benchr
Anthropic positioned Sonnet 4.6 as a balance of speed and intelligence. That is provider positioning, not evidence that it is the right tier for every deployment. Use it to define a test candidate, then measure task success, review effort, latency, and total cost.
The Claude model documentation lists a 1M-token context window, up to 64k output tokens, vision, function calling, structured output, prompt caching, and thinking features. Feature availability does not establish equivalent quality to Opus; test the behaviors your application depends on.
The decision is which tasks meet your acceptance criteria on Sonnet, which can use Haiku at a lower rate, and which show a measured benefit from Opus. Treat routing as an evaluation result, not a preset hierarchy.
When Sonnet is a candidate
Chat, drafting, routine coding, tool calling, and structured extraction are reasonable Sonnet test categories. Do not assume they are solved reliably: use fixed prompts and schemas, score accepted outputs and errors, and compare the same cases with another eligible tier.
The reason to default here rather than to Opus is pure economics. Sonnet's input is 60% of Opus's and its output is 60% as well, so every call you route down instead of up saves you real money at scale. The common mistake is defaulting to the flagship and barely touching the middle tier, then wondering why the bill climbs. The price-per-use-case breakdown shows how fast that adds up across chat, RAG, and agent workloads.
| Workload | Best tier | Why |
|---|---|---|
| Bulk classification, request routing, field extraction | Haiku 4.5 | Simple, high-volume; Sonnet's reasoning goes unused |
| Chat, drafts, routine code, tool use, most production traffic | Sonnet 4.6 | The balanced default: smart enough, priced for volume |
| Complex production code, long-horizon agents, dense docs, hard reasoning | Opus 4.8 | Worth the premium only where a mistake is costly to undo |
When to drop down to Haiku
Plenty of workloads never need what Sonnet brings. If you're tagging support tickets, routing requests to the right queue, pulling fields out of a document, or summarizing in bulk, the reasoning you'd pay Sonnet for sits idle. That's Haiku 4.5 territory, at a third of Sonnet's price.
The trick is to be honest about how many of your tasks are simple. A lot of "AI features" are classification wearing a trench coat, and they run fine on the cheap tier. The Haiku 4.5 review walks through where the cheapest model holds up and where the quality drop starts to cost you more than the savings.
When to climb to Opus
Complex multi-file bugs, long agent runs, and dense document analysis are useful escalation tests. Anthropic reports benchmark and code-review-behavior gains for Opus 4.8, but those provider claims do not show that it earns the roughly 1.67× input premium on your workload.
Escalate to the expensive model one task at a time, not one stack at a time.
Per-task escalation is one pattern to test: use a baseline candidate, identify failed task classes, and evaluate a higher tier on those cases. The routing mix should come from measured outcomes rather than a universal split.
The price math, plainly
Here's the shape of it. Sonnet output costs five times its input, the same ratio Anthropic runs on every tier, so the fastest way to cut a Sonnet bill is to shorten responses, not inputs. Cached input drops to about a tenth of the standard rate, and the Batch API takes 50% off both directions for anything you don't need answered in real time. Stack caching and batching on a repetitive workload and the effective rate falls hard.
For the levers that cut a token bill regardless of tier, benchr's guide to cutting your token spend goes deeper on caching, batching, and routing. The headline for Sonnet specifically: it's the tier where those savings compound, because it's the one you should be running most of your traffic through.
The verdict
Claude Sonnet 4.6 is a candidate baseline, not a guaranteed default. Compare it with Haiku and Opus on representative tasks, then route only where measured acceptance rate, review time, latency, and total cost support the decision.
A mixed-tier strategy may reduce cost, but equivalent output quality must be demonstrated on the same versioned workload.
Frequently asked
What does Claude Sonnet 4.6 cost?
Per Anthropic's pricing page, $3 per million input tokens and $15 per million output. Cached input drops to roughly a tenth of the input rate, and the Batch API takes 50% off both directions for asynchronous jobs.
Is Claude Sonnet 4.6 good enough for production?
It is a production candidate, not a guaranteed fit. Anthropic lists coding, tool use, a 1M-token context window, vision, structured output, and thinking features. Validate those behaviors and compare task success, errors, latency, and cost with another eligible tier.
When should you use Haiku 4.5 instead of Sonnet 4.6?
Haiku 4.5 is a lower-priced candidate for tagging, routing, and extraction. Its price is one third of Sonnet's, but do not assume equal quality; compare schema accuracy, false routing, latency, and review cost on matched tasks.
When is Opus 4.8 worth the jump from Sonnet 4.6?
Test Opus when mistakes are costly or Sonnet misses your acceptance threshold. Its input rate is about 1.67× Sonnet's, so pay the premium only where a matched evaluation shows enough improvement to offset it.
What is the context window for Claude Sonnet 4.6?
1M tokens, the same ceiling as Opus 4.8, with up to 64k output tokens per response on the standard API. Treat the long window as a retrieval surface for pulling facts back out rather than a one-shot summarization buffer.
Changelog
- July 24, 2026 — Replaced universal-default, tier-quality, and guaranteed-routing claims with provider attribution and matched-workload evaluation guidance; synchronized FAQ copy.
- May 30, 2026 — Originally published. Pricing, context window, output limit, and feature support verified against Anthropic's pricing page and the Claude model documentation.
References
- Anthropic, "Models overview," platform.claude.com, accessed May 2026.
- Anthropic, "Pricing," platform.claude.com, accessed May 2026.
- Anthropic, "Claude Pricing," claude.com/pricing, accessed May 2026.
- Anthropic, "Prompt caching," platform.claude.com, accessed May 2026.