Grok 4.6 Guide: Long-Running Agents and Dearer Cache Reads

Grok 4.6 arrived on 12 August 2026, four weeks after Grok 4.5, and was itself succeeded by Grok 4.7 on 21 September. A model with under six weeks as flagship is easy to dismiss. It is still worth understanding, partly because any integration that pinned it now faces a migration decision, and partly because it introduced a pricing change that the current flagship kept.

A note on names: xAI, the company behind Grok, was renamed SpaceXAI in July 2026 after SpaceX acquired it. This guide uses the xAI name throughout.

The headline is simple: better results than Grok 4.5 on every benchmark xAI published, at the same input and output price. The catch is in the cached-input rate, and for agent workloads that catch can outweigh the headline.

What Grok 4.6 added over 4.5

xAI's announcement frames Grok 4.6 around "long-running agents and more ambitious interactive and visual work". The changes it describes:

• A longer supplemental training run with curated model-generated reasoning data, more engineering data and an improved optimiser and recipe.
• Supervised fine-tuning trajectories regenerated across effort levels, agent harnesses and domains, with problematic traces filtered by model-based checks.
• More self-testing and verification on long tasks, with the model checking its own work before moving on.
• Stronger first drafts on visual and interactive projects.
• Safeguards recalibrated for the stronger model.

On the API, the shape barely moved from Grok 4.5: a 500K-token window, text and image input, function calling, structured outputs, reasoning efforts from low to xhigh with high as the default, no batch support, and 150 requests per second at the time of writing. The one functional addition is that xhigh genuinely applies from 4.6 onwards; on 4.5 it was treated as high.

No parameter count, architecture details or open weights have been published.

The knowledge-work numbers

The launch table compares Grok 4.6 at high effort with Grok 4.5, GPT-5.6 Sol at max and Fable 5 at max. Its strongest results are on knowledge-work evaluations rather than coding:

• Artificial Analysis Intelligence Index: 61, up from 56 for Grok 4.5, level with GPT-5.6 Sol at 61 and just behind Fable 5 at 62.
• GDPval-AA v2: 1,753, first in the table.
• AA-Briefcase: 1,577, first in the table.
• Harvey LAB (legal work, run by Vals): 15.8%, first in the table.
• APEX-Agents: 57.5%, up from 47.1% for Grok 4.5.

Read the GDPval row with care. xAI says third-party figures are the best of self-reported or public results, and the competitor scores in its table do not all match figures published elsewhere for the same models. Compare a model's GDPval score only against numbers produced under the same conditions. GDPval explained covers what the benchmark measures and why its Elo-style scores shift between runs.

Where it trails

The same table shows Grok 4.6 behind on agentic coding:

• DeepSWE v1.1: 65.9%, against 73% for GPT-5.6 Sol and 70% for Fable 5.
• Terminal-Bench v3.0: 26%, against roughly 34% for both competitors.
• CursorBench v3.2: 69.9%, ahead of GPT-5.6 Sol but behind Fable 5.

A useful sanity check: in the later Grok 4.7 announcement, xAI lists Grok 4.6 on DeepSWE v1.1 at 65.2%, not 65.9%. A difference of under a point between two runs from the same vendor is normal, and it is a good reminder that small gaps in any benchmark table are noise. Terminal-Bench explained covers why long terminal tasks are where models separate.

Cached input got dearer

Grok 4.5 lists cached input at $0.30 per million tokens. Grok 4.6 lists it at $0.50, and Grok 4.7 kept that rate. Input and output prices did not change.

For a chat app this is a rounding error. For an agent it is not, because an agent loop re-sends its growing transcript every turn and, with caching working, most of that input is billed at the cached rate. Take one million input tokens where 90% are cache hits:

• Grok 4.5: 100K uncached at $2.00 plus 900K cached at $0.30 comes to $0.47.
• Grok 4.6: 100K uncached at $2.00 plus 900K cached at $0.50 comes to $0.65.

That is about 38% more on input for a cache-heavy loop, before you count any change in output tokens. If Grok 4.6 finishes tasks in fewer turns, it can still come out cheaper. That is exactly the thing to measure rather than assume. Prompt caching savings math walks through the break-even. Above 200K prompt tokens every rate doubles, cached included, so the gap widens in absolute terms.

xAI also offers a fast variant at twice the price. Its documentation for the current generation describes it as the same model on faster infrastructure, available through Cursor and Grok Build rather than the public API.

Grok 4.6 now that 4.7 exists

Grok 4.7 is listed at identical rates, including the $0.50 cached input, and xAI reports it ahead of 4.6 on every row of its launch table. There is no price reason to stay on 4.6.

The reasons to stay are operational. Grok 4.7 always returns encrypted reasoning on the Responses API, which changes how multi-turn requests must be handled, and a model swap always needs re-validation. If 4.6 meets your quality bar today, schedule the move rather than rushing it.

How to evaluate it

1. Pick ten knowledge-work tasks and ten agentic coding tasks from real work.
2. Run both sets on Grok 4.5, 4.6 and 4.7 at the same effort level.
3. Log cached and uncached input separately, plus output and turns per task.
4. Compute cost per accepted result. The cache-rate change only shows up when you split input this way.