Grok 4.6 Is a 2 Dollar Model Until Token 200,001

Cached input rose 67 percent. And the launch table's clean rows are the losses.

SpaceXAI and Cursor released Grok 4.6 on August 12, thirty five days after Grok 4.5. It is not a bigger model. The foundation was held constant and the work went into post training: a longer supplemental run on curated model generated data, then Grok 4.5 regenerating the supervised fine tuning trajectories across reasoning efforts and agent harnesses, with problematic traces filtered by model based checks, then reinforcement learning in agentic environments spanning knowledge work, general coding, web development, computer aided design, and kernel optimization.

The pitch is long horizon autonomy. The model is described as staying with a task across many steps, and on longer trajectories the team reports seeing more self testing and verification before the model moves on. The model page lists 500,000 tokens of context, a February 1, 2026 knowledge cutoff, text and image input with text output, no stated output limit, and a reasoning ladder that adds a fourth rung above the three Grok 4.5 shipped with: low, medium, high as the default, and xhigh. No parameter count was published.

Headline pricing is 2.00 per million input tokens and 6.00 per million output tokens, identical to Grok 4.5. Nearly every writeup led with that. It is true, and for a long running agent it is close to the least useful number on the page.

A token price is three numbers, not one. Input, cached input, output. The published rate table holds two of them and moves the third. Cached input goes from 0.30 per million on Grok 4.5 to 0.50 on Grok 4.6, about 67 percent higher, at both tiers.

Then there is the second tier itself. Both models carry a long context threshold at 200,000 prompt tokens, above which the rates double. The billing rule is the part worth reading twice: once a prompt reaches the threshold, the long context rates apply to every token in that request, not to the excess above it.

Read those two facts together and the shape of the release changes. The context window is 500,000 tokens. The headline rate covers the first 200,000. Three fifths of the window that the marketing sells is billed at the second tier, and crossing into it is not gradual. It is a cliff, priced on the whole request.

A chat request is small and stateless. A long running agent is neither. It accumulates a system prompt, a tool schema, a growing transcript of tool results, and file contents, and it replays most of that on every step. That workload is cache heavy by construction, and its prompt grows monotonically until something truncates it. Both of the rates that moved are the rates that workload pays.

The model page says as much, in the form of advice. It recommends setting a prompt cache key, because without one a conversation scatters across servers, cache hits stop being reliable, and requests land at full input price on a cold server. It also points long agent loops at context compaction. Both notes describe the same underlying reality: the default behavior of a long loop is expensive, and staying under the threshold is something you engineer rather than something you receive.

Server side tools bill separately from tokens, and the vendor states plainly that the agent decides how many calls to make, so cost scales with query complexity. There is no batch reduction listed for this model, and priority processing multiplies every token type by two.

Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, a composite of nine benchmarks. That ties GPT-5.6 Sol Max and sits one point under Fable 5 Max, and it is five points above Grok 4.5. Every row in the launch table improves over 4.5, several of them by more than ten points. That part is real and it is the reason the release landed the way it did.

The rows are not equally informative, though. A vendor picks its comparison set, its benchmark versions, and its bolding. What a vendor cannot pick is which rows it loses. Sorting the table by evidence quality rather than by result puts the useful signal at the bottom.

Two of the three bolded wins are statistical ties. On the elo style rows, GDPval-AA v2 and AA-Briefcase, the gaps of twelve and three points sit inside the confidence intervals Artificial Analysis publishes, which means they are not leads in any sense a procurement committee should act on. The comparison set also leaves out Claude Opus 5, which currently sits above every model in the table on that index.

The losses, meanwhile, are the least flattering numbers the vendor could have printed, which is exactly what makes them worth trusting. DeepSWE v1.1 at 65.9 against 73.0, and Terminal-Bench v3.0 at 26.0 against 34.6, are both agentic software engineering evaluations, and both are the workloads most engineering teams adopt a coding model for.

CursorBench deserves its own paragraph, and credit where it is due. When Cursor launched Grok 4.5 in July, it published a footnote disclosing that the model held an advantage on CursorBench because an earlier snapshot of the Cursor codebase had been included in training by accident. The team said the exact impact was unclear, that the data had been removed for future models, and that a larger update to CursorBench was in progress, which is why the benchmark was left out of the 4.5 charts entirely.

In the Grok 4.6 table, CursorBench is back, at version 3.2, with 69.9 for the new model against 66.7 for the old one. The v3.2 label is consistent with the promised refresh, and scores on it are not comparable with the previous version. The 4.6 post carries no footnote at all.

Which leaves a question the launch material does not answer, and it is worth stating as a question rather than an accusation. The commitment was that the contaminated snapshot had been removed for future models. Grok 4.6 is described as a post training upgrade on the same foundation as Grok 4.5, not a new base model. If the snapshot entered through pretraining, removing it would mean retraining the base. The published material does not say whether that happened, and it does not say what the exclusion criteria for the v3.2 refresh were. Until it does, the honest read of the 3.2 point gain on that row is that its provenance is unresolved.

The capability half of the launch post is measured in ten benchmark rows. The safety half is measured in adjectives. Safeguards are described as improved and calibrated to capability, the pre deployment testing suite as the widest ever, the post deployment third party testing as extensive. No third party is named. No evaluation is named. No number appears anywhere in that section.

There is no model card at launch. That gap is real, and it is also temporary in the recent pattern: Grok 4.5 shipped July 8 and its model card arrived July 14, six days later. So the fair criticism is not that this vendor refuses to document its models. It is that documentation lands after the adoption window, and the adoption window is when the choice gets made. Doubled first week usage in Cursor and Grok Build, plus a reduced launch rate, both expire on roughly the same clock the card takes to appear.

One more item belongs in the gap column. The most interesting behavioral claim in the post, that on longer trajectories the model does more self testing and verification before moving on, is a vendor observation from internal testing. It is not an independently measured result, and no eval in the table isolates it. For anyone running unattended agents, that specific behavior is the one you would most want a number attached to, because self verification is the mechanism that decides whether a long run recovers from its own mistakes or compounds them.

Several things cut against the reading above, and they are worth stating at full strength. The long context tier is published, documented, and applied identically on Grok 4.5, so nothing here is concealed. It is a listed rate that most coverage skipped, which is a different failing and mostly ours. Cached input at 0.50 remains a 75 percent reduction against the uncached rate, so caching still pays, it pays less than it did. VentureBeat's read is that the API price sits below half of what GPT-5.6 Sol charges in standard mode, so the cost comparison can absorb a tier change and still land favorably.

The capability gains are also not in dispute. Every published row improves, DeepSWE by 11.9 points, APEX-Agents by 10.4, Terminal-Bench by 10.3, the composite index by 5, on a foundation that did not change. That is a strong post training result by any standard, and the model leads or matches the listed GPT scores on most rows in the table. And the CursorBench contamination note exists at all because Cursor wrote it voluntarily, in a footnote, about its own model, which is more disclosure than most vendors publish about their own benchmarks.

What would change the assessment: a Grok 4.6 model card with named third party evaluators and numbers attached to the self verification claim, and one sentence in the CursorBench row explaining what the v3.2 refresh removed. Both are inexpensive to publish. On the 4.5 timeline the first should land around August 18.

The number a vendor holds constant is the number it wants quoted. On a model sold for work that runs for hours, the rates that govern the bill are the cached rate and the threshold, and both of those moved while the headline stood still.