// benchmarks and budget

Opus 4.7 vs Sonnet 4.6 for coding

Both models can code. One costs 1.7× the other. Here's when spending 1.7× is worth it, and when it's a waste.

Opus 4.7

Sonnet 4.6

Input / M

$2.50

$1.50

Output / M

$12.50

$7.50

Context

200K

SWE-bench Verified

~72%

~65%

Latency

Higher

~2× faster

Tool-call accuracy

Better

Good

Benchmarks approximate Anthropic's published Opus 4.5+ scores. On llmrelay both models run at 50% off official list — the ratio between them is what matters here.

Use Opus 4.7 when

+You're in agent mode. Multi-file refactors, long tool trajectories, planner-executor loops. Opus's edge on SWE-bench compounds when the agent has 20+ tool calls in a task.
+You're debugging real production code. Sonnet often "fixes" the wrong thing; Opus reads the stack trace and finds the actual bug.
+Cursor Composer, Cline, Roo Code agent mode — anything where the model plans its own steps.

Use Sonnet 4.6 when

+You're chatting, not agenting. Ask a question, get a code snippet, move on. Sonnet is fine and 40% cheaper.
+You're building a RAG or long-context pipeline. Sonnet's speed matters more than Opus's reasoning depth.
+You're running batch inference over thousands of items. Save the $ on the boring 90%.

Real budget math

Heavy Cursor agent user: ~50M input, 5M output per month.

Opus 4.7 on Anthropic

$375/mo

Opus 4.7 on llmrelay

$187.50/mo

-50%

Sonnet 4.6 on llmrelay

$112.50/mo

-70% vs Anthropic Opus

Try both — a single key covers all models

The right answer is "use both — Opus for the tough tasks, Sonnet for everything else." One key. Prepaid. No expiry.

Get API key →