이미지: AI 생성 — METAL LAB
Summary
- xAI unveiled Grok 4.6 on the 12th. With an Artificial Analysis intelligence index score of 61, it matches GPT-5.6 Sol and enters the top tier, trailing only Claude Opus 5 and Fable 5.
- Its pricing — $2 per million input tokens and $6 per million output tokens — is less than half that of comparable models, and it completed agentic tasks using half the turns and a quarter of the tokens Opus 5 required.
- The gains came purely from strengthened post-training with the base model left unchanged, and with a larger-base Grok 4.7 already teased for release within weeks, the price-performance race is expected to intensify further.
An update that arrived just eight days after being teased
xAI unveiled Grok 4.6 on the 12th (local time). Alongside the announcement, it was immediately rolled out to the API, the coding tool Cursor, and xAI's own platform Grok Build, and it became instantly available through partner channels such as OpenRouter, Vercel, and Cloudflare. For the first week, xAI is also running a promotion doubling usage limits on Grok Build and Cursor.
This is the very model Elon Musk had teased earlier this month, saying it would arrive "sometime next week." For xAI, which is notorious for slipping roadmap deadlines, hitting the date this precisely was unusual.
Score of 61 — entering the top tier
Independent evaluator Artificial Analysis (AA) gave Grok 4.6 an intelligence index score of 61. That ties it with OpenAI's GPT-5.6 Sol, with only Anthropic's Claude Opus 5 (63) and Claude Fable 5 (62) ranking above it. It has effectively rubbed shoulders with the very top of the frontier.
The jump is especially steep: up 5 points from Grok 4.5 and 23 points from 4.3, and it has now also overtaken Kimi K3, previously known as a value-for-money leader. The catch is the price tag attached to this score.
| Model | AA Intelligence Index | Price (per million tokens, input/output) |
|---|---|---|
| Claude Opus 5 | 63 | $5 / $25 |
| Claude Fable 5 | 62 | — |
| Grok 4.6 | 61 | $2 / $6 |
| GPT-5.6 Sol | 61 | $5 / $30 |
| Kimi K3 | just below 61 | — |
The math behind the half-price tag
Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens — not a cent higher than 4.5. Compared with GPT-5.6 Sol or Opus 5 in the same score bracket, it's more than 60% cheaper. It's the same tutor delivering the same report card, but at half the tuition.
There is one quiet increase, though: the cache-hit rate rose from $0.30 to $0.50 per million tokens. According to AA's cost-per-task figure derived from running the full benchmark suite, it comes to $0.84 — matching Kimi K3's cost while delivering higher intelligence.
It gets even more impressive on agentic work
The real selling point of this version is agentic task performance. On GDPval-AA v2, a benchmark for practical knowledge-work tasks, it scored an Elo of 1753, placing second only to Opus 5. On Terminal-Bench v2.1, a terminal-task benchmark, it scored 88.4%. On τ³-Banking, a banking-operations simulation, it hit 50.7%, placing it at the top tier, and on AA-Briefcase, a comprehensive knowledge-work evaluation, it posted an Elo of 1577 — putting it on the same level as the leading models.
Its way of working stands out in particular. On long-running agentic tasks, Grok 4.6 produced results using an average of 53 turns and 500 million input tokens. Opus 5 needed roughly 103 turns and 2 billion tokens for comparable results. It's like an employee finishing the same errand in half the round trips and a quarter of the words — meaning the price gap widens even further on the actual bill.
Same body, different training recipe
According to foreign media analysis, Grok 4.6 keeps the same base model as 4.5 and swaps in an overhauled post-training process centered on SFT (supervised fine-tuning) and reinforcement learning. The context window remains unchanged at 500,000 tokens. Rather than redrawing the blueprint, xAI gained 5 points simply by changing the recipe.
The next step is already flagged. Word has also emerged of a roadmap for Grok 4.7, which will use a larger base model, arriving within weeks, and Musk has already declared it will surpass every existing model. Still, given xAI's track record of missed deadlines, whether it actually arrives on schedule this time remains the thing to watch.
Editor's take
The number that held my attention longest in this announcement wasn't the score of 61 — it was the 53 turns. When you run agents by piping several frontier models through their APIs, what actually determines cost isn't the price sheet but how many steps a model needs to finish a task. A model that takes twice as many turns for the same task ends up with a bill twice as large, even at an identical unit price. Grok 4.6 not only halved its unit price but also cut the number of round trips, so the real-world gap will likely feel far bigger than the benchmark table suggests.
The fact that a 5-point jump came from post-training alone is also worth dwelling on. Until last year, the standard way to raise scores was to scale the model up, but this year the frontier labs have all pivoted toward creating gaps through training technique at the same scale instead. It suggests the baseline capability has already leveled up across the board, and the real competition now plays out in how the model is "trained," not how big it is.
For teams here, this announcement is better read as "leverage for negotiation" than as "a reason to switch." In an environment where the price of frontier-level intelligence keeps dropping this fast within just six months, the moment you lock your pipeline to a single model, you start losing money. Building an adapter layer, maintaining your own evaluation set, and keeping prompts portable — lowering the cost of switching models — is right now the highest-return investment you can make. Indeed, the fact that DeepSeek struck back within hours with an even lower price sheet that very night shows this price war has only just begun.
