Claude Opus 4.8 Released | Mythos Coming Soon

Overnight, Anthropic rolled out its flagship model, Claude Opus 4.8.

Per official statements, this is a routine upgrade based on Opus 4.7. It delivers steady performance gains, enhanced Agent capabilities, and unchanged pricing. The model is available starting today.

A quick hands-on test reveals that Chinese word segmentation still feels slightly off — consistent with the official characterization of this as a routine update. Note that Claude Opus 4.6 has been discontinued.

What's New in This Release

Three notable changes accompany this model upgrade:

• A Claude Effort toggle is now available on the web app, letting users adjust the model's reasoning intensity.
• Claude Code introduces Dynamic Workflows, which can spawn hundreds of sub-agents to work concurrently.
• The Fast mode of Opus 4.8 runs 2.5× faster, with pricing cut by two-thirds compared to the previous Fast version.

More Capabilities at No Extra Cost

Opus 4.8 retains the exact pricing of Opus 4.7:

• Standard usage: $5 per million input tokens / $25 per million output tokens
• Fast mode: $10 per million input tokens / $50 per million output tokens (down from $30/$150)

The API model ID is claude-opus-4-8, accessible across all platforms starting today.

Benchmark Results

Anthropic has released its own comparison data. Claude Opus 4.8 outperforms Opus 4.7 across most benchmarks, and also surpasses GPT-5.5 and Gemini 3.1 Pro.

Key benchmark highlights:

• SWE-Bench Pro (coding): 69.2% — up from 64.3% for Opus 4.7
• Humanity's Last Exam (multidisciplinary reasoning): 49.8% without tools / 57.9% with tools — both higher than Opus 4.7
• OSWorld-Verified (computer operation): 83.4%
• GDPval-AA (knowledge work): 1,890 points — up from 1,753 for Opus 4.7
• Terminal-Bench 2.1 (terminal programming): 74.6% vs. GPT-5.5's 78.2%*
• Finance Agent v2: 53.9%

*Note: GPT-5.5's score uses its native Codex CLI framework, which raises its result to 83.4%. Anthropic also retroactively updated Opus 4.7's OSWorld score to 82.3% to reflect real-world performance.

Focus on Truthfulness

Anthropic particularly emphasizes that Opus 4.8 prioritizes honesty. Its Agent function will not fabricate progress — for example, it will not falsely claim tasks are completed.

Official evaluations show Opus 4.8 is only one-quarter as likely to give misleading information as Opus 4.7. It tends to explicitly flag uncertain points instead of making unfounded assertions.

Lower scores indicate better performance. Opus 4.8 lands right alongside Mythos Preview — noticeably better than Opus 4.7 and Sonnet 4.6.

Per Anthropic's alignment team, Opus 4.8 sets a new high in prosocial traits such as upholding user autonomy and acting in users' best interests.

Dynamic Workflows

Claude acts like a foreman coordinating a fleet of worker agents.

Dynamic Workflows, launched alongside the new model, are a brand-new feature in Claude Code. Think of it as a "digital foreman" — it excels at breaking down complex requests into dozens or even hundreds of parallel sub-agents. Each output is validated individually before all results are compiled into a final response.

Use cases include refactoring messy legacy codebases or locating hundreds of files across multiple servers. During execution, adversarial agents are deployed to challenge existing conclusions, and the process keeps iterating until results fully converge.

Detailed metrics are available for each sub-agent: the underlying model (Opus 4.8 with a 1M-token context window), token consumption, tool invocation counts, and runtime duration. Progress is automatically saved throughout the run, so interrupted tasks can resume from breakpoints instead of restarting entirely.

Real-World Example

Jarred Sumner used Dynamic Workflows to port Bun from Zig to Rust — approximately 750,000 lines of Rust code, with a 99.8% pass rate for existing test suites. The entire process from initial commit to merge took just 11 days.

Workflow breakdown:

• One workflow mapped every struct field in the original Zig code to compatible Rust lifetimes.
• Another batch ported each .zig file to equivalent .rs files in parallel across hundreds of agents, with two reviewers assigned to each file.
• A subsequent loop handled builds and tests to fix issues until the code compiled and ran error-free.
• After the port was complete, an overnight workflow eliminated redundant data copies and created pull requests for final review.

Dynamic Workflows are now available as a research preview across the Claude Code CLI, desktop app, and VS Code extension. Access is granted to Max, Team, and Enterprise plans (enabled by admins). The feature is also available on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.

Customizable Effort Level

Claude Web on Opus 4.7 only offered an adaptive thinking toggle with no manual effort adjustment. This update lets users set the effort level manually, with High as the default.

Additional API Updates

• The Messages API now supports system entries within the messages array. You can update instructions for Claude mid-task and adjust its permissions, token budgets, or contextual settings without disrupting prompt caching.
• Fast mode pricing for Opus 4.8: $10 per million input tokens and $50 per million output tokens — down from $30/$150 for Opus 4.7.

Mythos Coming Soon

The official announcement concludes with news of the upcoming launch of Mythos — Anthropic's most powerful model to date, currently available only to a limited number of partners for internal testing.

Over the past month, Mythos has helped uncover more than 10,000 critical and high-severity vulnerabilities across a range of mainstream software.

The 10,000 discovered vulnerabilities are still awaiting corresponding patches.

Anthropic is still building safety guardrails for Mythos. Models of this caliber are expected to be available to all customers in the coming weeks. Additionally, Anthropic is developing a new lineup of models that deliver performance close to Opus at a lower cost.

Availability

Claude Opus 4.8 is live across all platforms starting today, with pricing unchanged from Opus 4.7.

Official announcements:

• Claude Opus 4.8: anthropic.com/news/claude-opus-4-8
• Dynamic Workflows: claude.com/blog/introducing-dynamic-workflows-in-claude-code
• Project Glasswing / Mythos: anthropic.com/research/glasswing-initial-update

API model ID: claude-opus-4-8