Claude Opus 4.8 Is Here: Release, Benchmarks, and What Anthropic's New Model Means for Businesses

On May 28, 2026, Anthropic released Claude Opus 4.8 - described by the company as its most capable generally available model to date. Unlike the weeks of speculation around Sonnet 4.8, this is not a leak but an official release with a documented API ID, benchmarks, and immediate availability across all platforms.

This article summarizes what is verifiably known about the model on launch day - and what the jump from Opus 4.7 to 4.8 concretely means for businesses.

Key Release Facts at a Glance

• Model ID: "claude-opus-4-8" - a direct drop-in replacement for Opus 4.7, with no API breaking changes.
• Available from: immediately on May 28, 2026, everywhere - the Claude API, the Claude apps, Amazon Bedrock, and Google Vertex AI.
• Context window: 1M tokens by default on the API, Bedrock, and Vertex AI (200K on Microsoft Foundry).
• Max output: 128K tokens.
• Reasoning: adaptive thinking - the model reasons only when the turn requires it.
• Price: unchanged from Opus 4.7 - $5 input / $25 output per million tokens.

The bottom line in one sentence: Opus 4.8 arrives at the same price as 4.7 but performs better on nearly every measured dimension - effectively a free upgrade for existing users.

Benchmarks: Where Opus 4.8 Actually Improves

Anthropic positions Opus 4.8 as its new flagship and compares it directly with Opus 4.7, OpenAI's GPT-5.5, and Google's Gemini 3.1 Pro. The key published numbers:

• SWE-Bench Pro (agentic coding): Opus 4.8 69.2% · Opus 4.7 64.3% · GPT-5.5 58.6% · Gemini 3.1 Pro 54.2%.
• OSWorld-Verified (computer use): Opus 4.8 83.4% · Opus 4.7 82.8% · GPT-5.5 78.7% · Gemini 3.1 Pro 76.2%.
• Terminal-Bench 2.1: Opus 4.8 74.6% · Opus 4.7 66.1% · GPT-5.5 78.2% · Gemini 3.1 Pro 70.3%. GPT-5.5 still leads here - but Opus 4.8 gains over 8 points versus 4.7.
• GDPval-AA (economically valuable tasks): Opus 4.8 1890 · Opus 4.7 1753 · GPT-5.5 1769.
• Humanity's Last Exam (multidisciplinary reasoning): 49.8% without tools, 57.9% with tools - ahead of all three rivals.
• Online-Mind2Web (web agents): 84% - according to Anthropic a meaningful jump over both Opus 4.7 and GPT-5.5.
• Legal Agent Benchmark: the first model ever to break 10% overall on the strict "all-pass" standard.
• CursorBench: exceeds all prior Opus models across every effort level.

The pattern is clear: Opus 4.8 is strongest at long-horizon, agentic tasks - exactly where models code, research, or operate tools autonomously across many steps.

New in Opus 4.8: Features for Developers and Teams

Beyond raw benchmark gains, the release brings several practically relevant additions:

Mid-conversation system messages

Opus 4.8 accepts "system" messages immediately after a user turn within the messages array - with no beta header. This lets you append updated instructions in long-running conversations without restating the full system prompt. It preserves prompt-cache hits on the earlier turns and reduces input cost in agentic loops.

Effort default "high"

The effort parameter defaults to "high" on all surfaces - API and Claude Code alike. In claude.ai and Cowork, reasoning intensity can now also be controlled directly.

Fast mode (research preview)

With "speed: fast", the same model delivers up to 2.5x higher output tokens per second on the API - at a premium price of $10 input / $50 output per million tokens. According to Anthropic, that is roughly one-third of the previous fast-mode cost.

Lower prompt-cache minimum

The minimum cacheable prompt length drops to 1,024 tokens. Prompts that were too short to cache on Opus 4.7 now create a cache entry with no code change - a direct cost benefit.

Dynamic workflows in Claude Code

As a research preview, Claude Code now enables parallel subagents for large-scale tasks - such as migrations, audits, or broad code sweeps that are too large for a single context.

"Anthropic's Most Honest Model": Safety and Reliability

Anthropic explicitly markets Opus 4.8 as its most honest model yet. The concrete claims:

• Opus 4.8 is reportedly around four times less likely than its predecessor to let code flaws pass unremarked.
• The model shows improved honesty and flags uncertainties more actively rather than glossing over them.
• In internal alignment assessments, rates of misaligned behavior are substantially lower than on Opus 4.7.

For businesses this is more than a PR detail: a model that flags its own uncertainties and waves through buggy code less often reduces review and rework effort - and with it a central risk of deploying autonomous agents.

What Changes Technically - and What Stays the Same

Opus 4.8 inherits the API constraints of 4.7, so existing Opus 4.7 code runs unchanged:

• Sampling parameters: setting "temperature", "top_p", or "top_k" to non-default values still returns a 400 error. Guide behavior through prompting instead.
• Thinking: only adaptive thinking is supported. Fixed thinking budgets ("budget_tokens") are rejected with a 400 - control reasoning depth via the effort parameter.

There are also behavior changes that are not breaking changes but may warrant prompt updates:

• Fewer wasted thinking tokens at the same effort level, because the model decides per turn whether to think at all.
• Better tool triggering: Opus 4.8 less often skips a tool call the task required - an issue some 4.7 users reported.
• Better compaction and long-context quality: long agentic traces stay on task more reliably after compaction.

What the Release Concretely Means for Businesses

1. A free upgrade - but not a blind one

If you already run Opus 4.7 in production, you get a measurably better model at the same price structure. Still: don't swap the model ID blindly - test the workflows that matter to you. Because of more efficient thinking and more active tool triggering, outputs can shift subtly.

2. Parameterize model IDs

Hard-coded references to "claude-opus-4-7" should be replaced with a configurable variable. Then switching to "claude-opus-4-8" is a configuration change of minutes - with no redeploy.

3. Choose the right model stack

The best production architecture in 2026 is not a single all-purpose model but a routed stack:

• Haiku 4.5 for fast, simple triage tasks.
• Sonnet 4.6 for standard coding, agent workflows, and everyday analysis ($3 / $15 per million tokens).
• Opus 4.8 for complex reasoning, long-horizon agentic coding, high-autonomy work, and demanding vision tasks.

One important aside: a Sonnet 4.8 still does not exist. The only new 4.8 release is Opus; the current Sonnet model remains 4.6. If you rely on the cheaper Sonnet tier, don't let rumors drive your strategy.

4. Factor in compliance

If you deploy Anthropic models in customer service or process automation, the EU AI Act deployer obligations apply: AI disclosure, AI-literacy training, and clean documentation. A model switch is the ideal moment to review these points.

FAQ: Common Questions About Claude Opus 4.8

Is Claude Opus 4.8 already available? Yes. Since May 28, 2026, the model is available under the API ID claude-opus-4-8 across all platforms - at the same price as Opus 4.7 ($5 input / $25 output per million tokens).

What is the biggest difference from Opus 4.7? Opus 4.8 is stronger primarily at long-horizon, agentic tasks: fewer compactions, better tool triggering, and better benchmark scores (e.g., SWE-Bench Pro 69.2% instead of 64.3%). It also adds a clear focus on honesty and safety.

Does Opus 4.8 cost more than 4.7? No. Regular usage costs an unchanged $5 input / $25 output per million tokens. Optionally, a fast mode ($10 / $50) offers up to 2.5x higher output speed.

Is there now a Claude Sonnet 4.8 as well? No. As of May 28, 2026, Opus 4.8 is the only new 4.8 model. The current Sonnet model remains Sonnet 4.6.

Do I need to change my code to migrate? There are no API breaking changes versus Opus 4.7 - existing code runs. However, behavior changes such as more efficient thinking and more active tool triggering may make prompt updates worthwhile. temperature, top_p, top_k, and fixed thinking budgets remain unsupported, as on 4.7.

Which model should my company use? Usually a routed stack: Haiku 4.5 for fast triage, Sonnet 4.6 for standard coding and agent workflows, Opus 4.8 for complex reasoning, long-horizon agents, and vision.

Conclusion: Solid Progress at the Same Price

Claude Opus 4.8 is not a loud paradigm shift but a targeted, well-documented step forward - better agentic coding, more reliable behavior over long contexts, and a pronounced safety and honesty focus, all at the familiar Opus price. For businesses, the right response is not hype but pragmatism: test the workflows that matter, parameterize model IDs, and pick the right tier in your stack.

Want to know which Claude model offers the best price-performance for your workflows? Schedule a free initial consultation - we analyze your use cases and recommend the optimal model stack.

---

Sources

• Anthropic (May 28, 2026): Introducing Claude Opus 4.8 - Retrieved: 2026-05-28
• Anthropic Docs (2026): What's new in Claude Opus 4.8 - Retrieved: 2026-05-28
• OfficeChai (May 2026): Anthropic Releases Claude Opus 4.8, Beats Opus 4.7, GPT-5.5 On Many Benchmarks
• Inc. (May 2026): Anthropic Says Its Claude Opus 4.8 Model Is Its 'Most Honest' Yet