xAI Grok 4.3 and 4.6 API Tariffs: Long-Context Bills Double Each Category

Versioned short-context prices

The xAI USD developer rate card lists grok-4.3 at $1.25 per million input, $0.20 cached input and $2.50 output tokens. Grok-4.6 is $2.00/$0.50/$6.00. A model alias should be logged separately from the exact returned model identifier, because migration or retirement can change a deployment even if a URL continues to resolve. These are standard API prices rather than a subscription to the Grok consumer app.

The 200,000-token boundary

On both models, long-context rates begin when the request prompt reaches the 200k threshold. Grok 4.3 then uses $2.50/$0.40/$5.00 and 4.6 uses $4.00/$1.00/$12.00 per million input/cached/output tokens. The official wording says that the long-context tariff applies to all billed tokens in the qualifying request, not only the portion after token 200,000. A model of invoice cost must select the request band first, before multiplying each token counter.

Dated rollout and replacement caution

xAI’s release notes identify the 12 August 2026 Grok 4.6 API rollout. Other release notices explicitly mention retirements and redirects from earlier Grok endpoints. Historic usage attributed to an old alias is therefore not automatically evidence of a retired model’s earlier rate. Save a per-request model ID and timestamp when possible, and avoid backdating a current price page to the date an older version first appeared.

A 210k prompt example

Suppose a Grok 4.6 request has 210,000 uncached input tokens and produces 5,000 output tokens. The appropriate long-context arithmetic is 0.21×$4 plus 0.005×$12 = $0.90. Pricing those 210k prompt tokens at the short-context $2 rate would yield a wrong comparison. The example excludes cache, network, server tools and taxes. For planning, model the distribution of individual request sizes, not merely the month’s aggregate tokens.

A regional endpoint does not mean all products relocate

xAI documents a US inference endpoint with a 10% token premium, scoped to selected text models; global routing need not guarantee one execution geography. Regional controls have limited coverage and should not be projected onto Imagine video or speech without official evidence. Compare the same model and prompt length at the exact endpoint, then ask the provider about data-handling commitments instead of inferring them from the surcharge alone.

How to verify the actual invoice

Begin with a small representative workload rather than the cheapest published headline. Record the exact provider and billing channel, API model identifier, currency, account geography, plan and commitment length. The price or quota evidence in this guide is keyed to cached_input_tokens_per_1m, input_tokens_per_1m, output_tokens_per_1m. Do not combine proprietary creative credits, token meters or media seconds into a single currency or invent a conversion rate between them. If an enterprise or regional rate is not published, leave its price unknown and obtain a written quote.

Keep separate estimates for failed attempts, optional tools, storage, exports, human quality review and tax. Use source-verified line items from the companion CSV instead of typing permanent numbers into your forecasting formulas. Where an old plan or model was superseded, preserve its original historical record and create an additive later snapshot. This guide does not claim a product test, performance ranking, independent legal assessment or live procurement approval.

Availability, alternatives and editorial limits

Available plans, supported endpoints, trial eligibility, quota reset and contract rights may vary by region and account even when a public website displays a single global currency. Compare an alternative only with the same accepted task, input/output volume, model version and privacy requirements. Vendor-provided marketing claims are not independent evidence of quality. This historical pricing article may be cross-referenced by the model, media or enterprise procurement team, but its purpose is to explain the billing mechanics without duplicating their evaluations or inventing test scores.