Grok 4.6 matches GPT-5.6 Sol on frontier index, but the real story is agent economics

xAI released Grok 4.6 on August 12, and the headline number is a tie: the model matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite of nine benchmarks, at 61 points each. Fable 5 Max edges both at 62. That parity is notable, but it is not the part of the announcement that matters most for the industry.

The part that matters is the price. Grok 4.6 costs $2 per million input tokens and $6 per million output tokens, with a fast variant at double that. The model ships with a 500,000-token context window, no text output limit, and a knowledge cutoff of February 1, 2026. It is available immediately in Cursor, Grok Build, the xAI API, and through gateways including OpenRouter, Vercel, and Cloudflare. xAI is also doubling included usage inside Grok Build and Cursor for the first week.

This is a pricing statement disguised as a model release. The benchmark table in the official announcement shows Grok 4.6 winning or near-winning across agentic coding tasks: 69.9% on CursorBench v3.2, 65.9% on DeepSWE v1.1, 61.3% on FrontierCode v1.1 (Extended), and 57.5% on APEX-Agents. On DeepSWE, it beats GPT-5.6 Sol’s 73%? No, it does not. GPT-5.6 Sol scores 73% there, and Fable 5 Max hits 70%. Grok 4.6 trails both on that specific benchmark. The honest read of the table is that the frontier has compressed: the gap between the best and fourth-best model on most of these evals is now single digits.

What distinguishes Grok 4.6 is not raw capability. It is the training recipe and the economic posture. The announcement describes a longer supplemental training run than Grok 4.5, using curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer. Then xAI used Grok 4.5 to regenerate SFT trajectories across reasoning efforts, agent harnesses, STEM, software engineering, and knowledge work, filtering out problematic traces with model-based checks. The agentic RL stage covered knowledge work, general coding, kernel optimization, web development, and computer-aided design.

That is a deliberate bet on long-horizon tasks. The model is built for agents that stay alive across many steps, not for single-shot Q&A. The announcement says Grok 4.6 produces stronger first passes on visual and interactive projects than Grok 4.5, establishing structure and visual language in one pass. On longer trajectories, xAI says it observed more self-testing and verification, with the model checking its own work before moving on.

The economics of that bet are the real story. A long-running agent burns tokens continuously. At $2/$6 per million, a 500,000-token context window that gets rewritten every few steps becomes a serious cost line. xAI’s pricing undercuts the competition on raw tokens, but the docs add a crucial caveat: xAI recommends setting a prompt_cache_key to route a conversation’s requests to the same server, making cache hits reliable. Without it, the docs warn, you often pay full input price on a cache-cold server. That is a quiet admission that the headline price assumes disciplined caching discipline. Builders who ignore it will see effective costs multiply.

The docs also mention context compaction for long agent loops and a reasoning effort parameter that goes up to “xhigh.” These are the mechanics of agentic cost control, and they are where the margin war will be fought. The model that can sustain a 10,000-step agent loop at the lowest effective cost per completed task wins the developer workflow, regardless of who tops the composite index.

xAI’s safety section is worth reading closely. The announcement says safeguards were “improved and calibrated in line with the model’s capabilities,” with the widest-ever suite of pre-deployment testing and extensive post-deployment and third-party testing. The framing is notable for what it permits: vulnerability patching, accelerating the engineering design cycle, and augmenting AI research. That is a direct appeal to security teams and defense-adjacent buyers, and it signals that xAI is comfortable positioning Grok 4.6 for dual-use work that other labs might gate more heavily.

The Harvey LAB (Vals) benchmark, a legal-domain eval, shows Grok 4.6 at 15.8% versus GPT-5.6 Sol’s 2.5% and Fable 5 Max’s 11.3%. That is a striking outlier. On legal knowledge work, Grok 4.6 is more than six times better than GPT-5.6 Sol by this measure. Either the benchmark is flattering to Grok’s training mix, or the model genuinely generalizes better into professional knowledge domains. The legal vertical is high-margin and high-stakes, and a frontier model that scores well there is a commercial wedge, not just a research curiosity.

The competitive picture is now clear. The top of the market has four models within a few points of each other on composite intelligence: Grok 4.6, GPT-5.6 Sol, Fable 5 Max, and Grok 4.5 itself at 56. Differentiation is shifting from raw IQ to three axes: price per token, agentic reliability over long horizons, and vertical strengths. xAI is competing hard on the first, credibly on the second, and opportunistically on the third.

For builders, the practical implication is that model choice is becoming a cost-engineering decision. The days of picking the single best model and paying whatever it costs are ending. A 500,000-token context window at $2 per million input tokens makes certain agent architectures viable that were not viable six months ago. Long-running research agents, codebase-wide refactoring loops, and iterative design workflows all become cheaper to operate. But the caching caveat means the burden shifts to the developer: prompt_cache_key discipline, context compaction, and careful reasoning-effort tuning now determine whether the advertised price is real.

The agentic coding market is the battleground, and Cursor is the distribution channel. xAI’s decision to make Grok 4.6 available on all Cursor plans, not just premium tiers, is a direct shot at the incumbents. CursorBench v3.2, where Grok 4.6 scores 69.9% against GPT-5.6 Sol’s 67.2%, is the benchmark that Cursor users will actually feel. The model that makes a coding agent complete a task in fewer turns, with fewer failed tool calls, wins on total cost even at a higher per-token price.

What xAI has not disclosed is the training compute. The announcement mentions an improved optimizer and training recipe but gives no FLOP count, no cluster size, no timeline. That opacity matters because the longer supplemental training run and the expanded RL task mix imply a substantial compute bill. If xAI is spending more per model and charging less per token, the question is whether the unit economics work at scale, or whether this is a land-grab priced below cost.

The answer will come in the next earnings cycle or in xAI’s infrastructure announcements. For now, the observable fact is that a frontier model with a 500,000-token context window, no output limit, and competitive agentic benchmarks is being sold at $2/$6 per million tokens. That is a price that compresses margins across the entire model-as-a-service layer, and it forces every other lab to respond on price, not just on capability.

The tie on the AA Intelligence Index is the least interesting number in the announcement. The interesting number is $6 per million output tokens, because that is the price at which long-running agents become economically sane for a much wider class of applications. The model that wins the agent era will not be the one with the highest benchmark score. It will be the one whose effective cost per completed task, after caching, compaction, and reasoning tuning, is lowest. xAI has made its bid. The market gets to verify it.