Claude Opus 4.8

Spec

Attribute	Value
Developer	anthropic
Released	2026-05-28
Announced	2026-05-28
Context window	1M tokens
Pricing	same as Claude Opus 4.7
License	unknown
Availability	Claude API, Amazon Bedrock, Vertex AI
Max output	128k tokens
Adaptive thinking	yes

Release Date

2026-05-28 — released alongside the Anthropic $65B Series H funding announcement.

Benchmarks

Benchmark	Opus 4.8	Opus 4.7	GPT-5.5
SWE-bench Pro (agentic coding)	69.2%	64.3%	58.6%
SWE-bench Verified	88.6%	87.6%	—
USAMO 2026 (math)	96.7%	69.3%	—
GraphWalks long-context F1 (1M tokens)	68.1%	40.3%	—
Notable: the USAMO 2026 math improvement (+27.4 pp) is the largest single-generation jump in high-school olympiad math to date.

Key Improvements Over Opus 4.7

Honesty & Self-Correction (major focus)

• Uncritically reporting flawed results: 0% (first Claude model to achieve zero)
• Fails to raise important events to user: 3.7% (down significantly)
• Overconfidence: 10x reduction vs Opus 4.7
• More likely to flag uncertainties and less likely to make unsupported claims

Agentic Coding

• 4x less likely to miss flaws in code it produces
• Tool calling "meaningfully more efficient, fewer steps for the same intelligence" (CursorBench)
• Long-context performance: GraphWalks F1 at 1M tokens: 68.1% (was 40.3%)

Dynamic Workflows (New Feature — Research Preview)

Released simultaneously with Opus 4.8 for Claude Code:

• Orchestrates up to 1,000 parallel subagents in a single session
• Claude writes a JavaScript orchestration script; the runtime executes it in the background
• Session stays responsive while agents work
• Codebase-scale migrations: hundreds of thousands of lines
• Real-world case: Bun ported from Zig to Rust — 750,000 lines of Rust, 99.8% test pass, 11 days total
• Availability: Claude Code Enterprise, Team, and Max plans; requires Claude Code v2.1.154+
• Platforms: CLI, Desktop app, VS Code extension

Use Cases

• Agentic software engineering at project scale
• Long-document reasoning and Q&A (1M context)
• Math and scientific reasoning (USAMO 96.7%)
• Honest, self-correcting AI collaboration (near-zero hallucination in flagged outputs)

Compared To

Model	Org	SWE-bench Pro	Notes
Claude Opus 4.8	Anthropic	69.2%	Dynamic Workflows, 1M context
Claude Opus 4.7	Anthropic	64.3%	Previous frontier
Claude Mythos Preview	Anthropic	(WITHHELD)	GPQA 94.6%, SWE-bench V 93.9%; cybersecurity risk — no public release
GPT-5.5	OpenAI	58.6%	—

Sources

• Anthropic — Introducing Claude Opus 4.8 (source)
• TechCrunch — Dynamic Workflow tool
• The New Stack — Full breakdown
• MarkTechPost — Dynamic Workflows capped at 1,000 subagents