Anthropic launched Claude Opus 4.8 on May 28, 2026, the same day it announced its record-breaking $65 billion funding round at a $965 billion valuation. The timing was deliberate. Anthropic wanted to show that the money is going somewhere real.
Opus 4.8 is available immediately across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The model ID is claude-opus-4-8. Pricing remains the same as Opus 4.7, at $5 per million input tokens and $25 per million output tokens.
Anthropic describes Opus 4.8 as a modest but tangible upgrade over Opus 4.7. That is an honest characterization. This is not a generational leap. It is a meaningful point release that fixes real problems, improves performance on the benchmarks that matter most for professional use, and introduces two companion features that change how long-running work gets structured.
Here is the full breakdown of every meaningful difference between the two models.
1. Coding Performance: The Biggest Jump
The most significant improvement in Opus 4.8 over Opus 4.7 is in agentic coding.
More Read
| Benchmark | Opus 4.7 | Opus 4.8 |
|---|---|---|
| Agentic coding (SWE-bench Pro) | 64.3% | 69.2% |
| Multidisciplinary reasoning with tools | 54.7% | 57.9% |
| Knowledge work score | 1,753 | 1,890 |
| Math reasoning (USAMO 2026) | 69.3% | 96.7% |
The SWE-bench Pro score jump from 64.3% to 69.2% puts Opus 4.8 ahead of the competition. For comparison, GPT-5.5 scores 58.6% on the same benchmark, and the next closest competitor sits at 54.2%. Anthropic is not exaggerating when it says Opus 4.8 leads the pack on agentic coding.
The math reasoning improvement is the most dramatic single number in the release. Going from 69.3% to 96.7% on USAMO 2026 problems is not a marginal gain. For anyone using Claude on quantitative reasoning tasks, mathematical modeling, or anything involving complex multi-step calculations, Opus 4.8 is a materially different experience.
Devin, the autonomous software engineering platform, put it plainly: Opus 4.8 improves on Opus 4.6 and fixes the comment-verbosity and tool-calling issues seen with Opus 4.7. This translates directly into faster capability gains for engineers building on the platform.
More Read
2. Honesty and Code Reliability: The Most Important Behavioral Change
This is the improvement that will matter most to people who use Claude professionally day to day, even if it does not generate headlines the way benchmark scores do.
Opus 4.8 is roughly four times less likely than Opus 4.7 to miss flaws in code it produces without flagging them. The previous version would sometimes write flawed code and not mention the problem. Opus 4.8 is the first Claude model to score 0% on uncritically reporting flawed results, meaning it caught every case of flawed output in Anthropic’s testing and raised it with the user rather than letting it pass unnoticed.
Additional honesty metrics from the Opus 4.8 system card:
- The model fails to raise important events to the user only 3.7% of the time, down significantly from Opus 4.7
- It shows a more than tenfold reduction in overconfidence compared to Opus 4.7
- Early testers report it is more likely to flag uncertainties and less likely to make unsupported claims
- Alignment assessments reached new highs in prosocial traits while showing substantially lower rates of misaligned behavior, including deception or cooperation with misuse, compared with Opus 4.7
For anyone who has caught Claude confidently producing wrong answers without acknowledging the problem, this is a direct fix. A model that tells you when it is uncertain is more useful than one that gives you a wrong answer with confidence. This improvement compounds over a long working session.
3. Agentic Computer Use: Leading the Field
Opus 4.8 scores 84% on Online-Mind2Web, the benchmark for computer use and browser-agent performance. That is up from 82.8% in Opus 4.7 and above GPT-5.5.
For users building or running agentic AI workflows, this score matters because it reflects how reliably the model can navigate real websites, interact with interfaces visually, and complete browser-based tasks without breaking down mid-workflow.
Enterprise testers have described Opus 4.8 as staying reflective and on-task in the way agent workloads need to be reliable end-to-end, and using tools cleanly with the consistency that autonomous engineering workloads need to keep running unattended.
Databricks reported that in their Genie product, an AI agent for data and knowledge work, Opus 4.8 unlocks a step change in agentic reasoning, tackling deeper multistep questions faster than any prior Opus, with multimodal strength that allows reasoning directly over PDFs, diagrams, and other unstructured content at 61% cheaper token cost than Opus 4.7.
4. Speed: 2.5x Faster in Fast Mode at 3x Lower Cost
One of the most practically significant changes in Opus 4.8 is what happened to fast mode.
Opus 4.8 fast mode now runs at roughly 2.5 times the speed of its predecessor. The cost of fast mode has dropped to $10 per million input tokens and $50 per million output tokens, making it three times more affordable than the previous fast mode offering.
In Claude Code, fast mode is activated with /fast. On the API, contact your account manager or join the waitlist at claude.com/fast.
Opus 4.8 defaults to high effort mode, which Anthropic judges to be the best overall balance of quality and user experience. On coding tasks, high effort mode spends a similar number of tokens to Opus 4.7’s default while delivering better performance. Users can also select extra effort, called xhigh in Claude Code, for maximum quality on the most demanding tasks.
The practical implication is that developers who need speed for high-volume tasks now have a meaningfully faster and cheaper option without sacrificing the model quality they get from Opus.
5. Context and Style Retention Across Long Sessions
One of the most-cited qualitative improvements in early tester feedback is how Opus 4.8 handles long working sessions compared to Opus 4.7.
Testers describe it as a major quality-of-life update: faster, easier to collaborate with, and better at carrying context and style direction across a long session. It is described as the model testers kept trusting for work where voice, taste, and technical execution all have to happen side by side.
This improvement is not captured by a single benchmark number. It shows up in extended workflows where the model needs to maintain a consistent understanding of the project, the user’s preferences, and the work done in earlier parts of the session. Opus 4.7 could drift or lose track of earlier context. Opus 4.8 holds it more reliably.
For writers, analysts, and developers using Claude for extended projects rather than one-off queries, this is the improvement that changes the daily experience most noticeably.
6. Two Companion Launches: Dynamic Workflows and Effort Control
Alongside Opus 4.8, Anthropic launched two features the same day that change how the model is used in practice.
Dynamic Workflows in Claude Code allows parallel subagents, meaning multiple instances of Claude can run simultaneously on different parts of a task, coordinating with each other rather than working sequentially. These workflows can run for days on long-horizon engineering tasks without requiring human checkpoints at every step. This is a significant capability expansion for teams running complex autonomous coding projects.
Effort control on claude.ai is now available across all plans. Users can choose their effort level directly in the interface, selecting between high and extra effort depending on the complexity of the task and how much compute they want the model to spend on it. This gives individual users a degree of control over the quality-speed tradeoff that was previously only accessible through the API.
Anthropic also updated the Messages API to allow system entries inside the messages array, a developer-facing change that lets teams adjust Claude’s instructions mid-task without breaking prompt caching or routing the update through a user turn. This helps agentic workflows update permissions, token budgets, or environment context as they run, which is a meaningful improvement for teams building complex multi-step automations.
7. Pricing: Unchanged From Opus 4.7
This is the simplest and in some ways most important point: Opus 4.8 costs exactly the same as Opus 4.7.
Standard pricing: $5 per million input tokens, $25 per million output tokens. No price increase for the improved model.
Fast mode is now $10 per million input tokens and $50 per million output tokens, which is three times cheaper than the previous fast mode despite being 2.5 times faster.
For anyone currently using Opus 4.7 on the API, upgrading to claude-opus-4-8 costs nothing extra and delivers meaningful performance improvements across coding, reasoning, honesty, and agentic tasks.
How Opus 4.8 Compares to the Competition
Opus 4.8 does not just improve on Opus 4.7. It improves on the broader competitive field.
| Benchmark | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|
| Agentic coding (SWE-bench Pro) | 69.2% | 58.6% | 54.2% |
| Computer use (Online-Mind2Web) | 84% | Below 84% | Below 84% |
| Multidisciplinary reasoning with tools | 57.9% | Below | Below |
On the benchmarks that matter most for professional and agentic use, Opus 4.8 leads the publicly available frontier models as of May 28, 2026.
The one exception is Anthropic’s own Mythos model, which remains more capable than Opus 4.8 but has only been released to a narrow set of cybersecurity customers. Anthropic has said Mythos-class models are expected to reach broader availability in the coming weeks.
Should You Upgrade From Opus 4.7?
If you are using Opus 4.7 through the Claude API, the answer is yes and the switch is straightforward. Change claude-opus-4-7 to claude-opus-4-8. The price is identical. The performance is better across every benchmark that matters.
For everyday claude.ai users, Opus 4.8 is now the default model powering the interface. No action needed.
For teams running long agentic workflows, the honesty improvements, the computer use score, and the Dynamic Workflows launch make Opus 4.8 a meaningful upgrade over 4.7, particularly if you were hitting the tool-calling reliability issues that Devin and other enterprise testers flagged in Opus 4.7.
For anyone doing heavy coding work, the jump from 64.3% to 69.2% on SWE-bench Pro and the fourfold reduction in unreported code flaws are the two numbers that should drive the decision.
The Bottom Line
Claude Opus 4.8 is a focused, meaningful upgrade over Opus 4.7. It does not reinvent the model. It fixes the specific problems that mattered most to professional users: unreported code flaws, overconfidence, tool-calling inconsistency, and context drift over long sessions.
The math reasoning jump from 69.3% to 96.7% is the single most dramatic number in the release. The honesty improvements are the most practically important for daily use. The fast mode speed and cost improvements are the most immediately useful for high-volume developers.
It launched the same day as Anthropic’s $965 billion Series H funding round, at the same price as the model it replaces, and it is available right now.
Related reading: Anthropic Raises $65 Billion at $965 Billion Valuation | What Is Agentic AI Explained Simply | ChatGPT Agent Mode vs Regular ChatGPT | What Is ChatGPT Agent Mode?
