Claude Sonnet 4.6: Features, Pricing, and Agentic Coding
Claude Sonnet 4.6 is Anthropic's hosted AI model for coding, agents, computer use, and long-context work. Released on February 17, 2026, it provides a 1M-token context window and up to 64K standard output tokens. It supports adaptive and extended thinking, allowing a task to spend more reasoning effort when needed.
Sonnet 4.6 remains a practical evaluation tier even though the Claude family now includes later models. The API identifier is claude-sonnet-4-6, which helps keep usage and cost tied to the intended tier.
Claude Sonnet 4.6 Overview
The model is designed for a balance of capability, speed, and cost across everyday development and agent workflows. Its published capabilities include coding, computer use, long-context reasoning, planning, knowledge work, and design. Compaction can help an agent continue longer tasks when conversation state grows.
| Trial dimension | Sonnet 4.6 setting |
|---|---|
| Model identity | Use claude-sonnet-4-6 |
| Context | Test retrieval quality, not only maximum size |
| Thinking | Keep adaptive or extended-thinking settings consistent |
| Agent loop | Measure tool success, recovery, and reviewer corrections |
SWE-bench & Coding Scores
The model system card reports 79.6 on SWE-bench Verified with adaptive thinking and maximum effort across trials. A specialized prompt produced 80.2, but that should not be presented as the default score. Both figures are vendor-reported and tied to their evaluation conditions.
For a repository test, use prompts such as:
- “Fix this failing test, add regression coverage, and explain why the patch is minimal.”
- “Refactor this module without changing its public API, then run the relevant checks.”
- “Review this pull request for correctness, security boundaries, and missing edge cases.”
Sonnet 4.6 vs Opus 4.7
Sonnet 4.6 is a balanced tier for routine development and agent work. Opus 4.7 is a newer Opus tier intended for more demanding tasks. The comparison is therefore not a same-generation matchup: it asks whether the harder task benefits enough from Opus to justify moving beyond a Sonnet baseline.
Start with Sonnet for implementation, review, and iterative tool use. Try Opus 4.7 on the same repository snapshot when complexity, planning depth, or correction effort remains high, then compare accepted-task cost.
Agentic Capabilities
Sonnet 4.6 can call tools, reason over large inputs, use computers, and maintain longer workflows with compaction. These features fit repository exploration, multi-file implementation, visual checks, and iterative debugging.
Agent reliability still depends on permissions and task design. Give the model a bounded objective, specify files or systems that must not change, require tests, and stop for approval before destructive or external actions.
Cost vs Performance
The API price is $3 per million input tokens and $15 per million output tokens. Prompt-cache writes are $3.75 per million tokens and cache reads are $0.30. Long context, repeated tool output, and high reasoning effort can materially change total task cost.
Compare cost per accepted task, not price per token alone. Include retries, failed commands, review time, and any work needed to correct an over-broad diff.
Using Claude Sonnet 4.6 in Verdent
Verdent's built-in catalog includes Claude Sonnet 4.6 and describes it as a balanced choice for everyday development. The Verdent model-pricing page shows its rates alongside the other built-in tiers.
Verdent Manager can coordinate parallel workers. Use separate task boundaries and review each result before integration.
Establish a Balanced Baseline
Run one bug fix, one feature, and one review task with fixed acceptance criteria. Use the accepted-task cost and correction effort as your baseline.
Frequently Asked Questions
How much does Claude Sonnet 4.6 cost through the API?
The listed rates are $3 per million input tokens and $15 per million output tokens, with separate prompt-cache write and read rates.
What is the Claude Sonnet 4.6 context window?
Claude Sonnet 4.6 supports a 1M-token context window and up to 64K standard output tokens.
Should I choose Sonnet 4.6 or Opus 4.7?
Start with Sonnet 4.6 for balanced coding and agent work. Test Opus 4.7 on difficult tasks where lower correction effort may justify the higher-capability tier.
Is Claude Sonnet 4.6 available in Verdent?
Yes. Verdent lists Claude Sonnet 4.6 in its built-in model catalog for balanced everyday development.