Both of these shipped in the same week of August 2026 and they are not competing for the same job. Grok 4.6 measures 60.9 on the Artificial Analysis Intelligence Index at $0.8367 per completed task; DeepSeek V4 Pro’s GA build measures 53.0 at $0.0561 — 14.9× cheaper. Grok 4.6 is the successor to Grok 4.5 at an unchanged price; V4 Pro left preview at unchanged prices too, gaining 7.7 index points on the way. Our Grok 4.6 release notes track what shipped, when, and what it replaced.
Put them side by side and the shape of the decision is unusually clear.
The measurements
Artificial Analysis, checked August 13, 2026:
| Grok 4.6 (high) | DeepSeek V4 Pro 0813 (max) | |
| Intelligence Index | 60.9 | 53.0 |
| Cost per index task | $0.8367 | $0.0561 |
| AA-Omniscience | 30.5 | 0.8 |
| Median output speed | 67.6 tok/s | not published |
| Context window | 500k | 1M |
Grok 4.6 leads on capability, calibration and measured speed. V4 Pro leads on price by 14.9× and on context window by 2×.
The Omniscience gap is 29.7 points and it is the largest difference in the table. V4 Pro’s 0.8 is the lowest figure on that board; Grok 4.6’s 30.5 sits in the upper middle, behind Claude Opus 5 (37.1) and Claude Fable 5 (43.3). That index measures knowledge plus willingness to admit its absence, so a 0.8 means the model answers confidently rather than declining when it doesn’t know.
What each one is actually for
Grok 4.6 is for hard work in a chain. Frontier-band reasoning at 67.6 tokens per second — the fastest of any model scoring 59 or above — makes it the natural choice for agent loops whose individual steps are genuinely difficult. It also has the better calibration of the two by a wide margin, which matters more in a chain than anywhere else, since one invented intermediate result contaminates everything downstream.
V4 Pro is for volume where you supply the facts. At six cents a task with a 1M window and a cache-hit rate 120× below its miss rate, it makes patterns affordable that a frontier model forecloses: verifying every record instead of sampling, running three passes and comparing, reprocessing an archive after a prompt change. Its low Omniscience matters much less when the answer is in the prompt rather than in the weights — which is the architecture that pricing pushes you toward anyway.
Neither displaces the other. A system that uses one for the volume and the other for the hard tail is cheaper and better than a system that picks one, and the split is measurable rather than a matter of taste.
The two rate cards behave differently
Grok 4.6, from xAI’s model documentation, checked August 13, 2026:
• $2.00 / $6.00 per million tokens under 200,000
• $4.00 / $12.00 at or above 200,000 — applied to the whole request
• Cached input $0.50 / $1.00, up 67% from Grok 4.5’s $0.30 / $0.60
• 500k context, knowledge cutoff February 1, 2026
DeepSeek V4 Pro, from DeepSeek’s own API documentation, same day:
• $0.435 per million input on a cache miss, $0.003625 on a hit
• $0.87 per million output
• 1M context, 384K maximum output
• No tier threshold — one rate across the whole window
• A notice that pricing will rise significantly, with no date and no replacement rates published
That structural difference matters for long-context work in a way the headline numbers don’t show. Grok 4.6’s advertised 500k window has 60% of its range above the doubling point, so a long request bills at $4.00 / $12.00. V4 Pro’s 1M window bills linearly throughout. For document-heavy pipelines the effective gap is wider than 14.9×, and it widens further if your prompts have a stable prefix that hits cache.
A concrete routing split
Send to V4 Pro: classification, extraction into a schema, summarising supplied text, first-pass filtering, bulk document interrogation, anything with a mechanical check downstream. This is where 14.9× multiplies and where 0.8 Omniscience has the least to act on.
Send to Grok 4.6: the hard tail — multi-step reasoning where a 53-point model fails and a 61-point one succeeds, agent chains without verification steps, work where throughput compounds across hundreds of sequential calls.
Measure the boundary rather than guessing it. Take fifty tasks from the difficult end of your distribution and run both. Score cost per completed job, quality on your own criteria, and — most importantly — how each fails. If V4 Pro clears your hard tail, the rest of your traffic was never the problem. If it doesn’t, you have found exactly where the 14.9× is worth paying and can route accordingly.
That test is cheap in a way it wouldn’t have been a year ago: fifty tasks on V4 Pro costs under three dollars on these figures.
The takeaway
Grok 4.6 costs 14.9× more per completed task than DeepSeek V4 Pro and delivers 7.9 more index points, 29.7 more points of factual calibration, and a published output speed of 67.6 tokens per second where V4 Pro has none. V4 Pro answers with twice the context window, no pricing tier to cross, and a cache-hit rate 120× below its miss rate. They are complementary rather than competing: route grounded, verifiable volume to the cheap model and the hard, unverified tail to the expensive one, and measure the boundary on fifty real tasks rather than choosing one for everything.
Sourcing note: Intelligence Index, cost per index task, AA-Omniscience, median output speed and context windows are from Artificial Analysis, checked August 13, 2026, at the configurations labelled; that evaluator publishes no output-speed figure for the DeepSeek V4 Pro 0813 build. Grok per-token prices, tier thresholds, cached-input rates and the knowledge cutoff are from xAI’s own model documentation, checked the same day. DeepSeek’s context length, maximum output, per-token rates and price-increase notice are from DeepSeek’s own API documentation, checked the same day; that notice carries no date and no replacement rates.
