Claude Sonnet
ModelsDefinition
The balanced Claude model tier offering the best trade-off between capability, speed, and cost for most production applications. The recommended default for new systems — handles the majority of complex tasks effectively without Opus-level expense.
Example Usage
Use Sonnet as the default model for a multi-turn customer support agent or an agentic coding assistant — it handles orchestration, tool use, and generation without the cost overhead of Opus.
Why It Matters for the CCA-F Exam
Sonnet appears in CCA-F scenario questions about production system design — it is the recommended default tier, so questions testing model selection often have Sonnet as the correct answer when the scenario describes a typical production workload without extreme latency or extreme quality constraints.
In Depth
Sonnet as the Production Default
The Sonnet tier is deliberately positioned as the recommended starting point for production applications. The current model, claude-sonnet-4-6, pairs a 1M-token context window with 64K output tokens and delivers a latency and cost profile that makes it viable for user-facing features, agentic loops, and batch workloads that would be cost-prohibitive on Claude Opus.
The "balanced" label reflects a genuine engineering trade-off: Sonnet performs near Opus on the large majority of real-world tasks, while costing significantly less and returning responses faster. The quality gap only becomes material on tasks requiring deep multi-step reasoning with many branching constraints — exactly the workload Opus is designed for.
What Sonnet Handles Well
Sonnet is the right tier for:
- Multi-turn conversational agents where response latency affects user experience.
- Code generation and review — the bulk of everyday programming tasks sit well within Sonnet's capability ceiling.
- Document summarization and extraction at scale, including lengthy contracts or research papers within the 1M-context window.
- Orchestrator role in multi-agent pipelines — the coordinator that routes sub-tasks to Claude Haiku for cheap screening and escalates to Opus only when confidence is low.
The temp/top_p Mutual-Exclusion Constraint
On claude-sonnet-4-6, at most one of temperature or top_p may be set in a single request — never both together. Sending both returns a validation error. This is a deliberate API constraint, not a bug. On Opus 4.7+ and Fable 5, both parameters are removed entirely (HTTP 400). For Sonnet 4.6, the recommended approach is to steer behavior through prompting and structured outputs rather than relying on sampling knobs as a primary quality lever. See Structured Output via Tool Use & JSON Schemas for the recommended pattern.
When to Pick Sonnet Over Opus or Haiku
Sonnet over Opus: when task complexity is moderate-to-high but not in the extreme tail — code generation, document analysis, multi-step reasoning that does not require branching over many constraints simultaneously. Opus escalation should be reserved for measurable quality failures, not used as a default safety net.
Sonnet over Haiku: when the task involves meaningful generation, nuanced judgment, or multi-turn context that Haiku's shallower capability ceiling would degrade. Haiku is for volume; Sonnet is for quality at scale.
For the full cross-tier capability table including context windows, output limits, and sampling parameter support, see Claude Opus — Full Tier Comparison.
How It Compares
| Feature | Claude Sonnet 4.6 | Claude Opus 4.8 |
|---|---|---|
| Context window | 1M tokens | 1M tokens |
| Max output tokens | 64K | 128K |
| Sampling params | At most one of temp/top_p | Removed (HTTP 400) |
| Cost | Mid | Highest |
| Best for | Production default | Complex, high-stakes reasoning |
*For the complete three-tier comparison including Claude Haiku, see Claude Opus — Full Tier Comparison.*
How It's Tested & Common Confusions
Common Question Angles
Model selection scenarios: Given a task description (e.g., "a customer support bot that handles 50K messages/day"), identify the right tier. Sonnet is almost always correct when the scenario does not call out extreme quality requirements or extreme cost sensitivity.
Why not Opus?: Expect a distractor option that is Opus. The correct reasoning is: Opus is only justified when Sonnet demonstrably underperforms on the task, not just as a safer choice.
Why not Haiku?: Haiku distractors appear when the scenario involves tasks with meaningful complexity (multi-turn reasoning, code generation, nuanced analysis). Haiku is correct only for high-volume, low-complexity work.
Parameter pitfalls: A question may describe setting both temperature and top_p on Sonnet 4.6. This is a violation — at most one is allowed. On Opus 4.7+, both are removed entirely.
Routing architecture: Scenario questions often ask which tier should serve as the orchestrator in a multi-agent pipeline. Sonnet is the typical answer — capable enough to coordinate complex workflows, cost-effective enough to run continuously.
Common Confusions
- Confusing "balanced" with "mediocre" — Sonnet's capability ceiling is high; it handles most real tasks that practitioners mistakenly attribute to Opus.
- Assuming
budget_tokensapplies to Sonnet 4.6 — it is deprecated on Sonnet 4.6 just as on Opus 4.6. - Forgetting Sonnet has 64K output, not 128K — relevant when generating very long documents.
Frequently Asked Questions
Is Claude Sonnet good enough for agentic pipelines or do I need Opus?
Sonnet is the recommended default for agentic pipelines. It handles orchestration, tool use, multi-step reasoning, and most code tasks without needing Opus-level capability. Reserve Opus for escalation paths when Sonnet's output quality falls below threshold on a specific sub-task.
What is the current Sonnet model ID and what are its output limits?
`claude-sonnet-4-6` is the current Sonnet model. It supports a 1M-token context window with a 64K maximum output — half the output ceiling of Opus 4.8 (128K). For most tasks 64K output is sufficient, but if you need to generate very long single-turn artifacts, Opus may be required.
Can I set both temperature and top_p on Sonnet 4.6?
No. On claude-sonnet-4-6 you may set at most one of temperature or top_p in a single request — never both. Sending both returns a validation error. On Opus 4.7+ and Fable 5, both parameters are removed entirely.