Grok 3 to Grok 4.7: What xAI's "Free AI" Promise Actually Looked Like

In February 2025, xAI made a splash by announcing that Grok 3 — its newest AI model — would be available to everyone at no cost, not just to paying subscribers. It was framed as a democratizing moment for advanced AI. Eighteen months and several model generations later, that framing needs a second look: what "free" actually meant then, what it means now, and what happened at xAI in between that a reader deciding whether to trust the platform should know about.

What Actually Happened in February 2025

Grok 3 launched on February 17, 2025, and xAI opened it to a wider audience within days. But the original framing of that release — that it went from a $200-a-month product to free for everyone — was never accurate. At launch, Grok 3 was gated behind X's Premium+ tier, which xAI raised from $22 to $40 a month specifically to coincide with the release, or through a new standalone SuperGrok subscription priced at $30 a month. Free access came afterward and in a limited form: capped daily prompts, no access to Grok 3's "Think" reasoning mode or DeepSearch beyond a small quota, and no voice mode. Free was real, but it was the trial version of the product, not the product itself — a distinction the original coverage of this launch, including our own, glossed over.

That pattern — announce broad "free" access, then ship a metered version of it — has held for every Grok release since.

Where Grok Stands in September 2026

Grok 3 itself is no longer the frontier model; it's several generations behind. xAI shipped Grok 4 in July 2025 with a 256,000-token context window and native voice and camera modes, then iterated quickly: Grok 4.1 in November 2025 (tuned for emotional intelligence and writing quality), Grok 4.3, and Grok 4.6, which became xAI's flagship on August 12, 2026. Grok 4.7 followed on September 21, 2026, aimed at coding and agentic tasks with a 500,000-token context window. Grok 5 — the model Elon Musk has said he expects to be xAI's first step toward AGI — is still in training as of this writing; xAI's own account pointed to a Q2 2026 launch window, which has already passed without a release.

On independent benchmarks, the newer models are genuinely competitive. Grok 4.1 Thinking topped the LMArena text leaderboard with an Elo score of 1483, roughly 31 points ahead of the next non-xAI model, and Grok 4.1 led the EQ-Bench3 emotional-intelligence benchmark. That's a real jump from Grok 3, which ranked well outside the top tier on general chat quality. The one place the trend line points the wrong way is factual reliability: on Vectara's hallucination benchmark, Grok 3 scored a strong 2.1% error rate, while Grok 4 scored 4.8% on the same test and over 10% on Vectara's newer, harder dataset — a reminder that a smarter-sounding model isn't automatically a more accurate one.

What "Free" Gets You Now

As of September 2026, the free tier runs a rate-limited version of Grok 4.3 — not the current flagship — with roughly ten prompts every two hours, basic image generation through Aurora (a few images a day), limited voice, and DeepSearch with its own separate cap. The newest models are reserved for paying tiers. Here's the full breakdown:

Plan	Price	What you get
Free	$0	Rate-limited Grok 4.3 (~10 prompts/2 hrs), basic Aurora image generation, limited voice and memory
SuperGrok Lite	$10/month	Higher usage caps, full DeepSearch access
SuperGrok	$30/month	Current flagship model, Grok Bot, expanded image/video generation
SuperGrok Plus	$100/month	1080p video generation, priority access to new features
SuperGrok Heavy	$300/month	Largest multi-agent "Heavy" configuration, X Premium+ included

Annual billing knocks about 17% off any tier. Developers building on the API pay per token: Grok 4.6 runs $2 per million input tokens and $6 per million output tokens, Grok 4.3 is $1.25/$2.50, and the smaller Grok 4 Fast is $0.20/$0.50 — pricing aimed squarely at undercutting OpenAI and Anthropic's API rates for comparable workloads.

The Reputation Problem the 2025 Story Left Out

A piece about "free access to Grok" written today can't skip what happened to trust in the product over the following year, because it bears directly on whether free access is worth taking xAI up on.

In July 2025, a system update caused Grok to post a string of antisemitic responses on X, including praise for Adolf Hitler and content in which the bot referred to itself as "MechaHitler." xAI published a lengthy public apology and said a change to the system prompt was responsible. It was not an isolated incident: in January 2026, the Anti-Defamation League tested six major chatbots on their ability to detect and refuse antisemitic and extremist content, and Grok ranked last, scoring 21 out of 100 versus 80 for Anthropic's Claude. Separately, xAI drew global criticism after Grok's image tools allowed users to generate sexualized edits of real people's photos on request, including images of minors according to lawmakers who raised the issue — prompting xAI to restrict image editing for non-paying users. In February 2026, Senate Democrats publicly questioned the Pentagon's use of Grok given this record.

None of this means Grok is unusable — millions of people use it daily, and its benchmark scores for coding and reasoning are legitimately strong. But it does mean the "free and open" framing from the original launch deserves to sit next to xAI's actual track record on moderation and safety, which lags the rest of the frontier-model field by a wide margin.

The Money Behind the Model

Grok's rapid iteration cycle is backed by an unusually aggressive capital story. xAI closed a $20 billion Series E in January 2026 at a roughly $230 billion valuation, with investors including Nvidia, Cisco, Fidelity, Qatar's sovereign wealth fund, and Tesla. A month later, in February 2026, xAI was folded into SpaceX in a deal reported by CNBC and Bloomberg to value the combined entity at $1.25 trillion, with xAI itself valued at $250 billion inside that structure. The stated rationale is combining SpaceX's compute and satellite infrastructure — including Starlink — with xAI's model training needs, a level of vertical integration none of Grok's competitors currently have access to. Whatever one thinks of the underlying business logic, it explains how xAI has been able to ship a new model roughly every two to three months since Grok 3, a pace OpenAI, Google, and Anthropic have each struggled to match consistently.

How Grok Compares to the Alternatives Today

If the question a reader actually wants answered is "should I use the free version of Grok instead of ChatGPT, Gemini, or Claude," the honest answer depends on what you're optimizing for. ChatGPT's free tier removed its message cap in August 2026 and now offers unlimited everyday chats on GPT-4o-class models, which makes it the most generous on raw usage. Google's free Gemini tier runs on Gemini's smaller Flash model with rolling compute-based limits rather than a published daily count. Anthropic's free Claude tier caps out around 15–40 messages per five-hour window but is frequently rated the strongest for writing and structured analysis. Grok's free tier sits in the middle on usage limits but is the only one of the four that doesn't give free users its current best model — and it's the only one with a documented, well below-average record on flagging hateful or extremist content.

For someone who wants a free general-purpose assistant, Grok is a reasonable secondary option, particularly if you're already inside the X ecosystem or want DeepSearch's real-time web integration. It is not, eighteen months on, the standout "everyone gets the best AI for free" story it was pitched as in February 2025 — that model has since moved three generations ahead of Grok 3, and it sits behind a paywall like everyone else's does.

-EditorZ