THE VERDICT
● Under 200K input tokens: Grok 4.6 is materially cheaper from September 1, at 2 and 6 dollars against 3 and 15.
● Over 200K input tokens: Grok reprices the whole request to 4 and 12. The advantage narrows sharply.
● The hidden cost: Sonnet 5's tokenizer change adds 10 to 35 percent more tokens on code, on top of the rate rise.
● Do not switch on price alone. Run your own eval suite. These models are not interchangeable on output quality.
What changes on September 1
|
Now |
From Sept 1 |
| Claude Sonnet 5 input |
$2 / M |
$3 / M |
| Claude Sonnet 5 output |
$10 / M |
$15 / M |
| Sonnet 5 tokenizer |
Current |
+10 to 35% tokens on code |
| Grok 4.6, under 200K in |
$2 / $6 per M |
Unchanged |
| Grok 4.6, over 200K in |
$4 / $12, whole request |
Unchanged |
| Grok 4.6 context |
500K |
500K |
THE TOKENIZER CHANGE IS THE PART PEOPLE MISS
A 50 percent input rate rise is the headline. But if the same source file now tokenises 10 to 35 percent longer, your effective increase on a coding workload is well above 50 percent. Measure your own repos against both tokenizers before assuming the sticker figure.
Where each one actually wins
Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, leads APEX-Agents, and posts 69.9 percent on CursorBench against Grok 4.5's 66.7. It is live on the xAI API, Grok Build, Cursor, X, OpenRouter, Vercel and Cloudflare. Grok 4.7 has no model ID, no price and no benchmark card, with a window of early to mid September per Musk.
Claude Sonnet 5 remains the stronger pick where output quality on nuanced instruction-following matters more than per-token cost, and it sits inside a tooling ecosystem — Claude Code, agent frameworks, editor integrations — that Grok has not matched.
Who should do what
| If you are... |
Do this before Aug 31 |
| High-volume, short-context coding loops |
Price both. Grok 4.6 is the cheaper column from Sept 1 |
| Running agents with accumulating context |
Cap Grok input at 200K in code, or the saving disappears |
| Deep in Claude Code or the Anthropic SDK |
Migration cost likely exceeds the saving. Re-measure, do not assume |
| Undecided |
Route through OpenRouter now so the choice stays a config change |
| Tempted to wait for Grok 4.7 |
Do not plan around it. No price, no model ID, no benchmark |
FAQ
Exactly when does Sonnet 5 pricing change?
September 1, 2026. Input 2 to 3 dollars per million, output 10 to 15, with the tokenizer change landing alongside it.
Why does Grok double above 200K tokens?
xAI reprices the entire request, not just the overage. A 201K-token request costs roughly twice a 199K one end to end.
Should I switch purely to save money?
No. Price is one input. Run your own eval suite on tasks you actually care about — the two models are not interchangeable on output quality, and a cheaper model that needs two attempts is not cheaper.