Grok API Cost Explained: Grok 4.6 vs 4.5 Pricing
Current xAI token pricing including cached input, the 200K long-context doubling, and what a realistic Grok workload costs per month.
Updated Aug 11, 2026 · Reviewed by PromptCostLab
Token and pricing information changes frequently. Use this guide for planning and verify model-specific limits and prices with the provider before production billing.
Current Grok pricing (verified August 2026)
xAI's lineup as of mid-August 2026, from the provider's models page (last updated August 12):
- Grok 4.6 — new flagship: $2 input / $6 output / $0.50 cached, 500K context.
- Grok 4.5 — previous flagship, same $2/$6: cached input $0.30, 500K context.
- Grok 4.3 — value tier: $1.25/$2.50 with $0.20 cached and a 1M context.
- Grok Build 0.1 — agentic-coding model: $1/$2 with a 256K context.
- Long-context tier: prompts ≥200K tokens double all rates for the entire request.
- Older grok-4-fast and grok-3 models were retired in May 2026 and now redirect to Grok 4.3.
The 200K trap
Grok's headline prices apply only below 200K prompt tokens. Cross that line — easily done inside a 500K window with a large document — and the entire request bills at $4/$12 (Grok 4.6) rather than $2/$6. A 250K-token prompt that looks like $0.50 of input actually costs $1.00. Budgets built on list prices understate big-context Grok bills by exactly 2x.
If your prompts hover near the boundary, trimming retrieved context below 200K is a genuine 50% input discount, not an optimization nicety.
Caching on Grok
Prompt caching is automatic on xAI: repeated prefixes are billed at the cached rate ($0.50 on 4.6, $0.30 on 4.5 — 25% and 15% of input respectively) and the cache survives 5–15 minutes of inactivity, refreshing whenever it is used. There is no write surcharge, which makes stable-prefix patterns — fixed system prompts, shared documents first — pure savings.
Note the interaction with the long-context tier: cached rates also double on 200K+ prompts ($1.00 on 4.6).
What a realistic workload costs
Take a chat assistant doing 3,000 requests a day with 3,500 input and 400 output tokens each on Grok 4.6, with a 60% cache-hit share on input. Per request: fresh input ~$0.0028, cached input ~$0.0011, output ~$0.0024 — about $0.0063 per request, or roughly $570 per month. The same workload without caching is about $756; on Grok 4.5's cheaper cache, about $530.
Run your own numbers with the PromptCostLab Grok calculator — it applies the 200K tier and cache rates automatically and compares the workload across every model in the registry.
Where Grok fits
At $2/$6 with a 500K window, Grok 4.6 undercuts several flagships on output price while offering automatic caching. Grok 4.3 at $1.25/$2.50 with a 1M context is the value play for large-context work. Watch the extras that sit outside token pricing: Live Search runs $5 per thousand basic searches ($25 for premium sources), and priority processing doubles rates.
How much does the Grok API cost?
Grok 4.6 and Grok 4.5 both cost $2 per million input tokens and $6 per million output tokens, with a 500K-token context window. Cached input is cheaper: $0.50 per million on Grok 4.6 and $0.30 on Grok 4.5. Prompts of 200K tokens or more pay doubled rates on the entire request.
What happens if my Grok prompt exceeds 200K tokens?
Input and output prices double for the entire request — $4 input, $12 output, and $1.00–$1.20 cached per million depending on the model. This tier applies at 200K+ prompt tokens, well inside the 500K window, so large-context budgets must use the doubled rates.
Does Grok support prompt caching?
Yes, automatically. Cache reads cost 25% of input price on Grok 4.6 ($0.50) and 15% on Grok 4.5 ($0.30), with the cache lasting 5–15 minutes of inactivity and refreshing on use. There is no cache-write fee.
Prices, token behavior, and model limits change. These primary provider docs are the right place to verify the details before shipping.