GROK 4.5
Released July 8, 2026. GPQA 93.1, 99th-percentile intelligence. 500K context, vision and tools, native X/web/news live search. $2/$6 per million tokens.
Grok 4.5 is xAI's flagship model, released July 8, 2026. Its benchmark profile is genuinely frontier — 93.1 on GPQA and 99th-percentile intelligence ratings — and its structural advantage is unchanged: native live search across X, the web, and news gives it access to real-time social data no other lab's search stack reaches. It ships a 500K-token context window, vision input, and tool support, priced at $2 per million input tokens and $6 per million output — aggressive for a flagship; the cheapest premium-tier model in Council's lineup by output price. Versus its predecessor Grok 4.3 ($1.25/$2.50, released April 30, 2026), Grok 4.5 raises the reasoning ceiling substantially but costs more per token and, notably, halves the context window from 1M to 500K — if you need the bigger window, 4.3 and the 2M-context Grok 4.20 stay in the lineup. The case for Grok 4.5: current-events research, market and social-sentiment work, and strong reasoning at a mid-tier price. The case against: whole-repo coding (500K may not fit; Claude and Gemini go to 1M) and refactor quality, where the Claude line leads.
This is not a strict upgrade, and it's worth being precise. Grok 4.5 is the stronger reasoner by a clear margin (GPQA 93.1, 99th-percentile intelligence) — but it costs $2/$6 against 4.3's $1.25/$2.50, and its 500K context window is half of 4.3's 1M. Grok 4.3 (released April 30, 2026) stays in Council's active lineup precisely because of that: if your workload is long-document analysis at volume, 4.3 may still be the better xAI seat. Pick 4.5 for reasoning quality and live-search work; pick 4.3 for context size per dollar; pick Grok 4.20 when you need the full 2M-token window.
Grok 4.5 is available on every paid Council AI plan (Plus, Pro, Ultra) and usually takes the real-time seat in a council: when your question touches anything that happened after other models' training cutoffs, Grok's live X/web/news search grounds the debate in current data. It answers in parallel with models from up to 8 other labs, and the moderator's agreement score makes disagreement useful — when Grok cites something the others missed, you see exactly where the fresh information changed the answer.
$2.00 per million input tokens and $6.00 per million output tokens — the cheapest output price of any premium-tier model in Council's lineup. Council AI bundles it into monthly plan budgets.
500,000 tokens. Note that's smaller than Grok 4.3's 1M and Grok 4.20's 2M — xAI traded window size for reasoning strength in this release.
At reasoning, clearly — GPQA 93.1 and 99th-percentile intelligence ratings. But 4.3 is cheaper ($1.25/$2.50 vs $2/$6) and has double the context window (1M vs 500K), so 4.3 remains the better pick for long-document volume work.
It's the only frontier model with native live search across X as well as the web and news. For social sentiment, breaking news, and market chatter, it reaches data other models' search stacks don't index.
Yes — on all paid plans (Plus, Pro, Ultra), running in parallel councils with models from 8 other labs and consensus scoring on every answer.