Grok 4.7 Released: Benchmarks, Pricing, and the API

Grok 4.7 is SpaceXAI's flagship large language model for coding and knowledge work, released on September 21, 2026 as the successor to Grok 4.6. SpaceXAI is the name xAI now operates under, and on its API the model is called grok-4.7. SpaceXAI calls it the most capable model it has built, and its model docs now point developers to Grok 4.7 for everything except image, video, and voice, which have dedicated models.

Under the hood, it is not a tune-up. Grok 4.7 uses a new, larger base model than Grok 4.6, trained with a longer reinforcement-learning run on a harder mix of tasks, weighted toward problems that take many hours to complete. SpaceXAI says the result verifies its own work more carefully and manages long context better. It also trained the model to natively understand the Grok Bot harness (the agent product it launched in August), which it says improves conversational tasks and general knowledge work.

Grok 4.7 costs $2 per million input tokens and $6 per million output tokens on the xAI API, the same price and speed as Grok 4.6, with a 500,000-token context window. The API model name is grok-4.7, and it has a May 2026 knowledge cutoff. It takes text and image input and returns text with no text output limit, offers reasoning effort at low, medium, high (the default), or xhigh, and works on both the Responses API and Chat Completions, with function calling, web search, X search, and code execution as built-in tools.

Beyond the xAI API, it is the default model in Grok Build, SpaceXAI's coding agent, which has a free tier; it is available on every Cursor plan; and it is served through OpenRouter, Vercel, and Cloudflare. A US regional endpoint keeps inference in the United States at a 10% premium on token prices.

The fast variant needs a footnote. The launch post says SpaceXAI serves a version with twice the output speed at twice the price, but the model docs add that Grok 4.7 Fast is available only in Cursor and Grok Build, is not part of Grok Build's free tier, and is not on the public xAI API. If you are building on the API, the standard model is what you get.

On SpaceXAI's own table, Grok 4.7 beats Grok 4.6 on every benchmark shown and leads the field on electrical engineering and legal work, but trails Claude Fable 5.1 on CursorBench, Terminal-Bench, office work, and clinical reasoning. SpaceXAI ran Grok 4.7 at xhigh reasoning, Grok 4.6 at high, and the competitors at max effort.

Its two clear wins are in structured professional work. Grok 4.7 scores 64.0% on EEBench, an electrical-engineering benchmark, against 56.4% for Claude Fable 5.1, 39.4% for GPT-5.6 Sol, and 53.0% for Grok 4.6. On the Harvey Legal Agent Benchmark it scores 19.6%, against 6.7% for Fable 5.1, 2.5% for GPT-5.6 Sol, and 15.8% for Grok 4.6. On DeepSWE v1.1 it lands in the middle: 71.0% at high effort, just above Fable 5.1's 70.0% and just below GPT-5.6 Sol's 72.7%, up from 65.2% for Grok 4.6.

It trails Fable 5.1 on the remaining four. On CursorBench 4.0, Grok 4.7 scores 46.3% to Fable 5.1's 51.8% (GPT-5.6 Sol 41.7%, Grok 4.6 40.4%); on Terminal-Bench 4.0, 37.6% to 57.9% (Sol 37.3%, Grok 4.6 20.3%); on AA Briefcase v1.1, an office-work benchmark, 1,657 to 1,678 (Sol 1,487, Grok 4.6 1,546); and on HealthBench Professional, 56.7% to 62.1% (Sol 60.5%, Grok 4.6 48.5%).

Read the Terminal-Bench result carefully. Grok 4.7 nearly doubles Grok 4.6 there, from 20.3% to 37.6%, yet multi-hour terminal work is still its weakest showing against Fable 5.1's 57.9%, and HealthBench tells a similar story for clinical reasoning. Grok 4.7's wins are concentrated in structured professional domains, and price makes the rest of the case: at $2 and $6 per million input and output tokens, against $4 and $20 for GPT-5.6 Sol and $10 and $50 for Fable 5.1, its output tokens cost roughly an eighth of Fable 5.1's.

One more caveat about vendor tables. SpaceXAI lists Claude Fable 5.1 at 57.9% on Terminal-Bench 4.0; Anthropic's own launch post a day later lists Fable 5.1 at 55.8% on the same benchmark, and puts 57.9% against GPT-6 Astra instead. Vendors run and report competitor scores under different setups, so small discrepancies like this are common, and they are a good reason to treat every cross-vendor table as approximate.

Grok 4.7 is the cheaper model, and Claude Opus 5.5 is the stronger one on the benchmarks both vendors report. Anthropic released Opus 5.5 on September 22, 2026, one day after Grok 4.7, and the two launch posts share enough reference points to line them up.

Grok 4.7 costs $2 and $6 per million input and output tokens, has a 500,000-token context window, and defaults to high reasoning effort; Opus 5.5 costs $4 and $20, has a 1M-token window, and defaults to medium. On CursorBench 4.0, Grok 4.7 scores 46.3% at xhigh, while Opus 5.5 scores 57.8% at max effort and 52.5% at its default medium. On Terminal-Bench 4.0, both measured at xhigh, the gap is wider: 37.6% against 66.4%.

CursorBench is the more trustworthy of the two comparisons, because both launch pages report identical scores for the shared reference models (Fable 5.1 at 51.8% and GPT-5.6 Sol at 41.7%), which suggests comparable setups. Even at its default effort, Opus 5.5 scores about six points higher than Grok 4.7 at xhigh.

What neither vendor publishes is tokens per task for the other's model, and that is the number that turns a per-token price into a real bill. At $6 per million output tokens, Grok 4.7 can still be cheaper per task even if it needs more turns; you only find out by running both on your own workload. The companion piece "Claude Opus 5.5 Released: Benchmarks, Pricing, Migration" covers its pricing and migration details.

Switching is a one-line model change on the Responses API or Chat Completions, but three details in the docs affect cost and correctness. The request itself is ordinary: send it to the xAI Responses endpoint with your API key as a bearer token, set the model to grok-4.7, and pass your input.

First, set a prompt cache key. SpaceXAI strongly recommends prompt_cache_key on the Responses API, or the x-grok-conv-id header on Chat Completions, so a conversation's requests reach the same server; without it you often pay full input price on a cache-cold server. Second, pass reasoning back unchanged. On the Responses API, grok-4.7 always returns encrypted reasoning content, and a multi-turn conversation keeps the model's reasoning only if you send those reasoning items back in the next request. Third, choose the effort level on purpose. Reasoning runs from low to xhigh with high as the default, and SpaceXAI's headline coding scores were measured at xhigh, not at the default.

SpaceXAI says Grok 4.7 ships with an entirely new safeguard stack and is the strongest model it has tested on refusals and jailbreak resistance. On HackerBench v0.3, SpaceXAI's own benchmark of risky and malicious cyber tasks, it let through 3.3% of risky dual-use prompts, and on LatchBio's biosafety benchmark it scored 62.4%, the top result. Select cybersecurity partners also get invite-only access to its red-team capabilities for defense research.

Both numbers deserve the usual skepticism. HackerBench is SpaceXAI's own benchmark, and SpaceXAI publishes no false-positive rate behind its claim that the model rarely blocks legitimate security work. If you do security engineering, test it on your own tasks. The contrast this week is striking: Anthropic re-routes most cybersecurity tasks on Opus 5.5 to an older model, while SpaceXAI is pitching Grok 4.7's balance of cyber capability and low refusals.

Use it where price per token dominates and the work plays to its strengths: high-volume coding assistance inside Cursor or Grok Build, legal or engineering document work, and knowledge tasks that fit in a 500,000-token window. Be more careful with long, multi-hour terminal work, clinical reasoning, or anything where you need the strongest available coding model regardless of cost; there, SpaceXAI's own table points elsewhere.

The durable lesson from this week's two launches is that frontier pricing now spans nearly an order of magnitude, from $6 to $50 per million output tokens, for models that are often within a few benchmark points of each other. Picking a model is becoming an engineering decision rather than a loyalty choice: route by task, measure cost per completed task, and cache aggressively. That routing and measurement work is what we do, and it starts with running the candidates on your own work rather than on the vendor's table.