Claude Code Models: Which One to Use for Every Task in 2026

Quick Answer

Claude Code ships with three models. Default to Sonnet 4.6 for everything. Switch to Opus 4.6 when you need deep architectural reasoning or hit a problem Sonnet can't solve. Use Haiku 4.5 for simple cleanup tasks where speed matters more than intelligence.

70% of developers prefer Sonnet 4.6 over the previous flagship Sonnet 4.5. 59% prefer it over Opus 4.5. The newer Sonnet delivers what used to require the top-tier model.

Benchmark Comparison

SWE-bench Verified tests whether a model can resolve real GitHub issues across production codebases. It is the closest proxy we have for real-world coding ability. Here is how the Claude models compare, alongside competitors:

The top three frontier models (Opus, GPT-5.3-Codex, Gemini 3.1 Pro) now cluster within 1 point of each other on SWE-bench. The benchmark alone no longer differentiates them. Cost, speed, and tool integration matter more.

Where Opus does separate itself: Terminal-Bench 2.0 (65.4% vs Opus 4.5's 59.8%) and OSWorld for agentic computer use (72.7% vs 66.3%). Both measure multi-step agentic reasoning, exactly the scenario where Opus's extra inference cost pays off.

Pricing Per Model

You access Claude Code models through either a subscription plan or API key. Subscription costs are fixed monthly. API costs scale with token usage.

API Pricing (per million tokens)

Subscription Plans

The Pro plan at $20/month is the entry point. It includes Sonnet and Haiku but not Opus. To use Opus in Claude Code, you need Max ($100 or $200/month) or an API key with direct billing.

Speed Comparison

Output speed determines how long you wait for Claude Code to finish generating. In an agentic workflow where the model makes multiple tool calls per task, a 2x speed difference compounds across 10-20 steps.

Sonnet is 1.5-2x faster than Opus at standard pricing. Haiku is roughly 2x faster than Sonnet. For a 20-step agentic task where each step generates 500 tokens, the total wait time difference between Sonnet (~140s) and Opus (~250s) is nearly two minutes. That adds up across a workday.

When to Use Opus 4.6

Opus costs 67% more than Sonnet per token and runs 1.5-2x slower. It earns that premium in specific scenarios:

• Multi-file architectural refactors spanning 10+ files where the model needs to hold the entire dependency graph in working memory
• Complex debugging involving concurrency, race conditions, or subtle state management bugs that require reasoning across multiple execution paths
• First-attempt correctness on tasks where iteration is expensive (production deployments, database migrations)
• Layered architectural decisions where the model needs to surface tradeoffs before writing code (modular auth systems, backward-compatible API changes)
• Agent orchestration tasks where Opus coordinates multiple sub-agents or manages multi-step plans with branching logic

One developer put it well: "Opus catches rebuild issues, missing disposes, and async bugs that Sonnet completely skips. When you are close to shipping, Opus is worth the wait."

When to Use Sonnet 4.6

Sonnet 4.6 is the daily driver. It handles the work that used to require Opus 4.5:

• Feature development: REST endpoints, frontend components, database schemas, utility functions
• Code review and refactoring within a single module or 2-5 files
• Writing tests: unit, integration, and e2e test generation
• Bug fixes with clear reproduction steps
• Documentation generation from code
• Iterative development where you plan to review and refine the output over multiple rounds

Sonnet 4.6 is widely praised for frontend work. Multiple developers report it generates "pixel-perfect layouts" and handles state management, API integration, and component composition without needing Opus-level reasoning.

When to Use Haiku 4.5

Haiku costs 80% less than Sonnet and runs 2x faster. Its coding ability is limited (88.1% HumanEval vs Sonnet's 95%), but it excels at tasks that are well-defined and repetitive:

• Generating boilerplate code from templates
• Formatting and linting operations
• Simple file operations (rename, move, create from pattern)
• Writing test stubs with obvious assertions
• Commit message generation
• Quick explanations of code snippets

Do not use Haiku for multi-file reasoning, complex debugging, or tasks where the model needs to infer intent from ambiguous instructions. It will produce output quickly, but that output will often miss edge cases or architectural concerns that Sonnet catches.

How to Switch Models in Claude Code

Four ways to set your model, in order of priority (highest first):

1. Mid-session: /model command

# Type in the Claude Code prompt:
/model

# Or specify directly:
/model opus
/model sonnet
/model haiku

2. At launch: CLI flag

claude --model opus
claude --model sonnet
claude --model haiku

3. Environment variable

export ANTHROPIC_MODEL=claude-opus-4-6
# Or for a specific session:
ANTHROPIC_MODEL=claude-opus-4-6 claude

4. Settings file

Add "model": "claude-opus-4-6" to your Claude Code settings file for a permanent default. The /model command and CLI flag override this.

The Hybrid Workflow

The most cost-effective approach is not picking one model. It is switching between them based on task complexity. This workflow is now standard across production teams:

The opusplan alias

Claude Code includes a built-in version of this workflow. The opusplan alias uses Opus for planning and reasoning during plan mode, then automatically switches to Sonnet for code generation. Activate it with:

/model opusplan

This gives you Opus-level architectural thinking during the planning phase while keeping generation costs at Sonnet rates. Teams using this pattern report 60-80% cost savings compared to running Opus for everything, with minimal quality loss on the implementation step.

Using Non-Anthropic Models with Claude Code

Claude Code is not locked to Anthropic models. You can route it through OpenRouter or Claude Code Router to use any model.

OpenRouter setup

export ANTHROPIC_BASE_URL=https://openrouter.ai/api
export ANTHROPIC_API_KEY=your-openrouter-key

# Override specific model slots:
export ANTHROPIC_DEFAULT_SONNET_MODEL=openai/gpt-5.3-codex
export ANTHROPIC_DEFAULT_OPUS_MODEL=google/gemini-3.1-pro
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek/deepseek-v4

This gives you Claude Code's agentic workflow (tool use, file editing, terminal commands) powered by any of OpenRouter's 320+ models. Useful when a specific model handles your language or framework better, or when you want to compare providers.

Claude Code Router (open-source proxy)

For local models (Ollama, LM Studio, llama.cpp) or direct provider access without OpenRouter as intermediary, the open-source Claude Code Router project supports 8+ providers including DeepSeek, Gemini, VolcEngine, and SiliconFlow. It adds dynamic model switching with /model provider,model_name syntax and context-based routing.

What Developers Actually Say

Benchmark scores tell one story. Developer experience tells another. Here is what the community reports:

"Default to Sonnet for everything. Escalate to Opus only when you need the deepest reasoning or multi-agent coordination."

"Opus found rebuild issues, missing disposes, and async bugs that Haiku and Sonnet completely skipped. When close to release, Opus is worth the wait."

"Sonnet 4.6 generates pixel-perfect layouts. For frontend work, I never touch Opus anymore."

"The hybrid approach works. Sonnet handles 80-90% of my daily coding. Opus steps in for the hard 10%. My monthly bill dropped from $200 to about $60."

"Haiku is the sprinter, Sonnet is the steady builder, Opus is the careful reviewer. Use all three."

The consistent pattern: developers who initially used Opus for everything switched to the hybrid model once they saw Sonnet 4.6's quality. The 1.2-point SWE-bench gap does not justify a 67% cost premium for most tasks.

Frequently Asked Questions