Grok 4.5 Pricing

Grok 4.5 API pricing

Grok 4.5 API pricing is $2 per million input tokens and $6 per million output tokens, verified September 14, 2026.

Input (per 1M tokens)

$2

Output (per 1M tokens)

$6

Cached input (per 1M tokens)

$0.5

Context window

500,000 tokens

Released

July 8, 2026

Capabilities

Grok 4.5 supports extended reasoning, tool use.

No Grok 4.5 price changes have been recorded yet.

No price snapshots recorded yet.

Frequently asked questions

Short answers to the most common questions about Grok 4.5 pricing, limits, and capabilities.

Notes

1.5T param coding-focused model trained on Cursor data. Musk describes it as "Opus-class" — competitive with Anthropic's top tier. TIERED PRICING at 200K prompt threshold: ≤200K prompts at $2/$0.50 cached/$6 per MTok (input/cached/output); >200K prompts at $4/$1/$12. Configurable reasoning_effort (low/medium/high, default high). Supports function calling, web search, X search, code execution, prompt caching. Available in Cursor (all plans), Grok Build, xAI console, OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic. [2026-08-17] Input $2.00 and Output $6.00 re-verified unchanged (mem0.ai, benchlm.ai, layer3labs, costgoat). CACHE READ CONFLICT: mem0.ai/benchlm now report cached input at $0.30/M for Grok 4.5, while a second search citing docs.x.ai reports $0.50/M. Left Cache Read at $0.50 pending resolution — needs a manual check of docs.x.ai/developers/models/grok-4.5. Recheck next run. Also noted: xAI has released Grok 4.6 at the same headline rates as 4.5 — not added, see new-models list. [2026-08-24] Input $2.00 / Output $6.00 re-verified unchanged (mem0.ai, benchlm.ai, sentisight.ai, layer3labs). CACHE READ CONFLICT UNRESOLVED FOR A SECOND WEEK — sources still split between $0.30/M (≤200K) / $0.60/M (>200K) and $0.50/M; sentisight.ai explicitly notes "some sources list cached input at $0.50" and recommends confirming on docs.x.ai. Cache Read LEFT AT $0.50. ACTION FOR KHALED: this one needs a manual eyeball on docs.x.ai/developers/models/grok-4.5 — search snippets will not settle it. Note the $0.30 figure would make Grok 4.5 cache reads 40% cheaper than currently modelled in the calculator. [2026-08-31] Input $2.00 / Output $6.00 re-verified unchanged (mem0.ai, benchlm.ai, layer3labs, spheron.network). CACHE READ CONFLICT NOW IN ITS THIRD WEEK — search results this run explicitly acknowledge the split ("$0.50 per million, a 75% discount versus fresh input" vs "cached input runs $0.30 per million below the 200K token threshold") and again defer to docs.x.ai. Cache Read LEFT AT $0.50 for the third consecutive run. ESCALATING: automated verification cannot settle this field — recommend Khaled resolve it manually once and then we stop re-flagging it weekly. [2026-09-07] Input $2.00 / Output $6.00 re-verified unchanged (mem0.ai, benchlm.ai, eesel.ai, layer3labs). CACHE READ — FOURTH WEEK, BUT THE EVIDENCE HAS SHIFTED. This run produced the first source that explicitly contrasts the two models side by side rather than quoting one in isolation: "Cached input on 4.6 is 67% more expensive than on 4.5 — $0.50 versus $0.30." That supplies a mechanism for the whole three-week conflict: the $0.50 figure appears to be Grok 4.6's cached rate being mis-attributed to 4.5. A second source independently gives 4.5 cached at $0.30/M (≤200K) and $0.60/M (>200K). Cache Read STILL LEFT AT $0.50 — no primary docs.x.ai read was obtainable (Cowork's egress allowlist blocks direct page fetches), and three prior runs set the precedent of not flipping this field on aggregator evidence alone. RECOMMENDATION FOR KHALED: the balance of evidence now favours $0.30. One manual look at docs.x.ai/developers/models/grok-4.5 settles it; if it reads $0.30, change it and the calculator gets 40% cheaper Grok 4.5 cache reads. [2026-09-14] Input $2.00 / Output $6.00 re-verified unchanged (mem0.ai, benchlm.ai, layer3labs, costgoat). CACHE READ — FIFTH WEEK. Evidence again favours $0.30: a dedicated search this run returned $0.30/M for Grok 4.5 described as confirmed across multiple sources, and the side-by-side contrast with Grok 4.6 at $0.50/M held up, consistent with last week's finding that the $0.50 figure is Grok 4.6's rate mis-attributed to 4.5. Cache Read STILL LEFT AT $0.50 — no primary docs.x.ai read was obtainable. New this run: a direct read was attempted through the in-app browser pane rather than WebFetch, and was refused because docs.x.ai is not on the approved-sites list and no one was present to approve it mid-run. ACTION FOR KHALED — two ways to end this loop: (a) look at docs.x.ai/developers/models/grok-4.5 once and set the field yourself, or (b) approve docs.x.ai for the browser pane so a future run can read the primary source directly. Five runs of aggregator evidence will not clear the primary-source bar on their own.