Anthropic Claude Opus 및 Sonnet 계층

Claude is the family of large language models built by Anthropic, a US AI company. Instead of shipping one model per generation, Anthropic ships several sizes under poetic names. Haiku is the smallest and fastest, Sonnet sits in the middle, and Opus is the largest of the original three. A tier name describes a position in the lineup, not a fixed level of ability: a new Sonnet can beat an older Opus, and every tier changes with each generation. The version number, such as 3.5, 4.5 or 5.5, tells you the generation, and the tier name tells you the size and price point within it.

The naming scheme arrived with the Claude 3 family on March 4, 2024. Anthropic's announcement ordered the three models by capability and gave each a role: Haiku for near-instant responses, Sonnet for balancing intelligence and speed in enterprise workloads, and Opus for the most complex tasks. Launch prices made the gap concrete. Opus cost $15 per million input tokens and $75 per million output tokens, Sonnet $3 and $15, and Haiku $0.25 and $1.25, all with a 200,000-token context window. A token is a chunk of text, roughly three quarters of an English word, and providers bill by the million.

The tiers quickly showed that names do not lock in rank. On June 21, 2024, Anthropic released Claude 3.5 Sonnet at the mid-tier price of $3 and $15 and reported that it ran twice as fast as Claude 3 Opus while outperforming it on a range of evaluations. In an internal agentic coding test it solved 64% of problems against 38% for Claude 3 Opus. Claude 4, released May 22, 2025, kept the Opus price at $15 and $75 and reported SWE-bench Verified scores (a benchmark of real GitHub issues) of 72.5% for Opus 4 and 72.7% for Sonnet 4 without extended thinking. The two tiers were close on that benchmark even though Opus cost five times more. Claude Opus 4.5, published November 24, 2025, cut the Opus price to $5 and $25 and introduced an effort parameter.

The effort parameter changed how the tiers are used. It lets a developer ask one model to think less, for speed and cost, or more, for harder problems. Anthropic's current model-selection guide says tuning effort is often a better lever than switching models. Thinking itself also changed: newer models use adaptive thinking, where the model decides how much internal reasoning to do, steered by effort, rather than a fixed thinking budget set by the caller. On the current lineup, Opus 5.5 and Haiku 5.5 default to medium effort and Sonnet 5.5 and Fable 5.1 default to high.

As of October 2026, Anthropic's models overview lists four generally available current models. Claude Fable 5.1 is described as for demanding reasoning and long-horizon agentic work, at $10 input and $50 output per million tokens. Claude Opus 5.5 is for long-running agentic coding and knowledge work at $4 and $20. Claude Sonnet 5.5 is called the best combination of speed and intelligence at $2 and $10. Claude Haiku 5.5 targets high-volume, latency-sensitive tasks such as classification, extraction and routing, from $0.10 and $0.50. All four accept text and image input, support tool use, have a 1 million token context window and up to 128,000 output tokens, and list a reliable knowledge cutoff of June 2026. A separate model, Claude Mythos 5.1, offers the same capabilities as Fable 5.1 only to organizations verified through Anthropic programs such as its Cyber Verification Program.

Picking a tier is an engineering decision, not a brand preference. Anthropic's guide describes two starting strategies. Efficiency-first starts with Haiku, tests thoroughly, and upgrades only where a capability gap appears; it suits prototypes, tight latency budgets and high-volume simple work. Capability-first starts with Opus, tunes prompts, then lowers effort or moves work to cheaper models once the workflow is understood; it suits complex reasoning, science and high-autonomy agents. Either way, the guide says a task-specific evaluation set is the most important step. Many production systems combine tiers: a cheap model handles most calls and escalates hard decisions to a frontier model.

The tier system has real limitations. Names are relative, so 'Opus' in a 2024 article and 'Opus' in 2026 are different models with different prices and behavior. Benchmarks published by the vendor use the vendor's settings, and close scores on one benchmark, as with Opus 4 and Sonnet 4 on SWE-bench Verified, can hide large differences on other tasks. Higher tiers are slower; the overview labels Haiku fastest and Fable slower. Model behavior also shifts between versions, so prompts tuned for one model may need retuning, which is why Anthropic publishes migration guides with breaking changes.

The practical costs go beyond the per-token price. Batch requests are 50% cheaper, and prompt cache reads cost a fraction of base input, 5% on Opus 5.5 and Sonnet 5.5. Models also have lifecycles: Anthropic labels them active, legacy, deprecated or retired, gives at least 60 days' notice before retiring publicly released models, and lists each retirement on its deprecations page. Claude 3 Opus, for example, was deprecated June 30, 2025 and retired January 5, 2026. Teams that pin a model ID should track those dates, keep their evaluation suite ready, and budget time to retest when they move to a new tier or generation.