DeepSeek V4 Pro 0813: Pricing, Benchmarks and the Mac Reality Check

Checked: August 18, 2026

DeepSeek released DeepSeek-V4-Pro-0813 on August 13, 2026 as the GA version of its largest V4 Pro model. The headline numbers are unusual: 1.6 trillion total parameters, 49 billion active parameters per token, a one-million-token context window and three reasoning levels. The direct DeepSeek API still uses the simpler model ID deepseek-v4-pro.

The useful question is not whether the name changed. It is whether the 0813 checkpoint is materially better than the preview, whether it justifies its price premium over V4 Flash 0731, and what those 49 billion active parameters actually mean for local inference on a Mac.

The short version is straightforward: V4 Pro 0813 is much stronger on long-horizon coding and agent workloads, but it is not automatically the best default model. Flash is cheaper, often faster and currently sits only slightly behind Pro on broad independent intelligence scoring. And despite the 49B active figure, the full model still contains 1.6 trillion weights, which makes it unrealistic for a normal single Mac to run locally.

Core specifications

Specification	DeepSeek V4 Pro 0813
Release	August 13, 2026
Direct API ID	deepseek-v4-pro
Architecture	Mixture of Experts
Total parameters	1.6 trillion
Active parameters per token	49 billion
Context window	1M tokens
Maximum API output	384K tokens
Reasoning	low, high, max
Default	Thinking, high
Tool calls	Yes
OpenAI Responses-compatible API	Yes
Anthropic-compatible endpoint	Yes
Weight license	MIT

Sources: DeepSeek GA release, DeepSeek Models & Pricing, V4 Technical Report, Model Card.

0813 is the checkpoint revision, not the direct API name

The naming is easy to get wrong.

The released checkpoint is:

DeepSeek-V4-Pro-0813

The direct DeepSeek API uses:

deepseek-v4-pro

That distinction matters in real integrations. Copying the checkpoint name and turning it into deepseek-v4-pro-0813 is not what DeepSeek currently documents for its own API.

Third-party providers use their own IDs. Cloudflare, for example, exposes:

@cf/deepseek-ai/deepseek-v4-pro-0813

OpenRouter uses:

deepseek/deepseek-v4-pro-0813

Model IDs should therefore be treated as provider-specific configuration rather than a universal property of the checkpoint.

1.6 trillion total parameters does not mean a 49B local model

V4 Pro uses a Mixture-of-Experts architecture. Only part of the network is active for each token; DeepSeek reports 49 billion activated parameters per token.

That lowers computation compared with a dense 1.6T model. It does not solve the local memory problem. The expert weights still have to be available to the inference system.

This is where the 49B number can be misleading. 49B active does not mean the checkpoint can be loaded like an ordinary 49B model.

The V4 Technical Report also describes architecture changes aimed at long-context efficiency, including Compressed Sparse Attention and Heavily Compressed Attention. DeepSeek reports substantially lower inference FLOPs and KV-cache requirements at one-million-token context compared with V3.2. Those are first-party technical-report figures, not independent Mac measurements.

What changed in the 0813 release

The GA checkpoint is more than a renamed preview. DeepSeek strengthened its agent post-training and integrated DSpark, its speculative decoding system.

DSpark prepares multiple likely continuation tokens and then verifies several of them together. In the accompanying paper, DeepSeek reports roughly 60–85% higher per-user generation speed than its previous MTP-1 baseline at matched serving throughput.

That number needs to be interpreted narrowly. It describes a documented serving setup. It is not a promise that every hosted V4 Pro endpoint will answer 60–85% faster.

The biggest gains are in coding and agents

DeepSeek’s own 0813 evaluations show the clearest improvements on long tool-use and coding workloads:

Benchmark	Pro 0813	Flash 0731	Pro Preview
HLE with tools	60.0	51.5	48.2
Terminal-Bench 2.1	87.9	82.7	72.1
NL2Repo	61.5	54.2	38.5
CyberGym	83.3	76.7	52.7
DeepSWE	62.7	54.4	12.8
Toolathlon-Verified	74.1	70.3	55.9
AutomationBench Public	31.8	25.1	12.8

Source: DeepSeek V4 Pro 0813 Model Card.

DeepSeek says its coding-agent evaluations used DeepSeek Harness with max reasoning, temperature=1.0 and top_p=0.95, among other settings. That makes the numbers most useful for comparing DeepSeek checkpoints under similar conditions. They should not be treated as a substitute for independent cross-vendor evaluation.

The pattern is still meaningful: 0813 improves the workloads where an agent has to remain coherent over many sequential steps.

Independent benchmarks show a much smaller overall gap

Artificial Analysis currently places DeepSeek V4 Pro 0813 Max at 53 on its Intelligence Index. Flash 0731 Max is at 52.

That matters because launch coverage cited much larger differences. Benchmark dashboards change as new runs are added or methodology is updated. For a current article, the maintained benchmark page is more useful than freezing the launch-day number.

The practical reading is therefore:

• Pro has a clear advantage on some agent and coding workloads.
• The broad composite gap to Flash is small.
• Treating Flash as dramatically less capable across the board is too simplistic.

Sources: Artificial Analysis – DeepSeek V4 Pro, Artificial Analysis – DeepSeek V4 Flash.

Current pricing: Pro is about three times the cost of Flash on DeepSeek’s direct API

DeepSeek changed V4 pricing again in mid-August. The direct API now has peak and off-peak windows.

Peak hours:

• 01:00–04:00 UTC
• 06:00–10:00 UTC

All other hours are billed at half price.

Direct API pricing

Model	Token class	Off-peak	Peak
V4 Pro	Cached input	$0.022 / 1M	$0.044 / 1M
V4 Pro	Uncached input	$0.66 / 1M	$1.32 / 1M
V4 Pro	Output	$1.98 / 1M	$3.96 / 1M
V4 Flash	Cached input	$0.007 / 1M	$0.014 / 1M
V4 Flash	Uncached input	$0.22 / 1M	$0.44 / 1M
V4 Flash	Output	$0.66 / 1M	$1.32 / 1M

Source: DeepSeek Models & Pricing.

Under the current direct rate card, Pro is approximately three times the cost of Flash across cached input, uncached input and output.

That is an important update to launch coverage. Reuters reported roughly 9× input and 14× output pricing versus Flash on August 13. Those ratios reflected Flash’s pricing at the time and no longer describe the August 18 rate card.

What that means in real usage

Per-million-token pricing is difficult to reason about. A few sample workloads make the difference clearer.

Light usage

Assume:

• 1M uncached input tokens
• 2M cached input tokens
• 0.25M output tokens

Pro off-peak

1 × 0.66
+ 2 × 0.022
+ 0.25 × 1.98
= 1.199 USD

Rounded: $1.20

The same workload at peak pricing costs $2.40.

Medium usage

Assume:

• 10M uncached input tokens
• 10M cached input tokens
• 3M output tokens

Result:

• $12.76 off-peak
• $25.52 peak

Heavy usage

Assume:

• 50M uncached input tokens
• 50M cached input tokens
• 15M output tokens

Result:

• $63.80 off-peak
• $127.60 peak

These examples exclude taxes, exchange-rate effects and third-party provider markups.

For batch agents, scheduling can therefore matter almost as much as model choice. Moving a workload outside DeepSeek’s peak windows halves the direct API token cost.

Providers differ in both price and speed

V4 Pro 0813 is available through DeepSeek and several third-party hosts including Baseten, Fireworks, DeepInfra, Novita, Cloudflare and OpenRouter.

Artificial Analysis tracks provider performance separately. These numbers change quickly, but the lesson is stable: the cheapest endpoint is not necessarily the fastest one.

Production comparisons should include:

• Time to First Token
• output tokens per second
• reliability and error rate
• cache billing
• tool-call behavior
• region and data handling
• provider-specific features

Current provider comparison: Artificial Analysis – V4 Pro Providers.

Does V4 Pro support images? It depends on the provider

The documentation is inconsistent across hosts.

DeepSeek’s own Responses API explicitly says image and file inputs are not supported. Its official Codex configuration is also text-only.

Cloudflare, however, marks its Workers AI deployment of V4 Pro 0813 as Vision: Yes.

The safe conclusion is therefore:

Treat the direct DeepSeek API as text-only unless DeepSeek documents otherwise. Some hosted implementations may expose additional multimodal serving paths.

This is a good example of why provider behavior should not be generalized to the underlying checkpoint.

Sources: DeepSeek Responses API, Cloudflare Workers AI.

OpenAI Responses-compatible is not the same as OpenAI feature parity

DeepSeek supports an OpenAI Responses-style API, so the model can be called with the OpenAI SDK:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ["DEEPSEEK_API_KEY"],
    base_url="https://api.deepseek.com",
)

response = client.responses.create(
    model="deepseek-v4-pro",
    instructions="You are a careful coding assistant.",
    input="Review this function for correctness.",
    reasoning={"effort": "high"},
)

print(response.output_text)

That makes migrations easier, but it does not mean DeepSeek implements the entire OpenAI Responses platform.

DeepSeek lists several unsupported features, including:

• background
• store
• conversation
• previous_response_id
• file_search
• code_interpreter
• computer_use
• general MCP tools

Some unsupported parameters may be silently ignored. An agent migration therefore needs semantic testing, not just a successful HTTP response.

Source: DeepSeek Responses API.

Anthropic-compatible endpoint and Claude Code-style workflows

DeepSeek also exposes an Anthropic-compatible endpoint:

https://api.deepseek.com/anthropic

Its documented model mapping includes:

• claude-opus* → deepseek-v4-pro
• claude-sonnet* → V4 Flash
• claude-haiku* → V4 Flash

That makes V4 Pro easier to slot into workflows built for Anthropic-style APIs without rewriting every integration from scratch.

As with the OpenAI-compatible endpoint, transport compatibility should not be confused with full behavioral parity.

Source: DeepSeek Anthropic API.

Can DeepSeek V4 Pro 0813 run locally on a Mac?

For a normal single Mac, the practical answer is no.

The important number is not 49 billion active parameters. It is 1.6 trillion total weights.

Theoretical lower bound for weight storage

Own calculation:

Representation	Theoretical weight-only size
FP16/BF16	~3.2 TB
FP8	~1.6 TB
4-bit	~800 GB
2-bit	~400 GB

Formula:

1.6 trillion parameters × bits per parameter ÷ 8

Those figures are weight-only lower bounds. They exclude:

• quantization metadata
• KV cache
• runtime overhead
• temporary tensors
• macOS
• other running software

Even an aggressive 4-bit estimate lands around 800 GB before runtime memory is considered.

MoE sparsity reduces computation per token. It does not transform 1.6 trillion weights into a 49B checkpoint.

DeepSeek’s own vLLM example uses a 4×GB300 node for V4 Pro 0813, which is a much better indication of the intended hardware class than a consumer Mac.

Source: DeepSeek V4 Pro 0813 Model Card.

What about Ollama?

Ollama lists:

ollama run deepseek-v4-pro:0813-cloud

The important word is cloud.

That entry is not evidence that Ollama is loading the full 1.6T checkpoint into Apple unified memory. It is a cloud-hosted model.

Source: Ollama – DeepSeek V4 Pro 0813 Cloud.

For real on-device inference on Apple Silicon, substantially smaller open-weight models remain the practical category. Model size, quantization, desired context and available unified memory all need to be considered together.

Pro or Flash: which should you use?

The most useful decision rule is based on workflow rather than a single benchmark number.

Choose Flash 0731 by default when

• token cost matters,
• the agent performs many routine steps,
• low latency matters more than maximum reasoning depth,
• most tasks finish in a few tool calls,
• a failed run is cheap to retry.

Use Pro 0813 when

• an agent works across large repositories,
• many tool steps depend on one another,
• failures late in the task are expensive,
• terminal, repository or automation work is central,
• maximum reasoning depth matters more than token cost.

That suggests a practical routing strategy:

Flash for routine paths, Pro for difficult escalations.

For many agent systems, that is more efficient than routing every request to the most expensive model.

Conclusion: Agent and coding gains at higher requirements

DeepSeek V4 Pro 0813 is a meaningful upgrade over the preview, especially where modern agents tend to fail: long coding tasks, tool chains, terminal work and multi-step automation.

It is still not an automatic default choice. The current direct API is roughly three times the price of Flash, broad independent intelligence scoring places the two models close together, and Flash remains good enough for many routine workloads.

For Mac users, the conclusion is even clearer. V4 Pro is interesting as an API model, not as a realistic local 49B model. The 49 billion active parameters describe computation per token; local memory is still dominated by the model’s 1.6 trillion total weights.

A simple rule therefore works well:

Use Flash by default. Escalate to Pro when the cost of a failed run is higher than the model premium.