Claude Models Explained: Sonnet, Opus, Haiku and Every Version Compared (2026)

Anthropic names its Claude models after poetry forms: Haiku (short, quick), Sonnet (structured, balanced), and Opus (large-scale, complex). The names aren't just branding. Each tier reflects a real tradeoff between intelligence, speed, and cost.

If you've opened Claude's model picker and wondered which one to use, or seen mentions of "Opus 4.6" and "Sonnet 4.5" and weren't sure what changed, this guide covers every tier, every major version, and which one fits your work.

The Three Claude Model Tiers

Anthropic maintains three concurrent product lines, each updated on its own schedule:

Haiku (anthropic.com/claude/haiku) is the speed tier. It's designed for high-throughput tasks where latency matters more than deep reasoning. Think classification, simple extraction, quick chat. It's the cheapest model to run and the fastest to respond.

Sonnet (anthropic.com/claude/sonnet) is the balance tier, the workhorse. It handles coding, writing, analysis, and agent tasks at a price point that makes it practical for daily use and production deployments. Sonnet is the default model for free and Pro users on claude.ai.

Opus (anthropic.com/claude/opus) is the flagship tier. It's built for the hardest problems: multi-hour agent workflows, complex codebase reasoning, legal analysis, and tasks where accuracy matters more than speed or cost. Opus costs significantly more per token.

In mid-2026, Anthropic introduced a fourth tier above Opus: Fable and Mythos. Fable 5 is publicly available via API and represents their strongest general-purpose model. Mythos 5 is at a similar capability level but only available to select partners through the Project Glasswing security research program.

Full Version History: Claude 3 to Claude 5

The Claude model family has evolved rapidly. Here's the timeline of every major release:

Claude 3 era (March 2024):

Claude 3 Opus, Claude 3 Sonnet, Claude 3 Haiku
Context window: 200K tokens
First multi-tier release. Opus 3 established Anthropic as a serious competitor to GPT-4.

Claude 3.5 (June–October 2024):

Claude 3.5 Sonnet (June): massive performance jump, outperformed Claude 3 Opus at lower cost
Claude 3.5 Haiku (October): brought Haiku speed with near-Sonnet intelligence

Claude 3.7 (February 2025):

Claude 3.7 Sonnet: introduced Extended Thinking mode, the first "hybrid reasoning" Claude model

Claude 4 (May 2025):

Claude Opus 4 and Claude Sonnet 4
Both are hybrid reasoning models with extended thinking, tool use, and computer use capabilities
Opus 4 could sustain agent tasks for up to 7 hours. Context window: 200K for Sonnet 4, up to 1M for Opus 4.

Claude 4.x (August–November 2025):

Claude Opus 4.1 (August): improved coding benchmarks (SWE-bench 74.5%)
Claude Opus 4.5 and Claude Sonnet 4.5 (November): Opus 4.5 became the strongest coding model, beating human candidates on internal take-home tests

Claude 4.6 (February 2026):

Claude Opus 4.6: 1M token context window (beta), Terminal-Bench 2.0 industry high, Humanity's Last Exam leader
Claude Sonnet 4.6: near-Opus performance at 40% lower price. SWE-bench 79.6% vs Opus's 80.8%. Became the default model for free/Pro users.

Claude 4.7–4.8 (April–May 2026):

Claude Opus 4.7: new tokenizer, high-res vision (2576px), enhanced coding
Claude Opus 4.8: dynamic workflow features, improved reliability

Claude 5 (June–July 2026):

Claude Fable 5 and Mythos 5 (June 9): SWE-bench ~95%, Artificial Analysis IQ index leader
Claude Sonnet 5 "Fennec" (July 1): default model for all free/Pro users, SWE-bench Pro 63.2%, OSWorld 81.2%
Claude Opus 5 (July 24): AI IQ leader at 61 on Artificial Analysis, 80% fewer system prompt requirements

Claude Sonnet: The Workhorse

Sonnet is the model most people use most of the time. It's fast enough for real-time chat, smart enough for serious coding, and priced for production workloads.

The latest Claude Sonnet 5 (July 2026, codename "Fennec") became the default model for all free and Pro users immediately on launch. It scored 63.2% on SWE-bench Pro and 81.2% on OSWorld-Verified, the computer-use benchmark, putting it close to the Opus flagship tier for practical tasks.

Sonnet 4.6 (February 2026) was a turning point. Before it, there was a meaningful gap between Sonnet and Opus. After 4.6, that gap narrowed to the point where most users couldn't tell the difference for everyday work. On OfficeQA (enterprise document understanding), Sonnet 4.6 matched Opus 4.6. On GDPval-AA (economic knowledge work), Sonnet 4.6 scored 1633 vs Opus's 1606.

Sonnet pricing as of mid-2026: $3 per million input tokens, $15 per million output tokens. The newer Claude Sonnet 5 likely follows similar pricing.

Claude Opus: The Flagship

Opus is Anthropic's most capable tier, built for work where getting it right matters more than speed or cost. The kind of tasks where you'd pay an expert human hours to complete, and expect the AI to match or exceed that level.

Claude Opus 4.6 (February 2026) was the first Opus with a 1M token beta context window. It set industry records on Terminal-Bench 2.0 (agentic coding) and Humanity's Last Exam (cross-disciplinary reasoning), and scored 144 Elo points above GPT-5.2 on GDPval-AA.

Claude Opus 5 (July 2026) took a different direction. Rather than pushing benchmarking scores higher, Anthropic focused on reliability and safety. The most notable change: Opus 5 needs roughly 80% fewer guardrails in its system prompt than previous Opus versions. The model's own judgment handles edge cases that previously required explicit rules.

Opus pricing: $5 per million input tokens, $25 per million output tokens. Long-context pricing (>200K tokens) is higher at $10/$37.50.

Claude Haiku: Speed and Economy

Haiku is the model you use when speed or cost is the primary constraint. It handles classification, entity extraction, simple Q&A, and high-volume API workloads where you're processing thousands of items.

Claude Haiku 4.5 is the most recent version. It has a 200K token context window and supports 64K output tokens. Pricing is roughly $0.80 per million input tokens, $4 per million output tokens, making it about 6x cheaper than Opus for input.

Haiku isn't meant for complex reasoning or coding. If you ask it to debug a multi-file codebase, it'll try but likely miss context. For categorizing 10,000 support tickets or extracting names and dates from documents, it's exactly the right tool.

How to Choose a Claude Model

The choice usually comes down to three questions:

What's the task's complexity? Simple classification or extraction → Haiku. Everyday coding, writing, or analysis → Sonnet. Multi-step agent workflows, legal reasoning, or deep research → Opus or Fable.
How fast does it need to be? Haiku responds near-instantly. Sonnet is fast enough for chat. Opus takes noticeably longer on complex tasks, especially with extended thinking enabled.
What's your budget? At production scale, the cost gap is real. Processing 10 million input tokens with Haiku costs roughly $8; with Sonnet, $30; with Opus, $50. For a high-throughput pipeline, those differences add up fast.

A practical approach: develop and test with Sonnet, then switch to Opus for production runs of your hardest tasks. Many teams run Haiku for classification and routing, Sonnet for generation, and Opus for the 5% of queries that need deep reasoning.