grok-4.6Grok 4.6 is xAI's frontier model, built around long-running agents and multi-step work — sustaining a task across many turns of research, codebase edits, or building an artifact. It matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index (61) and leads its predecessor on every published eval. 500K context, text and image input, and four reasoning-effort levels (low / medium / high — the default — / xhigh). Priced at $1.765 input / $5.294 output per 1M tokens, 12% under xAI official ($2.00 / $6.00), with cached input at $0.441. Called through the OpenAI-compatible /v1/chat/completions endpoint; /v1/responses also works and is the one that runs live web search. Failed requests are never billed.
View complete API reference with all parameters and examples.
View complete API reference with streaming, thinking, and more.
Billing: Cost = (input_tokens * input_price + output_tokens * output_price) / 1,000,000
$1.765 input / $5.294 output per 1M against official $2.00 / $6.00, with cached input at $0.441. Same model, same weights — the price difference is not a lesser tier
The 4.6 release targets tasks that span many steps — researching a topic, working across a codebase, turning an idea into a finished artifact — with more self-checking along the way. 500K-token context
low / medium / high / xhigh via reasoning_effort. Measured on our channel: 464 / 1,720 / 2,663 / 3,341 reasoning tokens on the same prompt. Unlike GPT-5.6, xhigh really is the deepest tier here
12/12 correct across 12 concurrent calls using a nested schema with enums and arrays — arguments parsed cleanly with no cross-request mixing. Streaming, multi-turn and parallel tool calls all pass
Cached input bills at $0.441 instead of $1.765. We send a per-conversation routing key so repeat turns keep landing on a warm server — without it, cache hits drop back to cold on roughly every other repeat
The default reasoning_effort is high, which runs about 53 seconds — set a client timeout above 60s. And billed output includes reasoning tokens, so your bill can exceed the visible answer; lower the effort when you do not need the depth
Grok 4.6 は xAI の大規模言語モデル API です。Grok 4.6 is xAI's frontier model, built around long-running agents and multi-step work — sustaining a task across many turns of research, codebase edits, or building an artifact. It matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index (61) and leads its predecessor on every published eval. 500K context, text and image input, and four reasoning-effort levels (low / medium / high — the default — / xhigh). Priced at $1.765 input / $5.294 output per 1M tokens, 12% under xAI official ($2.00 / $6.00), with cached input at $0.441. Called through the OpenAI-compatible /v1/chat/completions endpoint; /v1/responses also works and is the one that runs live web search. Failed requests are never billed. APIMODELS のプラットフォーム経由なら、統一 API と明朗な従量課金でこのモデルを呼び出せます。 現在の料金: Input: $1.765, Output: $5.294 per 1M tokens。
問い合わせに自動で答える対話システムを構築し、対応の効率を上げます。
記事、メール、広告コピーなどの文章を自動で書き、制作の手数を減らします。
コードの記述、デバッグ、レビューを補助し、開発を前に進めます。
非構造化データを読み解き、要点を抽出して要約レポートにまとめます。
Grok 4.6 は APIMODELS 経由で Input: $1.765, Output: $5.294 per 1M tokens で利用できます。課金は従量制で、生成した分だけの支払いです。
APIMODELS に登録して API キーを取得し、統一エンドポイントを呼ぶだけです。cURL / Python / Node.js のサンプルを含む詳しいドキュメントを用意しています。
APIMODELS は同じ Grok 4.6 を集約プラットフォーム経由で提供します。統一された API インターフェースなので、プロバイダごとにアカウントを作る必要はありません。キー 1 本ですべてのモデルに届きます。
$1.765 per 1M input tokens and $5.294 per 1M output, with cached input at $0.441. xAI official is $2.00 / $6.00, so both sides are 12% cheaper here. Pay-as-you-go, no minimum top-up, no subscription, and failed requests are never billed. Its sibling Grok 4.5 is $1.5 / $5 — cheaper, one capability tier down.
POST https://api.apimodels.app/v1/chat/completions with Authorization: Bearer YOUR_API_KEY and body {"model":"grok-4.6","messages":[{"role":"user","content":"..."}]}. It is plain OpenAI shape, so pointing openai-python or openai-node at base_url=https://api.apimodels.app/v1 works as-is, and so do Cursor, Cline and Codex. Add stream:true for streaming. /v1/responses is supported too.
The 4.6 release is aimed at long-running agents: tasks that span many steps — working across a codebase, turning an idea into a finished artifact — hold together better, and the model does more self-checking along the way. It scores 61 on the Artificial Analysis Intelligence Index, level with GPT-5.6 Sol, and beats 4.5 (56) on every published eval. Take Grok 4.5 ($1.5 / $5) when budget matters and tasks are short; take 4.6 for sustained work. Same endpoint — switching is a one-field change.
Yes, and we stress-tested it rather than smoke-testing it: non-streaming, multi-turn with tool results fed back, streaming, and parallel tool calls in one turn all passed 5/5. Then 12 concurrent calls against a nested schema with enums and arrays came back 12/12 correct — arguments parsed cleanly, no cross-request mixing. Grok upstreams have historically had intermittent tool-call failures on mixed pools; this channel did not reproduce that.
low, medium, high and xhigh, with high as the default. On one fixed prompt we measured 464 / 1,720 / 2,663 / 3,341 reasoning tokens respectively — strictly increasing, and unlike GPT-5.6, xhigh genuinely is the deepest tier for this model. The default level takes about 53 seconds, so set your client timeout above 60 seconds rather than leaving it at 30.
Because billed output tokens include reasoning tokens. A measured example: the model replied with a single number — one visible token — after spending 1,052 tokens reasoning, and the call bills 1,053. This is not padding on our side; the upstream charges us on exactly that basis and we pass it through. Lower reasoning_effort when you do not need the depth and the cost drops immediately.
Image input works — in testing it read text out of a picture accurately. Live web search runs on the **/v1/responses** endpoint with tools:[{"type":"web_search"}]; there it genuinely executes, returning web_search_call entries and real citation URLs. Passing search_parameters to /v1/chat/completions currently does nothing — the model emits the tool call as plain text instead — so do not rely on it on that path.
500K tokens. It is directly reachable from mainland China with no xAI account and no VPN, and the same API key also reaches every other model on the platform. Long conversations hit prompt caching automatically, billing the cached portion at $0.441 instead of $1.765; we also keep per-conversation routing sticky so repeat turns keep landing on a cache-warm server.
APIMODELS では Grok 4.6 が 60 以上のモデルと同じ API キー・同じ残高の上に並んでいるので、選択は「相性」の問題であって「囲い込み」の問題ではありません。Frontier、Long-Running Agents、Tool Calling、500K Context、Vision に対応しており、他の大規模言語モデルモデルと価格・機能を並べて評価できます。乗り換えはモデル名の文字列を 1 つ書き換えるだけ。新しいアカウントも追加の実装も要りません。大規模言語モデルの選択肢と最新価格は apimodels.app/models で確認できます。
Grok 4.6 は次に対応しています: Frontier、Long-Running Agents、Tool Calling、500K Context、Vision。パラメータの全一覧と呼び出し例は APIMODELS のドキュメントをご覧ください。
はい。APIMODELS は Grok 4.6 を単一の統一 API とキー 1 本で提供します。プロバイダごとのアカウントも、各社の地域ごとのネットワーク経路を自分で面倒みる必要もありません。
Stripe（Visa、Mastercard などの国際カード）と Alipay に対応しています。支払い後、残高はすぐ反映されます。