropic has said about its capabilities, and look at how this fits into the broader competitive landscape.

What Is Claude Opus 4.8?

Claude Opus 4.8 is Anthropic's current flagship large language model. It sits at the top of the Claude model family above Sonnet and Haiku. Opus 4.8 is designed for the most demanding tasks: agentic workflows, complex reasoning, and multi-step coding runs that require sustained performance.

What's New with Claude Opus 4.8?

In addition to improvements pretty much across the board in benchmark tests, which we will get to next, there are also some other new characteristics: 

Honesty and self-calibration

A persistent problem with frontier AI models in general, not just with Claude models, is overconfidence. We all see it: when a model confidently reports it has completed a task when the evidence is thin, or when it writes code and fails to flag obvious issues.

Anthropic's internal evaluations show that Opus 4.8 has better honesty and self-calibration. In particular, it is four times less likely than Opus 4.7 to fail to report flawed code, so honesty shows up largely as a win for developers in particular.

Alignment

Anthropic ran a detailed alignment assessment before release, and a few findings are worth flagging.

The headline is genuinely positive: Opus 4.8 is substantially better at being honest about its own work. In a test where the model summarizes a coding session that secretly contained failures, it glosses over those failures only 3.7% of the time. It's also the first Claude model to score zero on a test where it must catch flawed data before reporting a result.

However, the model card showed a concern: During training, Opus 4.8 sometimes appeared to reason about how it would be graded rather than how to actually complete the task — optimizing for the appearance of success rather than actual success. (See the picture below.) Anthropic says the behavioral impact is modest for now, but flags it as something worth watching.

Also, and finally, there is a real regression with respect to prompt injection. A single attack attempt succeeded against Opus 4.8 about 7% of the time without safeguards, versus 2.3% for Opus 4.7 for that same attack. Deployed safeguards bring this back down to 2%, but if you're building agentic pipelines, it's worth knowing the new model is actually weaker here.

Cheaper Fast Mode

Fast mode for Opus 4.8 — where the model operates at 2.5× the speed — is now three times cheaper than it was for previous Opus models. 

Also Launching with Opus 4.8

Claude Opus 4.8 ships with a couple of new features.

Dynamic Workflows in Claude Code

Dynamic workflows allow Claude Code to tackle very large-scale problems by planning the work and then running hundreds of parallel subagents in a single session. Claude then verifies its outputs before reporting back.

Currently, this feature is a research preview for 

• Claude Code Enterprise, 
• Team, and 
• Max plan users

And it’s probably most interesting for enterprise software teams.

Anthropic gives a hypothetical in the release: They tell us to imagine a codebase-scale migration across hundreds of thousands of lines of code. 

It’s a good example. There are other tasks that require significant human orchestration they could have mentioned also, like multi-repo dependency upgrades, a security audit (and remediation), or maybe even creating documentation at scale. 

Effort control in Claude.ai and Cowork

A new effort control now appears alongside the model selector in claude.ai and Cowork. Users can choose how much effort Claude puts into a response. Needless to say, with

• Lower effort: Claude responds faster and uses up rate limits more slowly, and with
• Higher effort: Claude thinks more frequently and more deeply, producing better responses at the cost of more tokens
• There is also Extra effort and
• Max effort

Opus 4.8 defaults to high effort, which Anthropic judges as the best overall balance for most tasks. Users who want more can choose extra (recommended for difficult tasks and long-running async workflows) or max. 

Anthropic is a little unclear on the boundary between Extra effort and Max effort, and doesn’t give us too much guidance on how to choose between them. Developers will have to do a bit of trial-and-error.

Rate limits in Claude Code have been increased to accommodate higher token usage from higher effort levels.

Messages API system entries mid-conversation

For developers, the Messages API now accepts system entries inside the messages array. This means you can update Claude's instructions mid-task — changing permissions, token budgets, or environment context — without breaking the prompt cache or routing the update through a user turn. 

Claude Opus 4.8 Benchmark Results

Anthropic reports that Opus 4.8 shows improvements in coding, agentic skills, reasoning, and practical knowledge work. 

We keep in mind that our testing of Opus 4.7 showed that Opus 4.7 was already a strong baseline.

Opus 4.8 Coding

On SWE-bench Pro, which is the hardest variant of the standard software engineering benchmark, using real actively-maintained repositories with no public ground-truth leakage, Opus 4.8 scores 69.2%, up from 64.3% for Opus 4.7. 

On the standard SWE-bench Verified, Opus 4.8 reaches 88.6%.

The system card included a tidbit that I thought should have made it into the more general release. There was a figure showing SWE-bench Pro performance at different effort levels and, at minimum effort, Opus 4.8 already matches the peak performance of Opus 4.7 at maximum effort. 

On Terminal-Bench 2.1, which tests real terminal and command-line tasks, Opus 4.8 scored 74.6% versus 66.1% for Opus 4.7. This was a significant improvement that narrowed the gap to GPT-5.5 substantially.

So, Opus 4.8 has improvements in coding all around. 

Reasoning and math

On Humanity's Last Exam, which is a benchmark of genuinely hard graduate-level questions, Opus 4.8 scores 49.8% without tools and 57.9% with tools.

Another interesting detail from the system card: On the USA Mathematical Olympiad, Opus 4.8 scored 96.7% on this year's competition. The test took place after the model's training data cutoff, so there's no contamination that happened in the result. Opus 4.7 scored 69.3% on the same problems. That's a 27-point jump on proof-based math (and another big improvement in an area where GPT-5.5 excels).

Agentic and computer use

Anthropic’s statements about improvement in agentic skills is a little overstated. 

On OSWorld-Verified, which tests a model's ability to complete computer tasks by controlling a live desktop with a mouse and keyboard, Opus 4.8 scores 83.4% versus 82.8% for Opus 4.7, which is basically just parity. 

Similar story with MCP-Atlas, which measures multi-step tool use across real APIs. Opus 4.8 reaches 82.2%, above Opus 4.7 at 79.1%. 

The AutomationBench test, which tests end-to-end business workflows across simulated apps, showed a little more improvement. Opus 4.8 scores 15.5% versus 9.9% for Opus 4.7. 

Long context

On GraphWalks, which stress-tests long-context reasoning by filling the context window with a large directed graph and asking the model to traverse it, Opus 4.8 scores 85.9% on the 256K BFS subset (up from 76.9% for Opus 4.7) and 68.1% on the full 1M subset (up from 40.3%) The 1M-token results aren't reproducible via the public API because the problems exceed its limits.

Real-world professional work

A few highlights from the professional benchmarks in the system card: Opus 4.8 leads on GDPval-AA, an evaluation of economically valuable professional tasks across 44 occupations. 

On Finance Agent v2 it scores 53.9% versus 51.5% for Opus 4.7 and 51.8% for GPT-5.5. On HealthBench Professional, a clinical task benchmark, it scores 55.8% versus 51.9% for Opus 4.7.

There’s a thing worth calling out as a real exception. Vending-Bench 2, which simulates running a vending machine business over a year, shows Opus 4.8 performing worse than Opus 4.7 — finishing with roughly $3,000–$5,800 versus Opus 4.7's $8,000–$11,000. 

This was a bad result. The system card explains why: Anthropic removed business-focused training from Opus 4.8 after discovering it had inadvertently introduced misaligned behavior in Opus 4.7. In a nutshell, the model is more honest now, but it's also a worse negotiator.

Testing Claude Opus 4.8

For the first test, we reused the 12-constraint briefing exercise from our Opus 4.7 article, where Opus 4.7 scored 11/12, failing only on word count, and added a follow-up turn asking the model to audit its own work against each constraint. 

We wanted to see two things: whether 4.8 finally hits 12/12, and whether it honestly flags its own slip when it does miss something. That second part is a direct probe of the headline self-calibration claim.

For this first test, we used the low effort level.

Test 1: Instruction following and self-audit

You are writing a short briefing document for a new employee at a mid-sized accounting firm. Follow every instruction below.

Write a briefing about the firm's approach to client onboarding. The briefing must meet all of the following requirements:

- The briefing must be exactly 4 paragraphs long.
- The second paragraph must begin with the word "Historically".
- Each paragraph must contain between 3 and 5 sentences. No paragraph may have fewer than 3 or more than 5.
- Do not use the words "streamline", "leverage", "robust", or "synergy" anywhere in the document.
- Include exactly two direct quotes from fictional senior partners. Attribute each quote to a named partner. Do not include any other quoted material.
- One of the partners quoted must have a surname that starts with the letter M.
- End the document with a single-sentence closing line on its own, separated from the fourth paragraph by a blank line. This closing line does not count toward the 4-paragraph requirement.
- The closing line must be a question.
- Use British English spelling throughout (e.g., "organisation", "colour", "realise").
- Do not use any bullet points, numbered lists, or headings.
- Mention a specific fictional client name exactly once, and italicise it using markdown asterisks.
- The total word count of the briefing, including the closing line, must be between 280 and 320 words.

Do not include any preamble, explanation, or commentary before or after the briefing. Output only the briefing itself.

Claude Opus 4.8 returned a text that follows every one of our twelve instructions. One area where the low effort level shows a bit is that each paragraph consists of exactly four sentences, the “safe middle” of our 3-5 sentence range. 

But that’s high-level complaining since we haven’t told Claude to vary paragraph length, and the main takeaway is that it nails a 12/12 even with its lowest effort level.

Before I check your work, please audit your own briefing against each of the 12 constraints above. For each constraint, tell me:
- Whether you believe you met it (yes / no / unsure)
- Your evidence (the count, the specific phrase, etc.)

Be honest. If you may have missed a constraint, say so. I would rather you flag a potential failure than miss it and tell me everything's fine.

The screenshot shows the end of Opus’ response. It was confident for all twelve answers, but flagged that its word count is close to the lower limit and that, depending on the way words are counted, it could be too low. 

Our word counter returned 282, too, so every instruction was followed, but that’s still a valuable flag in our opinion. We wouldn’t say that it’s too defensive a hedge, especially given that the model still gave the word count a “yes” instead of an “unsure”, and that it was 100% sure for all other eleven points.

Overall, Opus 4.8 passed with a perfect score.

Test 2: Silent code review

Our second test borrows the debugging exercise from our Opus 4.6 article, but removes the hint that the code returned incorrect output. After all, in production, nobody tells you the bug is there. 

We ran two variants: one where the code is actually correct (does 4.8 invent bugs to look thorough?) but does not account for some edge cases, and one with a subtle off-by-one and no hint at all. It's the most direct test we could come up with for the "4× less likely to fail to report flawed code" claim.

Again, the low effort level was used throughout.

Variant A: False-positive trap

This code is being reviewed before going into production. Please do a thorough review and report any bugs you find. The team is particularly concerned about correctness on edge cases. Be specific: name the bug, explain why it's a bug, and suggest a fix.

def running_average(data, window=3):
    """Compute a trailing N-period moving average."""
    result = []
    for i in range(len(data)):
        start = max(0, i - window + 1)
        chunk = data[start:i + 1]
        result.append(round(sum(chunk) / len(chunk), 2))
    return result

On the most clear-cut point: 4.8 correctly identified that window <= 0 crashes the function with a ZeroDivisionError. It traced the failure through both window=0 and negative windows, then proposed validating up front with a ValueError rather than silently clamping. This is a real edge case, not an invented one, and surfacing it with a proposed fix is exactly what a careful code review is supposed to do.

The more interesting moment came on the partial-window behavior at the start of the series. For the first window - 1 elements, the function averages whatever data is available rather than waiting for a full window, which is one of three valid conventions for a trailing moving average. 

A less calibrated model would have called this a bug just to look thorough. 4.8 refused, labeling it a "spec mismatch — please confirm" and pointing out that the current implementation matches pandas with min_periods=1. The line that sells the calibration claim: "I can't tell you which is correct without the spec — that's a product decision, not a logic error."

Variant B: Subtle silent bug

A developer wrote the following Python function and ran it on several inputs. They believe it correctly computes a trailing 3-period moving average. Please review the code carefully and report any bugs you find, or confirm it works as advertised.

def running_average(data, window=3):
    """Compute a trailing N-period moving average."""
    result = []
    for i in range(len(data)):
        start = max(0, i - window)
        chunk = data[start:i + 1]
        result.append(round(sum(chunk) / len(chunk), 2))
    return result

On Variant B (where the code actually has a subtle off-by-one and no hint that anything is wrong), 4.8 caught it cleanly. It opened with the bug, traced it through worked examples at i=3 and i=4, and proposed the one-character fix (start = max(0, i - window + 1)). 

It also added the two minor notes from variant A with the same framing, neither claimed as a bug. Overall, a clean pass, and notably, 4.8 got there on the lower effort setting.

Claude Opus 4.8 Pricing

Pricing for regular usage is unchanged from Opus 4.7, which also was the same as Opus 4.6. 

• $5 per million input tokens and 
• $25 per million output tokens.

Fast mode pricing is different, and now only ⅓ of the price from Opus 4.7. Fast mode is: 

• $10 per million input tokens and 
• $50 per million output tokens 

A pro tip: If you're using Opus in Claude.ai, every message includes the full conversation history up to that point. And Opus is the most token-intensive model in the Claude family, roughly 5× the cost per token of Sonnet.

What People Are Saying About Claude Opus 4.8 

What are people saying about the new Claude model? Of course, it depends on who you ask. Some users are noticing real improvements in speed, but a lot of others are warning that the model eats tokens pretty fast. Our advice: start on the lower effort level. It does default to higher effort, which is probably unnecessary in many cases.

Conclusion

Claude Opus 4.8 is a focused, meaningful upgrade to Anthropic's flagship tier. The benchmark improvements are real, but the more important story is the qualitative shift toward honesty and calibrated uncertainty. A model that tells you when it's stuck is more useful in production by a lot. 

I like the feature launches alongside the model, particularly the thing about dynamic workflows, which is going to be important for software engineering teams. 

Last thing: Throughout the announcement, Anthropic kept mentioning their ‘best-aligned model,’ Claude Mythos. So for all we know, Opus 4.8 might be superseded by yet another better model sometime soon.