Skip to main content
All models

Anthropic: Claude Sonnet 4.6 Coding Benchmark

Sonnet 4.6 is Anthropic's most capable Sonnet-class model yet, with frontier performance across coding, agents, and professional work. It excels at iterative development, complex codebase navigation, end-to-end project management with...

Code Agent#67Kilo rank
Context1,000,000tokens
Max Output128,000tokens
Inputmodalities
Price$3.00/1M input
Ranks from Kilo Code LeaderboardPricing via OpenRouter

Try Anthropic: Claude Sonnet 4.6 in Kilo Code

Experience this model with the most popular open source coding agent. Free to start, pay only for AI usage. Use in popular IDEs like VS Code, JetBrains, command line, or cloud agents.

5M+

Downloads

500+

models supported

Free

to Start

Get Started with Kilo

Access 500+ models including Anthropic: Claude Sonnet 4.6 and many more in Kilo Code

Coding Performance

Coding benchmarks and performance metrics for development tasks

Kilo Bench

% Completion on Terminal Bench 2.0
55.1%
Cost per attempt (USD)
$53.37
Benchmark
Terminal Bench 2.0

Official Kilo eval results. Cost is averaged per complete benchmark attempt.

Security

Enkrypt AI red-team scores for Anthropic: Claude Sonnet 4.6. Each value is the share of successful attacks on a 0–100 scale — lower is safer.

Overall risk
32.5

Lower is safer

NIST

32.0/ 100

Average attack success across NIST-mapped tests: bias, harm, toxicity, CBRN, and insecure code.

OWASP

39.0/ 100

Weighted average of the same tests using OWASP Top 10 for LLMs 2025 risk rankings.

Attack categories

Percentage of successful attacks in each Enkrypt red-team category.

Bias81.1/ 100

Share of tests that elicited biased responses.

Harmful content0.6/ 100

Share of tests that produced dangerous, violent, or hateful content.

CBRN15.3/ 100

Share of tests that elicited chemical, biological, radiological, or nuclear assistance.

Insecure code32.9/ 100

Share of tests that produced vulnerable or malicious code.

Score
Value
Overall risk
32.5 / 100
NIST
32.0 / 100
OWASP
39.0 / 100
Bias
81.1 / 100
Harmful content
0.6 / 100
CBRN
15.3 / 100
Insecure code
32.9 / 100

Security scores from the Enkrypt AI Safety Leaderboard · Last checked Oct 19, 2026

Real-World Usage

Real-world usage statistics from the Kilo Code community

Weekly Token Usage

Agent Rankings (Last Week)

Where this model ranks by agent role, including historical Orchestrator usage

Code

Write, modify, and refactor code

#67

Ask

Get answers and explanations

#28

Debug

Diagnose and fix software issues

#16

Orchestrator (legacy)

Legacy coordination role, now handled by native sub-agent delegation

No data

Real-world metrics from the Kilo Code Leaderboard

OpenClaw Benchmarks

PinchBench measures how Anthropic: Claude Sonnet 4.6 performs on real OpenClaw agent tasks: multi-step execution, tool use, recovery, latency, and cost.

PinchBench run

Average score

69.6%

#42 of 50 official models

Average time

159m 49s

47 runs · per OpenClaw task

Average cost

$15.390

Per benchmark run

Category breakdown

Best verified PinchBench v2 run by OpenClaw task family.

Basic100.0% · 1/1 cleared
Calendar100.0% · 1/1 cleared
Coding100.0% · 1/1 cleared
File Ops100.0% · 3/3 cleared

Top task results

Highest-scoring benchmark tasks from the same submission.

Calendar
Calendar Event Creation
100.0%
File Ops
Create Project Structure
100.0%
File Ops
File Structure Creation
100.0%
Writing
Professional Email Drafting
100.0%
Basic
Sanity Check
100.0%
File Ops
Search and Replace in Files
100.0%

Autonomous task execution

Anthropic: Claude Sonnet 4.6 shows emerging average success across OpenClaw-style benchmark runs, useful for recurring research, browser, and file-based automations.

Tool use and recovery

PinchBench tasks stress multi-step planning, tool calls, and judge-verified completion rather than single prompt coding snippets.

Agent workflow fit

Its deliberate average runtime and premium run cost help set expectations for long-running agents and production workflows.

Agentic benchmarks from the PinchBench Leaderboard

Pricing

Cost per 1 million tokens

Input Tokens
$3.00
per 1M tokens
Output Tokens
$15.00
per 1M tokens

Example Cost

Analyzing a 10,000 line codebase (≈40k input tokens, 10k output tokens) costs approximately $0.2700

Coding Capabilities

Features and parameters relevant to coding tasks

Coding Features

Function Calling
Can call external functions/APIs
Tool Choice
Control over function selection
Structured Outputs
JSON schema validation
Reasoning Tokens
Extended thinking for complex problems

Pricing details from OpenRouter

Technical Details

Architecture and implementation specifications

Specifications
Model ID
anthropic/claude-sonnet-4.6
Created
February 17, 2026
Tokenizer
Claude
Input Modalities
TextImageFile
Context Window
1,000,000 tokens
Max Completion Tokens
128,000 tokens
Input Price
$3.00 per 1M tokens
Output Price
$15.00 per 1M tokens
Cache Read Price
$0.30 per 1M tokens
Cache Write Price
$3.75 per 1M tokens
Content Moderation
Enabled

Ready to try Anthropic: Claude Sonnet 4.6?

Install Kilo Code and start using Anthropic: Claude Sonnet 4.6 for your coding projects today. Choose from 500+ AI models with complete freedom.

  1. Install Kilo Code

    Get the extension from VS Code Marketplace, JetBrains Plugin Repository, or the CLI.

  2. Open the model selector

    Click the model name in the Kilo Code chat panel to open the selector.

  3. Choose your model

    Search or browse to find and select your preferred model.

  4. Start coding

    Use the Code, Ask, Debug, or Plan agent — the model is ready immediately.

Get Started with Kilo
Talk to enterprise sales