Сравнение моделей Grok: какую модель Grok выбрать?

Updated September 7, 2026.

Сайт official Grok model catalog recommends 4.6 for chat and code. That makes it a useful default, but your workload may favor a cheaper model, less reasoning or a different product. A long coding job, a quick email rewrite and a live news question should not all use the same settings.

Prefer a ready-to-use model picker? Grok 4.6 on GlobalGPT gives you a direct place to start, and GlobalGPT also lists Grok 4.3. Try the same brief on both before committing your recurring work to one model.

Попробуйте Grok 4.6 прямо сейчас

Best Grok model by task

Everyday chat and writing

For email, brainstorming, summaries and ordinary questions, start with the default model your Grok app offers. Move to a deeper reasoning option when the task has constraints that the first answer misses. An expensive plan is hard to justify solely for short rewrites if your existing access already delivers useful answers.

If you are choosing an API backend for those everyday tasks, Grok 4.3 with нет или низкий reasoning is a sensible budget candidate. Its official model specification supports both settings. Test factual accuracy and instruction following before using the cheaper configuration at scale.

For the basic account and chat workflow, the Grok getting-started guide provides useful orientation. Use the current model selector and account usage screen for the options available to you today.

Reasoning and difficult analysis

Choose Grok 4.6 for a demanding analysis, complex debugging explanation or problem with several dependent steps. Start at высокий; consider xhigh when another round of thinking is worth the wait. The official reasoning controls define these effort levels, including the extra latency of xhigh. More thinking still requires checking the answer.

Coding and repository work

For repository work and sustained agentic coding, start with Grok 4.6 in Grok Build, or use the same model in your own agent. Build supplies the coding environment; 4.6 is the model powering its current default workflow. Evaluate changes through your normal tests and code review.

For a cost-sensitive coding API, benchmark grok-build-0.1 on your own repair and implementation tasks. It has lower listed token rates but a smaller context window. The Grok 4.6 review covers the flagship in more depth when you need a closer look at its features and published results.

Real-time information

For news, company updates or recent public discussions, use a model with search enabled. The API exposes Поиск в Интернете for internet sources; a model responding only from its training cannot supply reliable current-event coverage.

Используйте X Search when the question specifically concerns posts, profiles or threads on X. Ask for dated sources and open the important citations. Grok 4.6 is a sensible synthesis model; the freshness comes from the retrieved material.

Speed and response time

For a latency-sensitive 4.6 application, start with low reasoning effort and concise output requirements. For simpler API tasks that can avoid reasoning entirely, compare Grok 4.3 with none. Measure both time to first useful text and time to a complete usable answer; streaming speed alone does not capture the wait for reasoning or tools.

Budget and token costs

Grok 4.3 is the lower-cost general-purpose candidate in this comparison. It also gives you explicit control over whether reasoning happens. Grok Build 0.1 has lower listed input/output rates still, but its coding focus makes it a different decision. Compare the total cost of successful jobs, including revisions and tool use.

API development

Start a new general agent with Grok 4.6. Keep 4.3 when lower cost, disabled reasoning or its larger context is a requirement. Consider the still-listed 4.20 variants when an existing integration depends on that version or when you need its specialized multi-agent research workflow. A currently supported model is not automatically the best starting point for a new project.

Limited permissions and account access

The practical winner is the model and tool combination your account exposes. A model available through the API or a separate platform does not establish access through a particular Grok app or X subscription. Confirm the exact model, enabled tools and usage allowance on the route you plan to use.

Current Grok models compared

The following are the seven text API entries listed on the официальная страница с ценами on September 7, 2026. Costs below are USD per 1 million uncached input / output tokens at standard processing rates with a prompt below 200,000 tokens. They are API units, not monthly subscription prices. Team and regional availability can differ.

The text models above accept text and image inputs and return text. An image input can support document or visual analysis; it does not make that model an image generator. A bigger context window also does not guarantee that every detail in a long document will be used correctly.

Reasoning, Fast and Heavy: what the names mean

Reasoning is either a setting or a named model variant

On 4.6 and 4.5, reasoning effort is a setting on the same model. Grok 4.6 offers low, medium, high and xhigh, with high as the default. Grok 4.5 supports low, medium and high; the documentation says xhigh is treated as high on models that do not support it. Neither model offers a none setting.

Grok 4.3 does support none. The 4.20 family separately lists reasoning and non-reasoning model IDs. Those distinctions matter in code: changing a setting, choosing a different variant and switching to a different generation are three different operations.

Fast has more than one meaning

The older Grok 4 Fast and Grok 4.1 Fast names were actual API model variants. They appear on the May 15, 2026 retirement list. Continuing to send an old Fast ID does not preserve the original model or its original price.

Сайт Grok 4.6 launch announcement also advertises a fast variant at twice the price. That launch phrase does not supply a separate grok-4.6-fast ID in the current official model table. Use the exact option documented by your provider instead of constructing a model name.

For the official API, Приоритетная обработка is a documented speed control: request service_tier: "priority" while retaining the model ID. It changes scheduling and has a 2x token-price premium when the returned tier confirms priority. Its availability depends on capacity. Do not infer a new model from a delivery-speed label.

Heavy is not the name of every premium Grok model

Сайт July 2025 Grok 4 announcement used Grok 4 Heavy for parallel test-time compute that considered multiple hypotheses, and SuperGrok Heavy for a subscription tier. That is historical naming context, not evidence that a Heavy model is the current default in your account.

The current API has a specifically documented Grok 4.20 multi-agent research beta. Its effort setting changes how many agents collaborate. It should not be treated as a synonym for a consumer Heavy subscription, or automatically preferred for a short chat.

Which Grok models can you actually access?

Grok is the assistant and model family; xAI is the provider behind the API. The current documentation carries SpaceXAI branding. The official Grok product overview identifies grok.com, iOS and Android as consumer entries and says Grok is free to start. SuperGrok and X Premium names describe paid access, not underlying model IDs.

Сайт Grok account and usage FAQ describes a shared weekly allowance for paid users, with usage visible in Settings. It also explains linking an X account to receive relevant benefits. A subscription name alone does not tell you the exact model, tool access or remaining allowance.

For everyday consumer use, confirm what the model picker actually offers. An automatic selection or a longer-thinking control describes product behavior unless the interface explicitly names the underlying model. The API catalog does not establish a one-to-one mapping for every Grok app mode or X tier, and it does not establish universal free access to 4.6.

On GlobalGPT, start from the verified Grok 4.6 entry and use the platform’s own plan and usage information. The current public catalog also lists Grok 4.3. Access to a chat model through GlobalGPT does not establish that every xAI API reasoning setting or search tool is exposed in the same way.

If the real decision is whether Grok should be your only assistant, the Grok vs ChatGPT work comparison broadens the choice to writing, research and coding workflows across providers.

The cost differences that change the choice

For ordinary uncached short-context API requests, 4.6 costs $2 per 1M input tokens and $6 per 1M output tokens; 4.3 costs $1.25 and $2.50. That difference makes 4.3 worth evaluating for repetitive work. The lower price is useful only if it does not create more retries or manual cleanup.

Grok 4.5 has the same base input/output rates as 4.6, so its version number alone gives you no base-rate saving. Its short-context cached-input rate is lower: $0.30 versus $0.50 per 1M cached tokens. A validated, heavily cached 4.5 workload can therefore have a reason to stay on that version.

Three details can change the bill: reasoning tokens are billed; search and other tools can add charges; and a prompt reaching 200 тыс. токенов triggers long-context rates for all tokens in that request. . official API price table doubles the input/output rates for the listed models at that threshold. A 1M-token capacity is not a promise of short-context pricing throughout the window.

Compare consumer subscriptions by included usage and access, and compare APIs by actual token and tool consumption. The current consumer FAQ even lists API in its usage breakdown, so verify your account’s specific arrangement in the console. A consumer plan name is insufficient evidence of unlimited API calls or a particular API credit balance.

For the broader subscription and billing comparison, use the existing Grok pricing guide. For model selection, the useful question is narrower: does the more capable or faster option save enough rework to justify its total cost?

How to choose a Grok model for an API project

Use explicit model IDs for a new integration. The Grok 4.6 API guide documents the Responses API and its supported reasoning configuration. For a quick coding explanation, a request body can start like this:

{
  "model": "grok-4.6",
  "reasoning": {
    "effort": "low"
  },
  "input": "Fix numeric sort: [10,2].sort()"
}

That is a configuration example, not a performance result. Set an output budget appropriate to the job, enable the search tools needed for fresh sources, and record actual token usage. Use medium or high for deeper work after checking whether low meets your acceptance criteria.

1. Use representative tasks: a normal request, an ambiguous one, a long input, a tool failure and a difficult reasoning or code task.
2. Keep the setup comparable: same prompts, documents, tools and acceptance criteria; record model ID and reasoning effort.
3. Measure useful outcomes: correctness, complete-job latency, revisions, input/output and reasoning tokens, tool calls and total spend.
4. Choose a fallback deliberately: do not silently switch a failed current model to a retired ID or to a configuration with different reasoning behavior.
5. Recheck access: the console’s team model list governs your actual availability; a public catalog is a capability reference.

Сайт Grok API integration guide is useful background for the developer workflow. When adapting older examples, replace the model ID using the current official catalog and migration guidance.

For an established 4.5 deployment, run the same acceptance tasks on 4.6 before moving production traffic. For a 4.20 integration, preserve its explicit variant when that behavior is part of your baseline. Product defaults and moving aliases can change independently of your application.

Older Grok models and retired API names

Grok 3, the original Grok 4 API release and the old Fast models belong in migration context. The May 15, 2026 retirement notice lists grok-3, grok-4-0709, both grok-4-fast-* variants, both grok-4-1-fast-* варианты и grok-code-fast-1.

• Retired reasoning variants: the notice routes them to Grok 4.3 with low reasoning effort.
• Retired non-reasoning variants and Grok 3: the notice routes them to Grok 4.3 with none reasoning effort.
• Grok Code Fast 1: the code-specific migration section routes it to Grok Build 0.1; the current Build 0.1 page also lists the old code ID as an alias.

The old request name may still resolve while a different model serves the request and different rates apply. Read the official retirement and redirect instructions before treating an old benchmark, API bill or Fast label as evidence of current behavior.

Сайт Grok 4 family overview provides background on the earlier flagship generation. For a fresh selection, use the current 4.6/4.3 comparison above; the still-listed 4.5 and 4.20 entries remain specific alternatives rather than retired models.

What about images, video and voice?

Сайт current official catalog routes image generation to Grok Imagine Image 2.0, video generation to Grok Imagine Video 1.5 and live voice work to the Grok Voice API. These are separate modalities and services. Selecting a text reasoning model is not the whole decision for a media application.

For a creation workflow, use the Grok image and video generation guide. Keep its media choices separate from choosing a model for chat, analysis or code; media costs use different units from text tokens.

Часто задаваемые вопросы

What is the best Grok model right now?

Grok 4.6 is the default starting point for demanding reasoning, coding and knowledge work. Grok 4.3 is the general-purpose budget alternative for API use, with a larger context window and an option to disable reasoning.

Which Grok model is best for everyday chat?

Start with the current default your Grok account offers. For a chat API, evaluate Grok 4.3 with none or low reasoning for routine work, and move to 4.6 when the task needs deeper analysis or more demanding coding.

Can Grok 4.6 run without reasoning?

No. The official API supports low, medium, high and xhigh reasoning for Grok 4.6, with high as the default. Grok 4.3 supports none when you need to turn reasoning off.

Which Grok model should developers choose for coding?

Start with Grok 4.6 for demanding coding and agent workflows. Evaluate grok-build-0.1 when lower token costs matter. Grok Build is the coding application, while grok-build-0.1 is a separate API model ID.

Is Grok Fast the same as Grok Heavy?

No. Earlier Fast API models are retired; the 4.6 launch also describes a paid fast variant. Grok 4 Heavy was a parallel-compute model, while SuperGrok Heavy is a subscription tier. Check the exact model and delivery option on your chosen route.

Does the latest Grok model automatically know current news?

No. Current information requires search tools or supplied sources. Enable Web Search or X Search when appropriate, request dated citations and verify important claims in the linked sources.

Does a SuperGrok or X subscription guarantee every API model?

A subscription name does not establish access to every API model. Check the linked account, its model selector and usage allowance, then confirm API team availability and billing in the xAI console.

Can I use Grok 4.6 on GlobalGPT?

Yes. GlobalGPT lists a Grok 4.6 entry and also lists Grok 4.3. Its plans, usage limits and exposed tools are platform-specific; an xAI subscription or API feature does not establish the same entitlement on GlobalGPT.

Can I keep using an old Grok Fast API name?

The retirement notice says old Fast slugs continue to resolve through redirects. That does not preserve the original model or pricing. Use an explicit current model ID and confirm the reasoning setting and billed rates.

Which Grok model should you choose?

Choose Grok 4.6 for a new demanding workload, Grok 4.3 for a lower-cost general API workflow, and evaluate Grok Build 0.1 when coding cost is the priority. Keep 4.5 or a 4.20 variant when you have a concrete compatibility or evaluation reason. For ordinary app use, start with your available default and spend more only when better results, higher limits or a specific tool justify it.

The best Grok model is the one that meets your task requirements at an acceptable total cost. Pair it with the right reasoning level, current-information tools and access route, then judge the completed work.