Quick answer. Grok 4.6 is a modest upgrade over Grok 4.5. Same 500K context, same $2/$6 pricing, better SWE-bench (95.6% vs ~86%) and reasoning. But cached input costs 67% more, time to first token roughly doubled (14.0s to 27.8s on Artificial Analysis), and agentic coding and SkillsBench both regressed. Upgrade for one-shot coding; test carefully for long-running agents.
Grok 4.6 arrived on August 12, 2026, just 35 days after Grok 4.5. Short gaps usually mean incremental change, and that is broadly what happened — but the changes are not uniformly improvements, and a couple of them will cost you money if you do not notice.
What stayed the same?
More than you might expect for a version bump:
- Context window: 500,000 tokens, unchanged
- Headline pricing: $2 per million input, $6 per million output — unchanged
- Modality: text and image in, text out
- Regions: us-east-1 and us-west-2 only, still no EU
xAI disclosed no architecture details for either model, so whether 4.6 is a new base model or 4.5 with extended post-training is unknown. Claims either way are speculation.
What actually improved?
SWE-bench, substantially
The strongest verified result in the release. Vals AI measures Grok 4.6 at 95.60% on SWE-bench, roughly nine points above Grok 4.5. That is a real gain on a benchmark that matters for practical software work.
Reasoning and general intelligence
Grok 4.6 (High) scores 44.31 on Artificial Analysis's Intelligence Index v4.3.2 (October 2026), against Grok 4.5's 38.81 on the same version — a genuine 5.5-point gain, and the clearest general-intelligence improvement in the release.
A correction to what this article said in August: we quoted 60.92 against xAI's claimed 61, and concluded Grok 4.6 was effectively tied with GPT-5.6 Sol Max. Those were index v4.1.1 figures. AA has since rebased to v4.3.2, which changes the benchmark basket and moves every model's score — there is no conversion factor between the two. On v4.3.2 the tie does not survive: Sol Max scores 46.97 to Grok 4.6's 44.31, a 2.66-point gap, and Grok 4.6 also stays behind Claude Opus 5 Max (50.78) and Claude Fable 5 (49.63).
A new reasoning level
Grok 4.6 adds xhigh to the reasoning_effort options. Grok 4.5 accepted the parameter but silently downgraded it to high, so this is the first release where the setting does anything.
Non-agentic coding
LiveBench's coding score rose by more than six points. For single-shot code generation — write this function, fix this bug, explain this file — 4.6 is meaningfully better.
What got worse?
This is the part that gets skipped in most upgrade write-ups.
Agentic coding regressed
| Metric | Grok 4.5 | Grok 4.6 | Change |
|---|---|---|---|
| LiveBench agentic coding | 56.5 | 54.2 | −2.3 |
| SkillsBench | 66.03 (1st) | 55.77 | −10.3 |
Both moved the wrong way, and the SkillsBench drop is large — Grok 4.5 ranked first on that board. Because non-agentic coding rose over the same period, this is a specific regression in multi-step tool-using work rather than general degradation.
It also sits awkwardly next to xAI's own launch framing about a focus on long-running agents. If you use Grok inside an agent loop, benchmark your own workload before upgrading rather than assuming a newer version is better.
Latency got much worse
Time to first token went from 14.0 seconds on Grok 4.5 to 27.8 seconds on 4.6 — roughly twice as long (Artificial Analysis, High tier on both sides, re-read 5 October 2026). Throughput moved the other way: median output speed improved from 59.6 to 72.9 tok/s. These are two different measurements and a model can be faster on one while slower on the other, so be careful which one a comparison is actually quoting. Here the regression is the first-token wait, and for interactive tooling a near-half-minute pause before anything appears is a genuine usability problem.
Cost per job went up, even though the price did not
| Grok 4.5 | Grok 4.6 | |
|---|---|---|
| Input / output per 1M | $2 / $6 | $2 / $6 |
| Cached input per 1M | $0.30 | $0.50 |
| AA cost per Index task (v4.3.2) | $1.04 | $1.86 |
Two compounding causes: Grok 4.6 emits materially more output tokens for the same work, and cached input costs 67% more. The result is a 1.79x increase in measured cost per Index task — $1.04 to $1.86, High tier on both sides, Intelligence Index v4.3.2 — despite an unchanged rate card.
The cache rise hits agent loops hardest, because every turn re-sends the accumulated conversation. And note the threshold that applies to both versions: any prompt of 200K tokens or more bills the entire request at $4/$12, not just the tokens above the line.
Should you upgrade?
Yes, if your workload is one-shot or few-shot coding. SWE-bench up nine points and LiveBench coding up six is a clear win, and the extra output tokens matter less on short tasks.
Yes, if you were already reaching for xhigh and getting high silently.
Test first, if you run long agent loops. Agentic coding and SkillsBench both regressed, output per task is up, and the cache price rose — three effects that all compound in exactly that workload.
Probably not, if latency is user-visible. Going from 14.0 to 27.8 seconds before the first token will be felt.
Reconsider the vendor entirely, if you are choosing purely on cost. Meta's Muse Spark 1.2, launched a week earlier, is about half the cost per Index task (AA v4.3.2: $0.97 against $1.86) with twice the context and is statistically tied with Grok 4.6 on the Vals Index.
The honest summary
Grok 4.6 is a sideways-and-slightly-up release. It is better at writing code and reasoning, worse at agentic loops and latency, and more expensive to run despite an unchanged price list.
The most useful way to think about it comes from developers who have actually run it: Grok 4.6 is roughly Claude Opus 4.8-class at about a third of the price, and the workflow people keep reporting is to plan with an expensive model and implement with Grok. That is a real and valuable role. It is just not the same as a frontier upgrade.
FAQ
Is Grok 4.6 better than Grok 4.5?
For coding and reasoning, yes — SWE-bench rose about nine points to 95.60% and LiveBench coding by over six. For agentic work, no: LiveBench agentic coding fell from 56.5 to 54.2 and SkillsBench from 66.03 to 55.77.
Did Grok 4.6 change price?
The headline rate is identical at $2/$6 per million tokens. But cached input rose from $0.30 to $0.50, and AA's measured cost per Index task rose from $1.04 to $1.86 on Intelligence Index v4.3.2, because Grok 4.6 produces materially more output tokens for the same work.
Is Grok 4.6's context window bigger?
No. Both are 500,000 tokens, which is the smallest of the current frontier models. Prompts of 200K or more bill the entire request at the doubled $4/$12 rate.
Is Grok 4.6 slower than Grok 4.5?
On time to first token, yes — it went from 14.0 to 27.8 seconds (AA v4.3.2, High tier). On raw throughput, no: median output speed improved from 59.6 to 72.9 tok/s. The first-token wait is the regression, not the output rate.
What is xhigh reasoning effort?
A new reasoning level in Grok 4.6. Grok 4.5 accepted the parameter but silently downgraded it to high, so 4.6 is the first version where it has an effect.
Should I upgrade my agent from Grok 4.5 to 4.6?
Benchmark first. Agentic coding regressed, output per task rose, and cached input costs 67% more — three effects that compound in agent loops specifically.