Anthropic to Release Mythos-class Opus 4.8 Model

🔑 Enhanced Key Takeaways

•Anthropic's Opus 4.8, released on May 28, 2026, introduces a 'Dynamic workflows' feature in Claude Code, enabling the model to run hundreds of parallel subagents for complex coding tasks.
•The new model offers a 'fast mode' with pricing three times cheaper than previous models, making high-throughput inference more accessible for latency-sensitive production workloads.
•Opus 4.8 incorporates an 'effort control' feature on claude.ai and Cowork, allowing users to manually adjust the model's processing intensity and associated token consumption for a given task.
•Anthropic highlights Opus 4.8's improved 'honesty,' noting it is less prone to making unsupported claims and more likely to flag uncertainties in its work, enhancing reliability for users.
•Opus 4.8 serves as a precursor to the more powerful 'Mythos-class' models, which are currently restricted to Project Glasswing partners for cybersecurity applications due to their advanced capabilities and the need for stronger cyber safeguards before general release.

📊 Competitor Analysis▸ Show

Feature/Metric	Anthropic Claude Opus 4.8 (May 2026)	OpenAI GPT-5.5 (May 2026)	Google Gemini 3.1 Pro (May 2026)
Input Pricing (per 1M tokens)	$5 (regular), $10 (fast mode)	$5	$2 (Preview >200K)
Output Pricing (per 1M tokens)	$25 (regular), $50 (fast mode)	$30	$18 (Preview >200K)
SWE-bench Pro	69.2%	58.6%	N/A
SWE-bench Verified	88.6%	N/A	N/A
Terminal-Bench 2.1	74.6% (high effort)	78.2% (Terminus-2 public harness), 83.4% (Codex CLI harness)	N/A
Online-Mind2Web	84%	N/A (beats GPT-5.5)	N/A
Context Window	1M tokens	N/A (under 272K input tokens for strong competition)	N/A
Key Capabilities	Advanced coding, complex agentic workflows, high-stakes enterprise tasks, dynamic workflows, effort control, multimodal (text, image, file inputs)	Strong on terminal/CLI workflows, web browsing, graduate-level science, native multimodal understanding (text, images, audio)	Enhanced reliability, dynamic workflow capabilities, user-adjustable 'engagement levels'

🛠️ Technical Deep Dive

Context Window & Output: Claude Opus 4.8 features a 1 million-token context window and supports a maximum output of 128,000 tokens.
Input Modalities: The model accepts text, image, and file inputs, producing text outputs.
Dynamic Workflows: A new research preview feature in Claude Code allows the model to plan and execute tasks by running hundreds of parallel subagents in a single session.
Effort Control: Users can adjust the model's 'effort' level (Low, Medium, High, Max, or adaptive thinking) on claude.ai and Cowork, influencing processing intensity and token usage.
Constitutional AI: Anthropic continues to train its models, including Opus 4.8, using 'constitutional AI,' a technique designed to improve ethical and legal compliance by guiding AI behavior with predefined principles.
Behavioral Improvements: Opus 4.8 demonstrates enhanced honesty, being less likely to make unsupported claims and more prone to flagging uncertainties, and shows improved self-correction in coding tasks.
API Enhancements: The Messages API now accepts system entries mid-conversation, allowing developers to update Claude's instructions without breaking the prompt cache or incurring higher costs.
Refusal Stop Details: The stop_details object on refusal responses is now publicly documented, providing categories of refusal to help applications route users appropriately.

🔮 Future ImplicationsAI analysis grounded in cited sources

The general release of 'Mythos-class' models will significantly escalate the AI capabilities available to a broader audience.

Mythos-class models, currently restricted to Project Glasswing, have demonstrated unprecedented cybersecurity capabilities, finding thousands of zero-day vulnerabilities, and their public release implies a new frontier in AI application.

Anthropic's focus on 'effort control' and 'fast mode' pricing indicates a strategic move towards optimizing AI cost-efficiency and performance for diverse enterprise workloads.

These features allow users to balance intelligence, speed, and cost, making advanced AI more adaptable and economically viable for various business applications.

The introduction of 'Dynamic workflows' in Claude Code suggests a future where AI agents can autonomously manage and execute highly complex, large-scale software engineering projects.

This feature enables Claude to plan and run hundreds of parallel subagents, indicating a significant step towards more sophisticated and independent AI agentic capabilities in development.

⏳ Timeline

2021-01

Anthropic is founded by former OpenAI executives.

2023-03

Claude, Anthropic's first AI-based chatbot, is released.

2024-03

Claude 3 model family (Haiku, Sonnet, Opus) is launched, introducing multimodal image support.

2024-06

Claude 3.5 Sonnet is released, outperforming Claude 3 Opus and introducing the Artifacts feature.

2025-05

Claude Opus 4 and Claude Sonnet 4 are released, with Opus 4 classified as a 'Level 3' safety risk.

2026-05

Claude Opus 4.8 is released, with 'Dynamic workflows' and 'effort control,' and teases the upcoming general release of Mythos-class models.

Anthropic to Release Mythos-class Opus 4.8 Model

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (20)

👉Related Updates

Anthropic Launches Claude Opus 5 with Enhanced Performance

Controversial AI-generated store signage sparks design debate