Grok 4.6 vs. DeepSeek V4 Pro: The Agent Price War Is Now the Story

A dramatic version of this week’s AI story is easy to tell: DeepSeek drops V4 Pro at an absurdly low price, Elon Musk answers with Grok 4.6, and the frontier model race turns into a same-day shootout.

The competitive pressure is real. The chronology is not quite that simple.

DeepSeek V4 Pro was originally released on April 24, 2026, as part of the DeepSeek-V4 preview family. What matters now is that its pricing has settled at a level that makes frontier-class agent workloads dramatically cheaper. Meanwhile, Grok 4.6 launched on August 12, and independent measurements place it directly back on the intelligence frontier.

The more interesting story is therefore not who released a model first. It is how quickly high-end agent capability is being compressed into lower prices.

DeepSeek’s weapon is the cost curve

DeepSeek’s current API pricing for V4 Pro is unusually aggressive. The official pricing page lists the model at $0.435 per million uncached input tokens and $0.87 per million output tokens, with cache-hit input falling to just $0.003625 per million tokens. The model supports a 1 million-token context window, tool calls, JSON output, and both thinking and non-thinking modes.

That pricing changes the practical calculation for agent systems.

A useful autonomous agent does not make one request. It reads files, carries context forward, calls tools, revises plans, inspects results, and may execute dozens of model turns before a task is complete. Once a workflow becomes long-running, token economics stop being a line item and become an architectural constraint.

DeepSeek is attacking that constraint directly.

Its own launch material describes V4 Pro as an open model with substantially improved agentic coding performance and says the V4 family is already used internally for agentic coding. Those are vendor claims and should be treated as such, but the combination of a million-token context window and sub-dollar output pricing is independently important even before debating benchmark leadership.

Grok 4.6 attacks from the other direction

Grok 4.6 is more expensive per token than DeepSeek V4 Pro, but its first independent benchmark results show why xAI can still make a strong economic argument.

Artificial Analysis reports that Grok 4.6 scores 61 on its Intelligence Index, a five-point gain over Grok 4.5 and roughly in line with GPT-5.6 Sol. Its strongest results are not isolated math or trivia tests, but agentic evaluations: long-horizon knowledge work, banking tasks with tool use, and terminal-based software work.

The model keeps the same headline pricing as Grok 4.5: $2 per million input tokens and $6 per million output tokens. Artificial Analysis measured an average cost of $0.84 per Intelligence Index task, putting Grok 4.6 on its cost-versus-intelligence Pareto frontier.

That is the key point. Grok 4.6 does not need to be the cheapest API token to be economically competitive. If a stronger model completes a task in fewer turns, with fewer retries and less accumulated context, its total cost can beat a cheaper model that wanders.

Artificial Analysis saw exactly that kind of behavior on its AA-Briefcase long-horizon benchmark. Grok 4.6 completed tasks in roughly 53 turns on average, while Claude Opus 5 at maximum effort used about 103 turns. For agent builders, turn efficiency can matter as much as the list price.

The Fable comparison needs more precision

The viral claim that Claude Fable 5 is “45 times more expensive for only 5% more speed” is too neat to publish as a fact.

Anthropic prices Fable 5 at $10 per million input tokens and $50 per million output tokens, with a 90% input-token discount for prompt caching. Compared with DeepSeek V4 Pro, that makes Fable dramatically more expensive on headline token rates, but the multiplier depends on whether a workload is dominated by input, output, or cached context. There is no single honest “45x” number that describes every agent workload.

The same problem applies to speed. Tokens per second, end-to-end task time, number of turns, retry rate, and successful completion rate are different measurements. A model that generates faster can still finish slower if it needs more iterations.

For agents, the correct unit is increasingly cost per successful task, not cost per token and not raw generation speed.

GPT-5.6 Sol is the benchmark that makes Grok 4.6 interesting

OpenAI’s GPT-5.6 Sol established one of the current reference points for frontier reasoning and agentic work. OpenAI prices Sol at $5 per million input tokens and $30 per million output tokens. In Artificial Analysis testing around its launch, Sol sat within roughly one point of the top intelligence score while leading the Coding Agent Index.

Grok 4.6 now lands in the same broad intelligence tier at a substantially lower headline token price. Artificial Analysis scores it at 61 and describes it as being in line with GPT-5.6 Sol, while charging $2/$6 instead of $5/$30.

That does not make Grok universally better. Benchmarks are workload-specific, harnesses affect agent results, and reliability under real production conditions matters more than a leaderboard position. But it does make Grok 4.6 a serious option for developers who previously treated xAI as outside the top efficiency frontier.

The frontier is becoming an economics problem

The most important change is not that one company briefly caught another on a benchmark.

It is that frontier-level capability is becoming available across increasingly different price structures.

DeepSeek V4 Pro pushes the raw token price down to a level that encourages experimentation with very long contexts and multi-step agents. Grok 4.6 shows that a proprietary frontier model can improve intelligence without raising headline pricing. GPT-5.6 Sol demonstrates that a more expensive model can remain competitive by using fewer tokens and delivering stronger agent results. Fable 5 remains a premium option for demanding long-horizon work, but its price now has to be justified task by task rather than by brand position alone.

For builders, this changes model selection from a leaderboard question into a systems question:

• How many turns does the agent need?
• How much context is repeatedly carried forward?
• How effective is prompt caching?
• How often does the model recover from its own mistakes?
• What is the cost of a completed task rather than a generated token?

That is where the next phase of the model war will be decided.

The frontier is still moving upward. The more disruptive trend is that its price is moving downward at the same time.