Claude Model Comparison & Selection Guide — Opus 4.8, Sonnet 4.6, Haiku 4.5, Fable 5

Compare current Claude API models (Opus 4.8, Sonnet 4.6, Haiku 4.5, Fable 5) by price, context, max output, thinking, and knowledge cutoff, with selection criteria by task type — based on official docs as of June 2026.

The Claude API offers several models, and which one you choose drives both cost and quality. Based on the official Models overview (as of June 2026), this article compares the current models' specs and pricing and lays out selection criteria by task type.

Claude 모델 비교 (2026-06 기준) 가격 = MTok당 USD · 공식 Models overview 기준 · 정책은 변동될 수 있음 Fable 5 claude-fable-5 가격(입력/출력) $10 / $50 컨텍스트 1M 최대 출력 128k 최상위 난이도 워크로드 Opus 4.8 claude-opus-4-8 가격(입력/출력) $5 / $25 컨텍스트 1M 최대 출력 128k 복잡 추론·에이전트 코딩 Sonnet 4.6 claude-sonnet-4-6 가격(입력/출력) $3 / $15 컨텍스트 1M 최대 출력 64k 속도·지능 균형 기본값 Haiku 4.5 claude-haiku-4-5 가격(입력/출력) $1 / $5 컨텍스트 200k 최대 출력 64k 최고속·저비용 대량처리

Current model lineup

Per the official docs, the main current models are below. All current Claude models support text and image input, text output, multilingual, and vision, and are available via the Claude API, AWS, Bedrock, Vertex AI, and Microsoft Foundry.

  • Claude Opus 4.8 (claude-opus-4-8) — Anthropic's most capable Opus-tier model, for complex reasoning and long-horizon agentic coding. Pricing $5 input / $25 output per MTok, 1M context, 128k max output. Supports the effort parameter (defaults to high).
  • Claude Sonnet 4.6 (claude-sonnet-4-6) — the best balance of speed and intelligence. $3 / $15, 1M context, 64k max output. Supports both extended and adaptive thinking.
  • Claude Haiku 4.5 (claude-haiku-4-5-20251001, alias claude-haiku-4-5) — the fastest model with near-frontier intelligence. $1 / $5, 200k context, 64k max output.
  • Claude Fable 5 (claude-fable-5) — Anthropic's most capable widely released model, for the most demanding reasoning and long-horizon agentic work. $10 / $50, 1M context, 128k max output. (Generally available from June 9, 2026.)

Note: Claude Mythos 5 / Mythos Preview are offered separately via Project Glasswing (invitation-only, defensive cybersecurity workloads) with no self-serve signup.

Pricing & spec comparison

Cost is priced per million tokens (MTok) of input/output. For the same task, unit price varies widely: Haiku $1/$5 → Sonnet $3/$15 → Opus $5/$25 → Fable $10/$50. Output is pricier than input, so prompt design that reduces output length directly affects cost.

  • Context window: Opus 4.8, Sonnet 4.6, Fable 5 = 1M tokens; Haiku 4.5 = 200k. (On Microsoft Foundry, Opus 4.8 is 200k.)
  • Max output: Opus 4.8 & Fable 5 = 128k; Sonnet 4.6 & Haiku 4.5 = 64k. (Synchronous Messages API; on the Batch API some models support up to 300k output via a beta header.)
  • Thinking: Sonnet 4.6 & Haiku 4.5 support extended thinking; Opus 4.8 & Fable 5 use adaptive thinking. (Haiku 4.5 has no adaptive; Fable 5 has adaptive always on.)
  • Knowledge cutoff: Opus 4.8 Jan 2026, Sonnet 4.6 Aug 2025, Haiku 4.5 Feb 2025.
어떤 모델을 고를까? 공식 권고: 잘 모르겠으면 복잡한 작업은 Opus 4.8로 시작 작업 난이도는? 비용 vs 품질 트레이드오프 대량·단순·저지연 → Haiku 4.5 일상 프로덕션 기본 → Sonnet 4.6 복잡 추론·에이전트 → Opus 4.8 최고난도 프런티어 → Fable 5 고려 포인트 · 긴 문서·코드베이스 → 1M 컨텍스트 (Opus·Sonnet·Fable) · thinking 필요 → Sonnet·Haiku (extended) / Opus·Fable(adaptive) · 워크플로 중 난이도 변동 → 단계별로 모델 라우팅 최신 스펙은 공식 문서로 확인

Choosing by task type

The official docs advise "if unsure, start with Opus 4.8 for complex tasks." In practice, a cost/quality trade-off lens works well:

  • High-volume, simple, low-latency (classification, short replies, real-time chatbots) → Haiku 4.5. Fastest and cheapest.
  • Everyday production default (most coding, generation, summarization) → Sonnet 4.6. Balanced speed/intelligence/cost.
  • Complex reasoning & long-horizon agentic codingOpus 4.8. When answer quality matters most.
  • Frontier-difficulty workFable 5. Only when the extra cost is justified.

For long documents/codebases, the 1M context (Opus, Sonnet, Fable) helps; for workflows where difficulty shifts step to step, routing different models per step is a valid approach.

Model IDs are pinned snapshots

A caveat: every Claude model ID is a pinned snapshot. IDs with a date (e.g., 20250929) are fixed to that release, and starting with the 4.6 generation the dateless format (e.g., claude-sonnet-4-6) is also a pinned snapshot, not an evergreen pointer. You can query model capabilities and token limits programmatically via the Models API (returns max_input_tokens, max_tokens, and a capabilities object).

Summary

A sensible model-selection flow: (1) gauge task difficulty and cost tolerance → (2) pick a tier among Haiku/Sonnet/Opus/Fable → (3) fine-tune by context, thinking, and output needs. To learn more, see Claude API Model IDs and Versioning and Context Window and Tokens.

This article summarizes public information from the official Anthropic docs (platform.claude.com/docs) as of June 2026. The model lineup, pricing, and specs change frequently, so always verify against the official Models overview and Pricing pages when choosing. This site is not an official Anthropic site.

Keep reading

Have a question or want to share how you use Claude?

Join the community to share tips with other users, or explore more guides.