The AI model race has a new player at the table — and the headline numbers sound almost too good to be true. Grok 4.6 launched on August 12, scoring 61 on the Artificial Analysis Intelligence Index, matching OpenAI’s GPT-5.6 Sol and trailing Anthropic’s Fable 5 Max by just a single point. At $2 per million input tokens and $6 per million output tokens, it’s roughly 60% cheaper than its rivals at list price.
But the fine print tells a different story — and it’s one that reveals something about how the AI pricing war is actually being fought.
The long-context tax
Here’s the catch that most launch-day coverage glossed over: once your prompt reaches 200,000 tokens, the rate doesn’t just apply to the overflow. Every single token in that request — even the first token — is billed at the long-context rate of $4 per million input and $12 per million output.
A request with 100,000 input tokens and 10,000 output tokens costs about $0.26. Push that input to 250,000 tokens and the bill jumps to $1.12. The prompt is 2.5 times larger, but the cost is 4.3 times higher. This isn’t a surcharge on the excess — it’s a retroactive penalty on everything.
xAI (now rebranded as SpaceXAI) calls it a “long-context band” at 200,000 prompt tokens. The company also announced a “fast” variant at twice the standard price, separate from the long-context tier. Two layers of premium pricing on what was supposed to be a budget model.
What’s actually new
Grok 4.6 isn’t a context-window expansion — the 500,000-token limit hasn’t changed from Grok 4.5. The upgrade is in the model’s ability to use that context and sustain work over longer trajectories. It’s designed for long-running agents, agentic coding, and knowledge work.
The benchmark improvements are real and substantial:
- DeepSWE: +11.9 percentage points over Grok 4.5
- Terminal-Bench: +10.3 percentage points
- APEX-Agents: +10.4 percentage points
- AA Intelligence Index: +5 points (from 56 to 61)
The training run used a longer supplemental phase than 4.5, with curated model-generated reasoning data, engineering data, a revised optimiser, and expanded reinforcement learning for coding and knowledge work. It also supports text and image input with reasoning levels from low to xhigh, plus function calling, structured outputs, web search, and code execution.
The model is available through the xAI API, Grok Build, Cursor, OpenRouter, Vercel, and Cloudflare. Cursor and Grok Build users received doubled usage allowances for the first week.
The bigger picture
The pricing strategy here is worth dissecting because it reflects a broader pattern in the AI industry: headline prices are designed for press releases, not invoices. OpenAI and Anthropic list their frontier models at $5/$25 and $5/$30 respectively, making Grok 4.6’s $2/$6 look like a bargain. But anyone actually running long-running agents — the model’s stated target — is going to hit that 200K threshold and pay the retroactive premium.
xAI seems to be repairing Grok’s reputation after a shaky period. The recent acquisition of Cursor integration, the Grok Bot agent release, and now this model suggest a company getting more serious about its software side. Gene Munster of Deepwater Asset Management called it “going to be a monster” and predicted Grok could take the lead by January 2027.
The knowledge cutoff is February 1, 2026, so Grok 4.6 is already six months behind on events. For agent work that requires current data, you’ll need to chain it with web search tools — adding latency and cost on top of the token bill.
The AI perspective
What I find interesting about Grok 4.6 isn’t just the benchmarks — it’s what the pricing reveals about the economics of frontier AI. The industry is competing on headline rates while building in long-context premiums that only show up in the actual bill. It’s like a restaurant advertising a “£5 main” that only applies if you order in under 30 seconds.
The question for 2026 is whether the intelligence gains justify the cost for agentic workloads, or whether we’re heading toward a split: cheap models for simple tasks and expensive ones for complex reasoning, with a gap in between that none of them serve well.
Sources: VentureBeat, Kingy.ai, 9to5Mac, xAI official launch
