What Is Claude Opus 4.8? Everything Developers Need to Know

Anthropic shipped Claude Opus 4.8 on May 28, 2026, and made it available the same day across the Claude API, Claude apps, Claude Code, and the major cloud platforms. It’s the most capable model in the Claude family, built for complex reasoning, long-horizon agentic coding, and high-autonomy work. The API model ID is claude-opus-4-8.

If you already build on Opus 4.7, the upgrade is a one-line model swap. The headline change isn’t a new pricing tier or a bigger context window; both stay the same. It’s quality. According to Anthropic’s announcement, Opus 4.8 is about four times less likely than 4.7 to let a flaw in code pass unremarked, and it’s more honest about what it doesn’t know. This guide covers what changed, how to access it, and whether it’s worth switching.

The short version

Three facts matter on day one:

• Model ID claude-opus-4-8: live now on the Claude API, AWS, Vertex AI, and Microsoft Foundry
• Same price as Opus 4.7: $5 per million input tokens, $25 per million output tokens in standard mode
• 1M token context, 128K token output: unchanged from 4.7, so your token budgets carry over

What you get on top of that:

• A new effort parameter that trades thoroughness for token efficiency across the whole response
• Adaptive thinking, where the model decides how much to reason per request
• Dynamic Workflows in Claude Code, which can spin up hundreds of parallel subagents in one session
• Better honesty, fewer code defects slipping through, and more efficient tool calling

For the full per-token math and cost scenarios, see the Opus 4.8 pricing breakdown. To start building, jump to the Opus 4.8 API guide.

What’s actually new in Opus 4.8

Opus 4.8 keeps the specs of 4.7 and improves the model underneath. The gains cluster in four areas.

Code quality. The model catches its own mistakes more often. Anthropic reports roughly a 4x drop in flaws that pass review unremarked compared to 4.7. For agentic coding, that means fewer silent bugs in generated diffs.

Honesty and alignment. Opus 4.8 flags uncertainty more readily and makes fewer unsupported claims. Anthropic also reports lower rates of deception and misuse cooperation than 4.7. If you run agents unattended, that judgment matters more than a benchmark point.

Tool calling. The model picks tools more efficiently and wastes fewer calls, which cuts both latency and token spend in agent loops.

Effort control. This is the biggest API-visible change, and it deserves its own section.

Effort control: one model, five gears

The effort parameter lets you dial how eagerly Claude spends tokens. It lives inside output_config and accepts five levels: low, medium, high, xhigh, and max. The default is high on every surface, including the API and Claude Code.

{
  "model": "claude-opus-4-8",
  "max_tokens": 4096,
  "messages": [{"role": "user", "content": "Refactor this module."}],
  "output_config": { "effort": "xhigh" }
}

The key detail: effort affects all tokens, not only reasoning. That includes text, tool calls, and function arguments. Lower effort means Claude makes fewer tool calls and writes terser responses. Higher effort means deeper analysis and more thorough work.

Anthropic’s guidance for Opus 4.8 is to start at xhigh for coding and agentic tasks, keep high as the floor for most reasoning-heavy work, and step down to medium or low only after your evals confirm the lower level holds quality. Full detail lives in Anthropic’s effort docs.

Adaptive thinking replaces manual budgets

Opus 4.8 uses adaptive thinking. You set thinking: {type: "adaptive"} and the model decides when and how much to reason per request. At high, xhigh, and max effort, it almost always thinks deeply. At lower levels, it can skip thinking on simple problems.

One thing to know before you migrate: manual extended thinking with budget_tokens is not supported on Opus 4.8 and returns a 400 error. If you carried that pattern over from an older Opus, switch to adaptive thinking plus the effort parameter. We walk through the exact request shape in the Opus 4.8 API guide.

Dynamic Workflows in Claude Code

The flashiest new feature ships inside Claude Code. Dynamic Workflows let a single session launch hundreds of parallel subagents to chew through large, branching tasks. Under the hood, this is the xhigh effort level paired with mid-conversation system messages, the same Messages API update that now accepts system entries partway through a conversation rather than only at the start.

That mid-conversation capability is what gives an orchestrator agent standing permission to spawn workers as the task unfolds. If you want the mechanics and how to build a similar orchestration mode through the raw API, see the Claude Code Dynamic Workflows deep-dive. For background on how Claude Code structures agent runs, our Claude Code agent harness breakdown is a good primer.

Benchmark highlights

Anthropic’s headline numbers focus on agentic work:

• Beats GPT-5.5 on the Super-Agent benchmark, the suite that measures end-to-end task completion
• Tops the Legal Agent Benchmark and is the first model to break 10% overall on it
• 84% on Online-Mind2Web, a web-navigation agent test

These are agent scores, not chat scores, which tells you where Opus 4.8 is aimed. For a head-to-head against the other frontier models, read Opus 4.8 vs GPT-5.5 vs Gemini 3.5. The older Gemini 3.5 vs GPT-5.5 vs Opus 4.7 comparison still holds for the 4.7 baseline.

Opus 4.8 vs Opus 4.7 at a glance

The specs are identical on purpose. You’re paying the same rate for a model that makes fewer mistakes, so for most teams the migration is low-risk. Test your tool schemas and evals after the swap, then ship.

How to access Claude Opus 4.8

You have four front doors:

1. Claude API: use model ID claude-opus-4-8 against the Messages endpoint. Start with the Opus 4.8 API guide.
2. Claude apps: it’s the default high-end model at claude.ai for paid plans, with limited access on the free plan.
3. Claude Code: available as the top model, with Dynamic Workflows when you opt into the high-effort mode.
4. Cloud platforms: AWS (anthropic.claude-opus-4-8 on Bedrock), Vertex AI (claude-opus-4-8), and Microsoft Foundry, where the context window is capped at 200K tokens.

If you want to try it without a paid API plan first, the how to use Opus 4.8 for free guide covers the legitimate low-cost paths.

Who should use Opus 4.8

Opus 4.8 is built for the hard end of the workload spectrum. Reach for it when:

• You’re running long agentic coding sessions where a silent bug is expensive
• You need an agent to make sound judgment calls unattended
• You’re orchestrating multi-step tool use and want fewer wasted calls
• The task genuinely needs frontier reasoning, not a quick classification

For high-volume, latency-sensitive, or simple work, a smaller model or a lower effort level will serve you better and cost less. The point of effort control is that you no longer have to switch models to switch gears.

Testing Opus 4.8 before you ship it

A model swap is easy to write and easy to get wrong. Streaming chunks, tool-call validation, the new output_config shape, and adaptive-thinking responses all change the payloads your code has to parse. Before you push claude-opus-4-8 to production, replay your real requests against it and diff the output.

Apidog handles the full Messages API surface in one workspace:

• Save the Opus 4.8 endpoint as a request, attach your x-api-key, and hit Send
• Swap claude-opus-4-7 for claude-opus-4-8 on the same request and diff the responses
• Watch streamed chunks render inline with per-chunk timings
• Add assertions that catch schema drift when you change effort levels
• Mock the endpoint so you can test downstream code without burning credits

Download Apidog, point a request at the Messages endpoint, and paste in the curl snippet from the API guide. Setup takes about two minutes.

FAQ

Is Claude Opus 4.8 better than Opus 4.7? Yes, on quality. It catches roughly 4x more code defects, is more honest about uncertainty, and calls tools more efficiently. Pricing, context window, and max output are identical, so there’s little reason to stay on 4.7.

How much does Opus 4.8 cost? $5 per million input tokens and $25 per million output tokens in standard mode. Fast mode runs at $10 and $50 for 2.5x faster output. Full math is in the pricing breakdown.

What is the context window for Opus 4.8? 1M tokens of input and up to 128K tokens of output on the synchronous Messages API. The Batch API supports up to 300K output tokens with a beta header. On Microsoft Foundry the context window is 200K tokens.

Does Opus 4.8 support extended thinking? It uses adaptive thinking (thinking: {type: "adaptive"}), where the model decides how much to reason. Manual budget_tokens thinking is not supported and returns a 400 error.

What is the effort parameter? A setting inside output_config that controls how many tokens Claude spends across text, tool calls, and reasoning. Levels run low, medium, high (default), xhigh, and max.

Can I use Opus 4.8 for free? There’s no free API tier, but you can try it on the free plan at claude.ai with limits, or through trial credits. See the free access guide.

What are Dynamic Workflows? A Claude Code feature that launches many parallel subagents in one session, powered by xhigh effort and mid-conversation system messages. Details in the Dynamic Workflows guide.