Claude Opus 4.8 is Anthropic’s most capable generally available model — and the most important upgrade isn’t raw intelligence, it’s honesty. Released May 28, 2026, it’s built for complex, long-running agentic workflows where a confident wrong answer is far more dangerous than an uncertain correct one.

Feature	Claude Opus 4.7	Claude Opus 4.8
Flawed code flagging	Baseline	~4x more likely to flag
Adaptive thinking	✅	✅ Improved
Dynamic Workflows	❌	✅ New
Effort control	Limited	✅ Full (low / medium / high)
Mid-conversation system messages	❌	✅ New
Context window	1M tokens	1M tokens
Max output	128k tokens	128k tokens
Input pricing	$5 / MTok	$5 / MTok
Output pricing	$25 / MTok	$25 / MTok
Migration effort	—	One string change

🚨 The Real Problem With AI Agents

Most people assume a smarter model just means fewer mistakes. The dangerous edge case is different: an AI agent that makes a mistake and sounds completely confident while doing it. In a short chat session, that’s annoying. In a 50-step autonomous run overnight, it compounds silently until you’re debugging output that was wrong at step three.

Anthropic’s training improvements make the model significantly more calibrated — its expressed confidence tracks more closely with its actual accuracy. The result: it’s approximately 4x less likely than Opus 4.7 to pass flawed code without flagging it. That’s a trained behavioral property, not a filter layer on top of the output.

📊 Where the Benchmarks Land

On SWE-Bench Pro the new model reaches 69.2%, Terminal-Bench 2.1 hits 74.6%, and agentic computer use scores 83.4%. Multidisciplinary reasoning lands at 49.8% without tools and 57.9% with tools. Knowledge work scores 1890 and agentic financial analysis comes in at 53.9%.

These numbers are worth noting, but they’re not the story. The honesty improvements — visible in the misaligned behavior chart, where Opus 4.8 scores approximately 1.83 versus Opus 4.7’s 2.48 (lower is better) — tell you how the model behaves when it’s wrong. That’s what matters in production.

🛠️ What Dynamic Workflows Actually Do

Dynamic Workflows are the structural addition that changes how the model scales to large tasks. Instead of one instance working step by step through a long task list, Claude can now plan a large task, dispatch many subagents to work in parallel, verify their outputs independently, and consolidate a single report.

Practical use cases include security audits across a monorepo (each module checked independently), bug hunts split across directories, framework migrations batched by file, and any workload with clear parallel structure. The coordinator merges results at the end — tasks that previously required sequential agent loops can now run in parallel, reducing wall-clock time significantly on large codebases.

Note: Dynamic Workflows require an orchestration layer (Claude Code, the Anthropic Agent SDK, or a custom tool loop). The model is the reasoning engine; you provide the runner.

⚡ Effort Control and Adaptive Thinking

Effort control is a practical API parameter (effort: "low" / "medium" / "high") that explicitly sets how much reasoning the model applies to a given task. The default is "high". Pair it with thinking: {"type": "adaptive"} and the model decides per turn whether extended reasoning is warranted — a quick lookup gets a direct response, a complex architecture question gets full deliberation.

This matters for cost control in agentic loops: not every step in a 50-step workflow needs deep reasoning. Routing simple steps to effort: "low" or to a cheaper model like Sonnet 4.6 keeps costs from ballooning on tasks that don’t need Opus-level depth.

💻 Coding Improvements and Claude Code

For developers using Claude Code, improvements span the full task lifecycle: better long-context code understanding across large files, stronger tool use accuracy, fewer derailments in long sessions after context compaction, and more consistent plan-following from start to finish. The model is more likely to catch its own bugs before handing you the result — which means fewer review cycles on long automated tasks.

The model ID is claude-opus-4-8. Migration from Opus 4.7 is, for most codebases, exactly one string change. No breaking API changes, no new required parameters, no prompt refactoring needed.

💡 When to Use It vs. Cheaper Models

Claude Opus 4.8 costs $5 per million input tokens and $25 per million output tokens — identical to Opus 4.7. It is not the right choice for every task.

Reach for it when the task is complex, when mistakes are expensive, when you need long-context reasoning across large codebases, or when you are building or running AI agent systems. Use Haiku 4.5 or Sonnet 4.6 when the task is simple, when speed matters more than reasoning depth, or when you are doing lightweight generation and cost is the primary constraint.

FAQ

What is new in Claude Opus 4.8?

Claude Opus 4.8 introduces honesty improvements (approximately 4x less likely to pass flawed code without comment), Dynamic Workflows for parallel subagent coordination, full effort control, mid-conversation system messages, a lower prompt cache minimum (1,024 tokens), and a research-preview fast mode. It shares the same pricing as Opus 4.7.

How does Claude Opus 4.8 compare to Opus 4.7?

The core API behavior is compatible — no breaking changes, same constraints (no temperature, top_p, or top_k). The main differences are the honesty and calibration improvements, Dynamic Workflows, and the new effort control parameter. Code running on Opus 4.7 runs on the new model with just the model name change.

What are Dynamic Workflows in Claude Opus 4.8?

Dynamic Workflows allow the model to plan a large task, spawn and coordinate many subagents in parallel, verify their outputs, and report back with a consolidated result. They are useful for security audits, bug hunts, migrations, and any task with clear parallel structure. They require an orchestration layer such as Claude Code or the Anthropic Agent SDK.

How much does Claude Opus 4.8 cost?

Claude Opus 4.8 costs $5 per million input tokens and $25 per million output tokens — the same pricing as Opus 4.7. There is no price increase for the upgrade.

Should I upgrade from Claude Opus 4.7 to 4.8?

For most developers, yes. The migration is a single string change (claude-opus-4-7 → claude-opus-4-8), there are no breaking API changes, and the honesty and reliability improvements are meaningful for agentic workflows. The only reason to delay is if you are on a pinned version for compliance reasons.

What is effort control in Claude Opus 4.8?

Effort control is an API parameter (effort: "low", "medium", or "high") that sets how deeply the model reasons before responding. The default on Opus 4.8 is "high". Setting lower effort on simple steps in an agentic loop reduces latency and cost without sacrificing quality where it matters.

✨ Key takeaways

• ✅ Anthropic released its most capable generally available model on May 28, 2026
• ⚠️ It is approximately 4x less likely than Opus 4.7 to pass flawed code without flagging it
• 🚀 Dynamic Workflows enable parallel subagent coordination for large codebases and complex tasks
• ⚡ Effort control lets you tune reasoning depth per task, keeping costs under control in long agentic loops
• 💻 Migration from Opus 4.7 is one string change — no breaking API changes
• 💡 Use it for complex coding, agentic workflows, and high-stakes automation; use cheaper models for simple tasks

For complex coding, agentic pipelines, and any workflow where the cost of a silent mistake is high, Claude Opus 4.8 is the model worth reaching for.

---