Grok 4.6 vs Grok 4.5: what the point release changed

Every column matches except the release date

grok-4.6 and grok-4.5 carry a 500,000-token context, a 65,536-token output ceiling, and text, tool-calling and thinking support. Both answer on POST /v1/chat/completions. Neither is a trimmed version of the other, and neither is missing a capability the other has.

Pricing is identical too. Official rates are $2 per million input tokens and $6 per million output on both ids, and both are billed at the 0.1x multiplier, which puts consumer cost at $0.20 per million input and $0.60 per million output either way. There is no price argument for or against the switch.

What separates them in the catalog is a date and one rate: grok-4.5 was released 2026-06-03, grok-4.6 on 2026-08-12, 70 days later. Cached reads are the one price that differs — $0.30 per million official on 4.5 against $0.50 on 4.6, which at 0.1x is $0.03 against $0.05. Any claim about one reasoning better than the other is not something the catalog can support, and we are not going to make it up.

Because the price is the same, trying costs nothing

When two ids differ in price, a switch is a budget decision you can reason about on a spreadsheet. Here the fresh-token rates match exactly, so the spreadsheet has almost nothing to say and the only evidence that counts is your own output.

The request shape is the same on both, so the change is one string. Send a batch of your real prompts through each id, diff the answers, and keep whichever you prefer — the bill will be the same either way, which is unusual enough that it is worth taking advantage of.

If you run an agent loop, run the comparison as a loop rather than a single prompt. A one-shot answer hides the failure mode that matters most in agents: a tool call that goes wrong on step four and drags every later step with it.

curl https://token-share.app/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN_SHARE_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "grok-4.6",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Refactor this function and explain the change."}
    ]
  }'

The 500,000-token window is the shared constraint

Both ids stop at 500,000 tokens of context and 65,536 tokens of output. If your prompt does not fit, moving between 4.5 and 4.6 does not help you — that is a different decision, made against the ids in this catalog that carry a 1,000,000-token window.

The same applies to output length. A job that needs more than 65,536 tokens in one response has to be split regardless of which Grok id you pick, or run against a model with a higher ceiling.

Within the window, the two behave the same way at the till, including the step in the rate card: the listed $0.20 per million input and $0.60 output are the sub-200K tier, and xAI doubles both at or above 200,000 tokens in a single prompt, on 4.6 and 4.5 alike. Sending more context is a cost decision, not a quality one — and past 200,000 tokens it is a steeper one than the headline rate suggests.

Common questions