Shipped the same night, but two different games
On the night of August 12, Hangzhou and Los Angeles dropped major models almost simultaneously. DeepSeek quietly flipped V4 Pro from preview to general availability, and the same day Musk's SpaceXAI released Grok 4.6. On the surface it looks like a China-vs-US collision. The strategies couldn't be more different: one squeezes price down, the other stacks capability up.
Tellingly, DeepSeek didn't even post an announcement. It just changed one version line in the API docs. That quiet ship is exactly its style.
The cards, laid out
- DeepSeek V4 Pro (build 0813, GA): 1.6 trillion total parameters, about 49 billion active per token, a 1-million-token context, and up to 384K tokens of output. Pricing held from preview: roughly 3 yuan per million input tokens (0.025 yuan on cache hit), 6 yuan per million output. Agent scores jumped hard - Terminal-Bench 2.1 went from 72.1 to 87.9, just 0.1 behind Fable 5's 88.0. The tradeoff: no vision input yet.
- Grok 4.6: 2 dollars per million input tokens, 6 dollars output, flat versus the previous generation. On GDPVal-AA v2 it scored 1753 Elo, above Fable 5 Max's 1741 and GPT-5.6 Sol Max's 1728, landing in the top tier. It tunes for long-horizon agents, complex coding, and knowledge work.
The real read: don't get distracted by parameters
DeepSeek pushed agentic coding to Fable 5's doorstep at roughly one-sixtieth of the price. That is not marketing - Terminal-Bench is a real coding benchmark. Grok 4.6 takes the other road: pricier, but a higher overall ceiling and steadier on long reasoning and English knowledge tasks.
Plainly, one is "buy 80 percent of the capability for loose change," the other is "pay up for the full build." Both work, as long as you know which you need.
Which one to open
- Pick DeepSeek V4 Pro for high-frequency production, agent loops, cost sensitivity, Chinese-first work, cheap but capable.
- Pick Grok 4.6 for top-tier overall, long-horizon reasoning, English knowledge work, and a budget that tolerates it.
What this means for you
You are probably not a tuning engineer, so don't let "1.6 trillion" scare you. Do the math: for the same 1 million output tokens, DeepSeek costs about 6 yuan, Grok about 43 yuan (6 dollars) - over seven times apart. For a small team running daily business, DeepSeek's value is close to a landslide; reach for the expensive tier only when you need top reasoning on long chains.
One concrete case: a customer-service agent emitting 500K tokens a day costs under 10 yuan a month on DeepSeek, versus over 60 yuan on Grok - a 600-700 yuan gap a year, enough for a monitor. For a business just starting, that math beats any leaderboard.
Why both bet on agents
The same-night launch is no coincidence. Through 2025 the big labs proved that post-training on tool use and long-horizon tasks is where capability gains now live. DeepSeek ran a heavy agentic RL pass over its April base and DeepSWE jumped from 12.8 to 62.7; Grok 4.6 also names long-horizon agents and complex coding as its main target. The base-model arms race is cooling; finishing the job is the new battleground.
Two caveats up front: DeepSeek has warned API prices will rise significantly soon, so today's rate may not last; Grok 4.6 is strong but can't clear some human-verification flows and costs more; and DeepSeek has no vision, so image-understanding jobs are out.
Wrap
The next winner in this race may not be the highest scorer, but the one that actually finishes the job at a controllable cost. These two launching the same night simply put "cheap and capable" and "pricey but full-blooded" side by side for you.