Grok 4.6 Reaches the AI Frontier Without Raising Its Price

The interesting thing about xAI's Grok 4.6 release is not that the model got better. It is that the bill did not.

Grok 4.6 arrived at $2 and $6 per million input and output tokens, exactly what its predecessor cost a month ago, while independent evaluator Artificial Analysis moved its Intelligence Index score up five points to 61 — a composite of nine benchmarks that now places xAI level with GPT-5.6 Sol, two points under Claude Opus 5, one under Claude Fable 5, and a step ahead of Kimi K3. Measured against Grok 4.3, the jump is 23 points.

Frontier capability gains have generally arrived with frontier prices attached, so a generation that holds rates flat is worth pausing on. The two models sitting within two points of Grok 4.6 charge $5/$25 and $5/$30 respectively, and because output tokens dominate spending on anything reasoning-heavy, that gap translates into a measured cost per task of $0.84 and a position on the intelligence-versus-cost Pareto frontier.

For buyers, that changes the shape of a familiar decision. The trade-off in agent deployments has usually been posed as capability against budget: pick the strongest model and absorb the cost, or accept a weaker one and run more of it. A model that lands within a couple of points of the leaders at roughly a third of the output price makes that framing harder to sustain. Whether the two points matter is now an empirical question for each workload rather than an assumption.

Built for tasks that do not finish in one turn

xAI framed the release around long-running agents and more ambitious interactive work, and the evaluation numbers follow that framing rather than cutting against it. The agentic scores are where the model separates itself: an Elo of 1753 on GDPval-AA v2, second only to Claude Opus 5 and statistically tangled with Claude Fable 5 and Qwen3.8 Max; 50.7 percent on the tool-using customer service benchmark τ³-Banking, a hair behind Qwen3.8 Max at 51.3 percent; and 88.4 percent on Terminal-Bench v2.1, level with the field's leaders.

Few models manage to be simultaneously competitive at knowledge work, customer service and terminal operation, which is what puts Grok 4.6 on the cost-performance frontier for every agentic evaluation in the index rather than just one of them. Breadth of that kind is harder to engineer than a single headline score, because the three categories reward different things: judgment, patience with a human counterpart, and precision against an unforgiving shell.

Its debut on AA-Briefcase, a private benchmark of long-horizon knowledge work, produced an Elo of 1577 — respectable rather than leading, and consistent across rubric grading, presentation and analytical quality rather than lopsided. The efficiency figures underneath that score are more revealing. Artificial Analysis clocked the model finishing tasks in roughly 53 turns and half a billion input tokens on average, where Claude Opus 5 at maximum effort took about 103 turns and two billion.

Long agentic runs accumulate context fast, so halving the turns and quartering the input volume compounds into a cost advantage well past what per-token pricing alone suggests. It also shortens wall-clock time, which is the constraint that actually bites when an agent sits in a product loop with a person waiting at the other end.

What changed underneath

The recipe, by xAI's account, was refinement rather than reinvention: a longer supplemental training run than Grok 4.5 received, fed with curated model-generated reasoning data and high-quality engineering data, plus a reworked optimizer. The previous model was then put to work regenerating supervised fine-tuning trajectories across different reasoning efforts and agent harnesses, with model-based checks discarding bad traces before they reached the next stage.

Reinforcement learning ran across domain-specific environments — kernel optimization, web development and computer-aided design among them. xAI reports that on longer trajectories the model began verifying its own output before continuing, which is the sort of behavior that decides whether a fifty-step task ends in a result or a mess. The company says safeguards were recalibrated alongside those capabilities, with its widest pre-deployment testing suite to date.

The caveats worth keeping

Artificial Analysis is one evaluator, and a composite index compresses nine different things into a single number that no individual workload resembles. Parity on an index is not parity in production, and the gap between a benchmark harness and a messy internal codebase has embarrassed better scores than this one. The turn-efficiency figures, likewise, describe averages over a private task set rather than a guarantee.

What is harder to argue with is the cadence. Roughly a month separated Grok 4.5 from Grok 4.6, and the release moved the score without moving the price. If that pattern holds across the field, the interesting competition over the next year may be less about who tops an index than about who can hold a position on it while charging less to stay there.

Where to find it

Cursor and Grok Build carry the model now, with API access through OpenRouter, Vercel and Cloudflare, and double the included usage in Grok Build and Cursor through the first week. The context window holds at 500,000 tokens. One number did rise: cached input now bills at $0.50 per million tokens rather than $0.30. A faster variant costs twice the standard rate.