Detailed Model Pricing

Updated November 2, 2026 — Imagine. As of November 2, 2026, grok-imagine-image-quality is retired: requests are served by Imagine Image 2.0 (grok-imagine-image-2.0, $0.04 per image) at low quality. Update the ID in your examples and set the desired quality; grok-imagine-image (version 1.0) is unaffected. Imagine Image 2.0 accepts up to five reference images and the auto quality setting, which is now the default. Updated November 3, 2026 — Voice: The voice agent costs $0.08 per minute ($4.80 per hour), not $0.05, according to the Pricing page. Updated November 3, 2026 — Models: The grid includes grok-4.7 (the model recommended by xAI since September) and grok-4.6, according to the Models page.

Choose the Right Model for Your Budget

Model selection is the technical decision that impacts your Grok bill the most: the gap between two models in the range is measured in factors, not percentages, and the most expensive one often provides no perceptible improvement for a given use case.

The xAI grid adds a specific detail to integrate from the start: tiered pricing. Each text model applies a base rate as long as the request stays under 200,000 context tokens, then a doubled rate beyond that. Input tokens served from cache largely escape this logic.

Text and Code Models

In each cell, the first figure applies under 200,000 context tokens and the second beyond that. A grok-4.7 request carrying 250,000 context tokens is therefore billed at $4.00 per million on input and $12.00 per million on output: the 50,000 tokens exceeding the threshold doubled the price of the entire request, not just that of the excess portion. Watch this tier closely on applications with long history, where context grows conversation after conversation.

Reading this grid gives four benchmarks. grok-4.7, the model xAI recommends, shares the highest rate and adjustable reasoning with grok-4.6 and grok-4.5; only the cached input rate distinguishes them. grok-4.3 offers the best value while having the largest window, 1 million tokens. The three variants of grok-4.20-0309 remain in the catalog at the same rate as grok-4.3. Finally, grok-build-0.1 is both the cheapest and fastest on code tasks, at the price of a shorter window of 256,000 tokens.

Visual Generation: Grok Imagine

For image and video, billing abandons the token in favor of the unit produced or the second generated, which simplifies estimation considerably. Count $0.02 per image with grok-imagine-image (Imagine 1.0) and $0.04 with grok-imagine-image-2.0 at low quality; medium quality is billed at its own rate, and grok-imagine-image-quality, at $0.05, is retired as of November 2, 2026. On the video side, grok-imagine-video comes to $0.05 per second, grok-imagine-video-1.5, available since June 2026, at $0.08 per second, and grok-imagine-video-1.5-lite at $0.02 per second.

Voice API: Voice Agent, TTS, and STT

Voice services bill per time or character. The real-time voice agent (grok-voice-3.0) costs $0.08 per minute, or $4.80 per hour of conversation, plus $0.004 per text input in real-time session. Text-to-speech comes to $15.00 per million characters. Speech recognition distinguishes its two modes: $0.10 per hour of audio in batch processing via REST, $0.20 per hour in streaming—real-time costs double, a trade-off to make depending on whether you are transcribing a live feed or an archive. Since July 2026, the Voice Agent Builder lets you assemble voice agents without writing code, and 21 new voices have joined the catalog.

The Input Cache Lever

Return to the “Cached Input” column in the table: it is the most profitable of all, since a token already in cache costs five to seven times less than a standard token. If your assistant reuses the same system prompt and the same instruction preamble on every request, this prefix is billed at $0.20 per million on grok-4.3 instead of $1.25.

You still need to structure your prompts for the cache to operate: strictly stable start—system prompt, instructions, examples—and everything that varies pushed to the end, user question or daily data. A variable inserted at the head, even just a timestamp, invalidates the entire prefix and makes you pay full price for tokens you could have had at a reduced rate.

Estimate Your Monthly Bill

Take a customer service application handling 10,000 requests per day on grok-4.3, with 2,000 tokens on input and 500 on output per request—therefore well under the 200,000 token tier—and a cache rate of 60%. Over thirty days, input totals 600 million tokens and output 150 million.

The same traffic served by grok-4.5 would come to about double: $108 cached input, $480 full input, and $900 output, or nearly $1,488 per month. Before arbitrating, ask yourself the only question that counts: do these additional $741 buy a response quality your users will notice? On a customer service well-scoped by its instructions, the answer is often no.

Rates as of August 5, 2026 — prices evolve regularly: before any budget calculation, check the current grid on the official xAI models and pricing page.

Key Takeaways

• Tiered pricing: base rate under 200k context tokens, doubled rate beyond—watch this threshold
• grok-4.3 ($1.25 / $2.50 per M) is the best value; grok-4.7 ($2.00 / $6.00) the recommended model, at the same rate as grok-4.6 and grok-4.5; grok-build-0.1 ($1.00 / $2.00) the cheapest for code
• Input cache divides the cost of stable prefixes by 5 to 7—structure your prompts accordingly
• Multimedia: image from $0.02, video from $0.05/s; voice: $4.80/hour voice agent, TTS $15.00/M characters, STT from $0.10/hour
• Estimate your bill by combining volume, cache rate, context tier, and model choice
• Rates evolve: check the official models page (docs.x.ai) before any budget commitment