← Back to API models

Claude Opus 4.8

by Anthropic · Opus flagship · released May 28, 2026

Anthropic's Opus-tier flagship. Vision, code, and careful reasoning — one step below the new Fable 5.

Input $5 / 1M
Output $25 / 1M
Context 200K tokens
Released May 2026
Anthropic Console ↗ Updated June 10, 2026
§ API pricing

Per-token rates.

Input
$5/1M tokens
Prompt tokens
  • Matches GPT-5.5 on input
  • Vision inputs billed as tokens
  • Prompt caching reduces re-sent inputs
Output
$25/1M tokens
Completion tokens
  • $5 cheaper per 1M than GPT-5.5 output
  • Fast-mode output $50 / 1M (3× cheaper than 4.7)
  • Half the output rate of Fable 5
Context
200Ktokens
Window
  • Same window as Sonnet & Haiku
  • ~150K words of practical input
  • Fable 5 is the 1M-window Claude
Caching
~90%discount on cache hits
Prompt caching
  • Cached input billed at fraction of rate
  • Best when re-sending long system prompts
  • 5-minute and 1-hour TTLs supported

Why Opus 4.8 exists

Anthropic ships Opus when they want to push the quality ceiling without worrying about cost. Opus 4.8, released May 28, 2026, replaces Opus 4.7 in Claude Pro and Max for the harder tasks Sonnet 4.6 cannot consistently nail: nuanced legal and policy reading, novel research synthesis, multi-file code refactors with subtle dependencies, and any domain where the wrong answer is more expensive than another $5 of API spend.

At $5 input and $25 output per 1M tokens, Opus 4.8 holds 4.7's pricing while raising the agentic-coding score from 64.3% to 69.2%. Fast mode dropped sharply — $10/$50 per 1M (down from $30/$150 for 4.7) at roughly 2.5× standard speed. Versus GPT-5.5 ($5/$30), Opus is cheaper on output; the trade-off is context: 200K vs GPT-5.5's 1M.

Since June 9, 2026, Opus 4.8 is no longer the top of the Claude family — Fable 5 sits above it at $10/$50 per 1M with a 1M window. Opus 4.8 remains the price-performance flagship of the Opus tier, and it plays a second role in the new lineup: requests that Fable 5 declines for high-risk topics (cybersecurity, biology and chemistry, model distillation) are automatically routed to Opus 4.8.

Capabilities

Opus 4.8 is strongest at deliberate, multi-step reasoning. It writes prose with fewer tics and more structural coherence than the GPT or Gemini families, which is why it dominates serious editorial and research workflows. Anthropic reports it is roughly four times less likely to overlook code flaws than 4.7, scores 84% on Online-Mind2Web, and is the first model to clear 10% on the Legal Agent Benchmark all-pass standard. It is also a strong vision model for diagrams, charts, and screenshots.

The honest weakness: no image, video, or audio generation. Anthropic stays focused on text and analysis. If you need multimodal output, pair Opus 4.8 with another model for that step.

Typical use cases

  • Long-document analysis: contracts, papers, filings, 200-page PDFs
  • Software engineering: multi-file refactors, code review, architecture critique
  • Editorial and research writing where tone and structure matter
  • Agent loops where one wrong step costs more than a careful one
  • Vision tasks on technical diagrams, charts, and product screenshots

Sibling and rival comparison

ModelInput / 1MOutput / 1MContext
Claude Fable 5$10$501M
Claude Opus 4.8$5$25200K
Claude Sonnet 4.6$3$15200K (1M β)
Claude Haiku 4.5$1$5200K
GPT-5.5$5$301M

Within the Claude family, Sonnet 4.6 gets you most of the way at 60% the price — switch to Opus 4.8 only when Sonnet's mistakes are unacceptable, and step up to Fable 5 when even Opus isn't enough and the 2× price is justified. Versus GPT-5.5, Opus is cheaper on output but limited to a 200K window. Choose Opus for quality-critical short-to-medium context; choose GPT-5.5 or Fable 5 when you need the full 1M.

← See all Anthropic / Claude plans