Claude Sonnet 4.6
Language Available Full comparison ↗ anthropic/claude-sonnet-4-6 · by Anthropic
· decoder-only-transformer
Pricing — 1 offering(s)
Input tokens
- $3.00 / 1M tokens (input) Current 2026-06-12 → present
Output tokens
- $15.00 / 1M tokens (output) Current 2026-06-12 → present
Cache hits and refreshes
- $0.30 / 1M tokens (cached input) Current 2026-06-12 → present
5-minute cache writes
- $3.75 / 1M tokens (cache write) Current 2026-06-12 → present
Batch input tokens
- $1.50 / 1M tokens (input) Current 2026-06-12 → present
Batch output tokens
- $7.50 / 1M tokens (output) Current 2026-06-12 → present
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Claude Sonnet 4.6 fits into a cost-aware routing setup
See how →Capability profile
How Claude Sonnet 4.6 rates across core capability dimensions, with the task-level evidence behind each rating.
SWE-bench Verified (500 human-verified GitHub issues; patch must pass the repo's own tests; no file-location hints): resolves 79.6% — fails roughly 1 issue in 5. Vendor-reported (Anthropic system card, averaged over 10 trials at max effort).
1M-token context window at standard pricing — confirmed live against Anthropic's pricing docs (2026-07-20): 'Claude Fable 5, ... Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6 include the full 1M token context window at standard pricing.' (Sonnet 4, the previous generation, was 200K.)
Coding benchmarks
View the full coding benchmark leaderboard →| Benchmark | Score | Harness | Source |
|---|---|---|---|
| swe-bench-verified ↗ | 79.6% 2026-02 | agentless 2.0 | 🏢 vendor ↗ |
| swe-bench-pro | 58.1% 2026-06 | agentless | 🏢 vendor ↗ |
| terminal-bench-2-1 | 67.0% 2026-06 | mini-swe-agent | 🏢 vendor ↗ |
Operator guidance
The price/performance sweet spot of the Claude 4.x line prior to Sonnet 5's release. Step up to Opus 4.8 for the hardest tasks; drop to Haiku 4 for cost/latency-sensitive work.
Use cases
- The default workhorse: strong quality at a moderate price/latency
- Production coding agents, RAG, and tool use at scale
Limitations
- Not as strong as Opus 4.8 on the very hardest reasoning/agentic tasks
- Reasoning/tool-use/instruction-following/multilingual ratings are qualitative (vendor + community signal, inherited from the Sonnet 4 family's established profile); coding is now benchmark-cited directly
- Uses the PREVIOUS tokenizer, not the Opus-4.7+/Sonnet-5 one — Anthropic's pricing docs list Sonnet 4.6 and earlier as using the older tokenizer, so token counts for this model are NOT inflated ~30% the way Sonnet 5's are for the same text