Anthropic released Claude Opus 4.8 on May 28, marking a significant leap in code safety and reasoning capabilities. The new model scores 88.6% on SWE-Bench Verified (a widely used software engineering benchmark), up from 87.6% on the previous version, while introducing a critical safety improvement: it is four times less likely than its predecessor to let flaws in its own code go unflagged. Token pricing remains unchanged at $5 per million input tokens and $25 per million output tokens, making the upgrade immediately accessible to existing Claude users.

What Makes Opus 4.8 Different From Previous Claude Models?

Beyond the benchmark numbers, Opus 4.8 introduces what Anthropic calls "honest evals," a measure of how often the model flags potential problems in its own work. For developers building agentic systems,where an AI agent works autonomously on complex tasks,silent failures are far more costly than slower performance. The 4x improvement in flagging code flaws means fewer surprises when Claude-powered agents run unattended.

The model also excels at specialized coding tasks. On agentic coding benchmarks, Opus 4.8 scores 69.2%, significantly ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%). On the Super-Agent benchmark, Anthropic reports that Opus 4.8 is the only model that completes every test case end-to-end without human intervention.

A new "Fast mode" runs at 2.5 times the speed of standard inference for $10 per million input tokens and $50 per million output tokens, roughly 3 times cheaper than the previous Fast tier. This creates a practical two-tier option: use standard Opus 4.8 for complex reasoning tasks and Fast mode for rapid iteration or lower-stakes work.

How Can Developers Use Dynamic Workflows to Scale AI Work?

One of the most ambitious new features is dynamic workflows in Claude Code, a research preview that lets developers orchestrate multiple AI agents working in parallel. A developer writes a natural language prompt, and Claude generates a JavaScript orchestration script that runs in the background. The system can spawn up to 16 concurrent subagents and up to 1,000 total subagents per workflow, enabling complex multi-step projects to be broken into parallel subtasks.

The real-world impact is striking. Developer Jarred Sumner used dynamic workflows to port Bun, a JavaScript runtime, from the Zig programming language to Rust, a task involving roughly 750,000 lines of code. The result: 99.8% of the existing test suite passed, and the entire migration took 11 days from first commit to merge. That kind of "weekend megaproject" capability is what Anthropic is positioning dynamic workflows to enable.

Steps to Get Started With Claude Opus 4.8 and New Features

• Access the Model: Claude Opus 4.8 is available today across Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry, and GitHub Copilot. Pricing remains flat at $5/$25 per million tokens, with no premium surcharge for the new version.
• Try Dynamic Workflows: Dynamic workflows are available in research preview in Claude Code CLI, Desktop, and VS Code extension for Max, Team, and Enterprise plans. Track workflow runs using the /workflows command.
• Adjust Effort Levels: Claude.ai and Cowork now expose an effort selector (Low, Medium, High, or Max) for each request. Choose Low for faster, cheaper responses on simple tasks and Max for complex legal drafts, reports, or code review.
• Leverage Fast Mode: For rapid iteration or lower-stakes work, use Fast mode at 2.5x speed and roughly 3x lower cost than standard Opus 4.8, ideal for brainstorming or exploratory coding.
• Monitor Code Quality: Use the improved /code-review and /simplify commands in Claude Code 2.1.154, which now separates bug-hunting from cleanup tasks and defaults to xhigh effort for the hardest problems.

Why Is Opus 4.8's Rollout Across Multiple Platforms Significant?

Anthropic shipped Opus 4.8 to its API, Bedrock, Vertex AI, Microsoft Foundry, and GitHub Copilot all on the same day, May 28. This simultaneous availability across competing platforms signals a shift in how frontier AI models are distributed. The differentiation between coding environments like GitHub Copilot and Cursor is no longer about which model they use, but rather the user experience, automation features, and integrations built on top of the model.

GitHub made Opus 4.8 generally available to Copilot Pro+, Business, and Enterprise users on May 28. Until usage-based billing flips on June 1, Opus 4.8 carries a 15x premium request multiplier, meaning heavy users may want to keep a cheaper model in rotation for routine tasks. After June 1, the pricing aligns with the standard $5/$25 tier.

The broader context is Anthropic's $65 billion Series H funding round, closed on May 28 at a $965 billion post-money valuation. The company disclosed a $47 billion run-rate revenue earlier in May, and the funding will support alignment research, expanded compute capacity, and continued scaling of Claude Code and Cowork, Anthropic's enterprise collaboration platform.

What Do the Benchmark Improvements Actually Mean for Developers?

Opus 4.8 posts strong results across multiple benchmarks: 74.6% on Terminal-Bench 2.1 (a test of command-line reasoning), 93.6% on GPQA Diamond (a knowledge benchmark), and 1890 Elo on GDPval-AA (a reasoning benchmark). However, the most practical improvement for developers is the "honest evals" metric. A model that flags its own mistakes four times more often than its predecessor reduces the cognitive load on developers reviewing AI-generated code.

The 1 million token context window is now the default across all Claude API platforms, with a maximum output of 128,000 tokens. In practical terms, this means Claude can process roughly 100,000 words at once, enough for a full technical specification, a large codebase, or an entire research paper in a single request.

For teams using Claude Code, the leaner system prompt is now the default for every model except Haiku, Sonnet, and older Opus versions. This reduces token usage and makes Claude less likely to surface multiple-choice prompts when it already has enough context to act directly. The /simplify command has been re-scoped to focus on cleanup and efficiency, while heavier bug-hunting work stays in /code-review --fix.

The practical takeaway is that Opus 4.8 represents incremental but meaningful progress on the frontier of AI-assisted coding. The safety improvements matter most for autonomous agents, the speed improvements matter for rapid iteration, and the pricing stability matters for cost planning. For developers already using Claude, the upgrade is immediate and free; for those evaluating coding AI tools, Opus 4.8's performance on agentic benchmarks and its availability across multiple platforms make it a strong contender.