Anthropic introduces Claude Opus 4.8, an update that promises to be faster, more reliable and more collaborative without raising the price. What does that mean for you? Less frustration when you work with an AI assistant and better results on complex tasks like code migrations, legal analysis, or agents browsing the web.
What's new in Claude Opus 4.8
Opus 4.8 improves on version 4.7 in several dimensions: judgment in agentic tasks, reasoning, efficient use of tools, and cost per token in multimodal scenarios. Anthropic shares evaluations where the model stands out on code, agent, reasoning, and practical-work benchmarks.
On top of that, Opus 4.8 brings a more explicit focus on honesty: it's four times less likely than its predecessor to let errors in code pass without flagging them. That’s not just a stat — it means less time chasing faulty outputs and more confidence handing off technical work.
How it improves collaboration
Do you work with agents or integrated tools? You’ll notice concrete changes. Early testers report better judgment: the model asks the right questions, detects its own mistakes, and questions shaky plans before making big changes.
In the Super-Agent benchmark, Opus 4.8 completed all end-to-end cases, outperforming previous models and matching GPT-5.5 at a similar cost. Impressive, right?
In legal and financial tasks Opus 4.8 raises the bar: it scored highest on the Legal Agent Benchmark and crossed the 10% all-pass threshold for the first time. If you rely on precision and consistency, that improvement translates into real work you can trust someone else to handle.
Important new features
-
Dynamic workflows: available in research preview for Claude Code. Lets Claude plan and run hundreds of subagents in parallel within a single session, and verify their outputs before returning to you. Ideal for large-scale codebase migrations with automated tests as the success criteria.
-
Effort control on claude.ai and Cowork: now you can choose how much Claude “thinks.” At low settings it answers faster and uses fewer tokens; at high settings or
xhigh(extra) andmaxit spends more tokens for deeper responses. Opus 4.8 ships by default at high effort. -
Messages API: the API now accepts system inputs inside the messages array, so you can update instructions mid-task without breaking prompt caching. That helps when you need to change permissions, token budgets, or operational context while an agent is running.
Performance and usage examples
-
Opus 4.8 reached 84% on Online-Mind2Web, standing out in browser use and tasks that require structured web access.
-
In trials with Databricks Genie, Opus 4.8 improved agentic reasoning and handled multimodal content (PDFs, diagrams) with a 61% lower cost per token compared to Opus 4.7.
-
For financial orchestrators like Hebbia, teams reported more accurate citations and better token-retrieval efficiency — handy when you analyze dense case files.
Costs and availability
Claude Opus 4.8 is available globally today, at the same base price as Opus 4.7. Public rates announced:
- Regular use: $5 per million input tokens and $25 per million output tokens.
- Fast mode: a model 2.5× faster and now with adjusted pricing — $10 per million input tokens and $50 per million output tokens. Anthropic says this fast mode is three times cheaper than previous models’ fast modes.
Developers can access the model via the API with the identifier claude-opus-4-8.
A pragmatic step toward more powerful models
Opus 4.8 positions itself as a tangible upgrade, not a revolution. It brings practical improvements: more honesty, better long-context handling, tool-call efficiency, and user control options. These are the kind of changes that ease daily work and make it safer to delegate more complex tasks.
Anthropic also says it's working on even more capable models under Project Glasswing and on rolling out cybersecurity protections to offer them safely. Meanwhile, Opus 4.8 is a solid option for teams that need reliable agents today.