Grok 4 and Grok 4.20 — Benchmarks, Pricing, and an Honest Model Comparison

Overview of xAI's Grok Model Family

Grok 4 (base) is a proprietary multimodal model from xAI, released in July 2025. It supports text and image input, outputs text, and works with a context window of 256 thousand tokens. For the most complex tasks, it uses a multi-agent "Heavy" mode, where several copies of Grok discuss and combine an answer.

Grok 4 Heavy is a configuration of several agents — in early reports, five or more — and it's built for deep research and long reasoning. It's available as SuperGrok Heavy at a corporate price, around 300 dollars a month or 3,000 dollars a year on some plans.

Grok 4.20 (the 2026 flagship) is billed as xAI's cutting-edge model in 2026, with a 2-million-token context, improved programming, and a better price-to-quality ratio. It targets the SWE-bench, MMLU, and coding benchmarks while remaining cheaper than the top-tier GPT-5.x models.

Benchmarks and Intelligence of Grok 4 and Grok 4.20

According to aggregators, the base Grok 4 shows an average score of around 60–62% in some tallies. The model is strong on MMLU-Pro and full MMLU — that is, in advanced knowledge and reasoning — as well as on math and complex logic tests like ARC-AGI and AIME-style problems, where it often ranks among the leaders.

One aggregator reports high scores for Grok 4 — around 95% on HELM IFEval for instruction following, around 94% on Fiction. LiveBench for creative and narrative tasks, and over 85% on HELM MMLU-Pro. A separate composite metric, the Artificial Analysis Intelligence Index, gives Grok 4 a score of 33, which is above average among comparable models across reasoning, knowledge, math, and coding.

A 2026 benchmark review from Tokenmix reports that Grok 4.20 scores 78% on SWE-bench, 91.2% on full MMLU, and 1,385 Elo on the coding arena, placing it in the front ranks for programming. The model's context window reaches 2 million tokens, which is roughly twice that of some competitors. At the same time, Tokenmix notes that Grok 4.20 comes in third after GPT-5.4 in pure accuracy, but leads in context volume and is roughly 60% cheaper in output cost.

Source: Tokenmix benchmark

Pricing and Cost-Efficiency of the Grok Models

Different sources cite different prices depending on the plan and version, but several clear patterns are visible. According to early reports, the API cost for Grok 4 was estimated at up to about 5.50 dollars per 1 million input tokens and 27.50 dollars per 1 million output tokens, which is noticeably more expensive than the current public rates of around 3 dollars per 1 million input and 15 dollars per 1 million output tokens. The context window, meanwhile, is on the order of 256 thousand tokens.

By subscription, the SuperGrok plan with Grok 4 is listed in a number of reviews at roughly 300 dollars a year or 30 dollars a month. The SuperGrok Heavy plan with Grok 4 Heavy is estimated in community discussions at 300 dollars a month or 3,000 dollars a year, and it's designed for advanced users and companies.

A separate comparison is worth making between Grok 4.20 and GPT‑5.4. According to aggregators, the API for Grok 4.20 costs around 2 dollars per 1 million input tokens and 6 dollars per 1 million output tokens, with a context of up to 2 million tokens. Analytical reviews note that Grok 4.20's output cost is roughly 60% lower than that of comparable models like GPT‑5.4, with similar accuracy and a larger context. This makes Grok 4.20 one of the most cost-efficient cutting-edge models of 2026 at comparable quality.

Grok's Strengths and Weaknesses by Task

Grok 4 is a model that feels especially at home where rigorous logic is required. According to reviews and feedback, its strong suit is tasks with a clear structure. For example, analytics, complex "why is it built this way" questions, document analysis, working with up-to-date data from the internet, and interdisciplinary queries where you need to connect several fields of knowledge. The model was originally trained not just to "memorize texts" but to use tools — search, code, a calculator — so Grok more often behaves like a researcher: it checks facts, turns to external sources, and holds a long context better than many classic LLMs. This makes it convenient for deep analysis, business analytics, and situations where what matters isn't elegant phrasing but a well-grounded conclusion.

Grok's second strong area is reasoning and math problems. On benchmarks like ARC‑AGI, the model has shown results that reviews describe as "the first public AI to cross the threshold of conditional living reasoning," where it didn't just guess answers but searched for patterns and solved new types of problems. In demonstrations and real-world tests, Grok confidently handles chains of logic and complex "understanding the meaning" questions, helps break down technical topics, and can adjust its answer style to the request — from dry analytics to a livelier explanation. For those who work with texts and data, it feels like a "smart analyst conversationalist" rather than just a phrase generator.

But Grok also has its weak spots. Compared with the leading competitors, the model looks less strong on purely creative tasks. For example, its texts can be useful and well-structured, but not always as vivid and "literary" as those of models tuned for creativity. In multimodal scenarios and image generation, Grok also often loses to flagships like GPT and Gemini, where it has a harder time precisely following visual prompts, maintaining fine details, and neatly arranging objects in a picture. Some reviews also note the political bias of the base version, where on sensitive topics the model noticeably leans on Elon Musk's stance and related sources, which doesn't suit everyone for neutral analytics.

Another nuance is specialization and task balance. Grok 4 was conceived as a tool for complex research, programming, and analysis — hence the emphasis on logic, up-to-date data, and working with large volumes of information. This makes it a strong choice for experiments, business cases, and deep analysis, but not always the best option if you need quick creativity, complex multimodality, or a maximally neutral tone on sensitive topics. In the end, Grok is most often chosen as a "brain for serious tasks" rather than a universal storyteller for all of life's occasions.

Want to Try Grok Yourself

As you can see, Grok 4 and Grok 4.20 are serious models for math, science, complex reasoning, and working with a large context, and version 4.20 is also noticeably more economical than many competitors. The choice between them and the GPT or Claude models depends on your task and budget, and the best way to understand the difference is to test them on your own real-world queries.

That's exactly what Unitool.ai exists for, where all the leading neural networks are available under a single subscription in one place. You don't need to sign up for separate xAI, OpenAI, and other services' plans — just open the platform and switch between models in a couple of clicks. Sign up at Unitool.ai, choose Grok, and start comparing models on your own tasks today.