DeepSeek-V4-Pro-0813 GA: The Agent Model That Hits Fable-Level Coding at 1/57th the Price

Executive Summary

On August 13, 2026, DeepSeek launched the official general availability version of DeepSeek-V4-Pro-0813, marking a dramatic step-function in open-weight agentic coding capability. The release represents the culmination of a four-month evolution from the April 2026 V4 preview through the July 31 V4-Flash-0731 official release, with the 0813 checkpoint delivering massive post-training gains specifically targeted at agent workloads.

The most striking result: V4-Pro-0813 achieves 62.7 on DeepSWE (up from 7.3 in the April preview — an 860% increase), 87.9 on Terminal Bench 2.1 (within 0.1 points of Claude Fable 5's 88.0), and 80.6% on SWE-bench Verified (level with Gemini-3.1-Pro). These scores place an open-weight model within striking distance of the proprietary frontier on coding benchmarks, at approximately 1/57th of Fable 5's output pricing ($0.87 vs. $50 per million output tokens).

The release is bundled with DeepSeek Harness v0.1, an open-source agent framework that provides a runnable alternative to Claude Code and OpenAI Codex, and a new peak/off-peak pricing model that makes DeepSeek the first major API provider to offer structural time-based pricing — with off-peak rates at 50% of peak.

This article provides a comprehensive analysis of the V4-Pro-0813 release, the Harness framework, the pricing evolution, benchmark performance, integration story, and strategic implications for the open-weight agent ecosystem.

---

1. The Release: From Preview to Production

1.1 The Timeline

The DeepSeek-V4 journey has been a study in iterative improvement:

Date	Release	Key Changes
April 24, 2026	V4 Preview (Pro + Flash)	1.6T MoE architecture, 1M context, MIT license, initial benchmarks
July 31, 2026	V4-Flash-0731 GA	Retrained checkpoint, agent benchmarks beat Pro Preview, Responses API
August 13, 2026	V4-Pro-0813 GA	Major agent upgrades, Harness v0.1, peak/off-peak pricing, Codex integration

The jump from the April preview to the 0813 checkpoint is not an architecture change — it is entirely post-training improvement. The same 1.6T MoE with 49B active parameters, the same hybrid attention, the same mHC connections. What changed is the data, the alignment, and the agent-specific fine-tuning.

1.2 What's New in 0813

DeepSeek's official announcement highlights four key upgrades:

1. Major Agent upgrades — Strong production gains across coding, tool-use, and multi-step reasoning benchmarks
2. Flexible reasoning effort — Low for simple tasks, high for daily agent workflows, max for complex problems
3. Native OpenAI Responses API — Optimized for Codex with one-click setup
4. DeepSeek Harness v0.1 — Open-source agent framework released alongside the model

1.3 The Naming

The model identifier follows DeepSeek's versioning convention: DeepSeek-V4-Pro-0813 where 0813 is the release date (August 13). The API model name remains deepseek-v4-pro — existing integrations continue to work without changes, but now point to the 0813 checkpoint.

---

2. Architecture: The Same Foundation, Better Training

2.1 Core Architecture

DeepSeek-V4-Pro-0813 retains the architecture introduced in the April 2026 preview, as described in the technical report (arXiv:2606.19348):

Component	Specification
Architecture	Mixture-of-Experts (MoE)
Total parameters	1.6 trillion
Active parameters	49 billion
Context length	1,000,000 tokens
Max output	384,000 tokens
Attention	Hybrid: Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA)
Residual connections	Manifold-Constrained Hyper-Connections (mHC)
Optimizer	Muon
Training data	32+ trillion tokens
License	MIT

2.2 Hybrid Attention: CSA + HCA

The key architectural innovation is the dual-path hybrid attention mechanism that replaces V3.2's Multi-Head Latent Attention (MLA):

• Compressed Sparse Attention (CSA): Compresses the KV cache for frequently-accessed tokens while maintaining full precision for critical positions
• Heavily Compressed Attention (HCA): Aggressively compresses less-critical tokens to minimize memory usage

In the 1M-token context setting, this design achieves dramatic efficiency:

Metric	DeepSeek-V4-Pro	DeepSeek-V3.2	Improvement
Single-token inference FLOPs	27% of V3.2	100%	3.7× reduction
KV cache memory	10% of V3.2	100%	10× reduction

This is what enables routine 1M-token context support without prohibitive memory costs — a prerequisite for long-horizon agentic workloads that require sustained context across hundreds of tool calls.

2.3 The Post-Training Difference

What separates 0813 from the April preview is entirely in the post-training pipeline:

1. Agent-heavy data: Curated multi-step tool-use trajectories, coding workflows, and long-horizon task completions
2. Reasoning effort tuning: Training the model to calibrate its reasoning depth based on task complexity (low/high/max)
3. Tool-use alignment: Improved instruction following for tool calls, error recovery, and multi-turn agent loops
4. Context management: Better handling of 1M-token contexts with selective reasoning_content preservation

---

3. Benchmarks: The Step-Function in Agentic Coding

3.1 The DeepSWE Revolution

The most dramatic improvement is on DeepSWE, DeepSeek's own software engineering benchmark:

Model	DeepSWE Score	Improvement
V4-Pro Preview (April)	7.3	—
V4-Flash-0731 (July)	12.8	+75%
V4-Pro-0813 (August)	62.7	+390% over 0731

The jump from 7.3 to 62.7 in two releases is extraordinary. As one observer noted: "The jump on DeepSWE is probably the most interesting number here. 7.3 → 62.7 in two releases is kind of wild."

3.2 Coding Benchmarks vs. Frontier

On established coding benchmarks, V4-Pro-0813 competes directly with the proprietary frontier:

Benchmark	V4-Pro-0813	Claude Fable 5	Claude Opus 4.8	Gemini-3.1-Pro	GPT-5.6 Sol
Terminal Bench 2.1	87.9	88.0	—	—	—
SWE-bench Verified	80.6	80.8	—	80.6	—
LiveCodeBench	93.5	—	—	—	—
DeepSWE	62.7	—	—	—	—

The Terminal Bench 2.1 score of 87.9 is within 0.1 points of Fable 5 — effectively tied. The SWE-bench Verified score of 80.6% places it level with Gemini-3.1-Pro and a tenth of a point behind Opus 4.6.

3.3 Reasoning Effort Impact

The flexible reasoning effort control (low/high/max) shows different performance profiles:

Benchmark	High Effort	Max Effort	Delta
LiveCodeBench	89.8	93.5	+3.7
HLE	—	—	+3.2
MRCR 1M	—	—	+0.2
SWE-bench Verified	79.4	80.6	+1.2

The diminishing returns from high to max effort on coding tasks (only +1.2 on SWE-bench) suggest that for most production coding workflows, high effort is the sweet spot — delivering 95% of max performance at lower latency and cost.

3.4 Important Caveats

These benchmarks are vendor-reported by DeepSeek. No third-party evaluator has independently replicated the scores as of August 14, 2026. The benchmarks also compare against Fable 5 (released April 2026) rather than the latest Anthropic models. Terminal Bench 2.1 is approaching saturation, where marginal improvements become increasingly difficult to achieve and may not reflect real-world capability differences.

---

4. DeepSeek Harness v0.1: The Open-Source Agent Framework

4.1 What Is Harness?

Released alongside V4-Pro-0813, DeepSeek Harness v0.1 is an open-source framework for building and running AI agents. It is positioned as a direct alternative to Claude Code and OpenAI Codex, but with a key difference: it is fully open-source and works with any model that supports the Responses API.

4.2 Key Features

Feature	Description
Plugin architecture	Extensible tool system for custom integrations
Traceable sessions	Full observability of agent decision-making and tool calls
Multiple runtime modes	Terminal, browser-based interface, and API
Cache-first loop	Optimized for cost-efficient agent execution
Model escalation	Default to V4-Flash, escalate to V4-Pro for complex tasks
Codex-compatible	Works with OpenAI's Codex client via Responses API

4.3 The Development Timeline

Harness moved from team formation to developer preview in approximately five months — aggressive even by Chinese tech standards:

4.4 Strategic Positioning

Harness represents DeepSeek's attempt to own the agent orchestration layer, not just the model layer. By providing an open-source framework that works best with their own models but is compatible with others, DeepSeek creates a sticky ecosystem:

• Developers who adopt Harness are incentivized to use V4-Pro/V4-Flash for best results
• The open-source nature lowers adoption barriers compared to proprietary tools
• The plugin architecture encourages community contributions and custom integrations

---

5. Integration Story: Responses API and Codex

5.1 Native Responses API Support

V4-Pro-0813 now supports the OpenAI Responses API format natively, with the base URL https://api.deepseek.com. This enables:

• One-click Codex integration
• Compatibility with any tool that supports the Responses API
• Seamless switching between DeepSeek and OpenAI models in agent workflows

5.2 Codex Integration

DeepSeek provides a one-click setup for Codex integration:

{
  "models": [
    {
      "slug": "deepseek-v4-flash",
      "display_name": "DeepSeek-V4-Flash",
      "description": "Latest frontier agentic coding model.",
      "context_window": 1048576,
      "supports_parallel_tool_calls": true,
      "multi_agent_version": "v2",
      "default_reasoning_level": "high",
      "supported_reasoning_levels": [
        {"effort": "low", "description": "Fast responses with lighter reasoning"},
        {"effort": "high", "description": "Extra high reasoning depth for complex problems"},
        {"effort": "max", "description": "Maximum reasoning depth for the hardest problems"}
      ]
    },
    {
      "slug": "deepseek-v4-pro",
      "display_name": "DeepSeek-V4-Pro",
      "description": "Flagship agentic coding model.",
      "context_window": 1048576,
      "supports_parallel_tool_calls": true,
      "multi_agent_version": "v2",
      "default_reasoning_level": "high",
      "supported_reasoning_levels": [
        {"effort": "low", "description": "Fast responses with lighter reasoning"},
        {"effort": "high", "description": "Extra high reasoning depth for complex problems"},
        {"effort": "max", "description": "Maximum reasoning depth for the hardest problems"}
      ]
    }
  ]
}

5.3 Anthropic API Compatibility

DeepSeek also supports the Anthropic API format via the base URL https://api.deepseek.com/anthropic, enabling compatibility with Claude Code and other Anthropic-native tools.

5.4 Reasoning Effort Control

The thinking mode now supports three effort levels across both models:

Requested Effort	Actual Mapped Effort	Use Case
low	low	Simple tasks, quick answers
medium	high	Daily agent workflows
high	high	Complex reasoning tasks
xhigh	high	Complex reasoning tasks
max	max	Hardest problems, maximum depth

The mapping collapses medium/high/xhigh to the same "high" level, with only low and max being distinct. This suggests the model has two primary reasoning modes: lightweight and deep.

---

6. Pricing: The Peak/Off-Peak Revolution

6.1 Current Pricing (Pre-August 16)

Model	1M Input (Cache Hit)	1M Input (Cache Miss)	1M Output
V4-Flash	$0.0028	$0.14	$0.28
V4-Pro	$0.003625	$0.435	$0.87

6.2 New Peak/Off-Peak Pricing (Effective August 16)

DeepSeek is introducing the industry's first structural peak/off-peak pricing model:

Model	Time	1M Input (Cache Hit)	1M Input (Cache Miss)	1M Output
V4-Flash	Off-Peak	$0.007	$0.22	$0.66
V4-Flash	Peak	$0.014	$0.44	$1.32
V4-Pro	Off-Peak	$0.022	$0.66	$1.98
V4-Pro	Peak	$0.044	$1.32	$3.96

Peak hours: 01:00–04:00 UTC and 06:00–10:00 UTC (covering Asia and Europe working hours) Off-peak hours: All other times (16 hours per day)

6.3 Pricing Analysis

The new pricing represents a significant increase from current rates:

Model	Metric	Current	Off-Peak	Peak	Change (Off-Peak)
V4-Pro	Input (miss)	$0.435	$0.66	$1.32	+52%
V4-Pro	Output	$0.87	$1.98	$3.96	+128%
V4-Flash	Input (miss)	$0.14	$0.22	$0.44	+57%
V4-Flash	Output	$0.28	$0.66	$1.32	+136%

However, the off-peak discount (50%) creates a powerful incentive for workload scheduling:

• Batch processing: Run expensive agent workflows during off-peak hours
• Async agents: Schedule long-horizon tasks to start during off-peak windows
• Hybrid approach: Use V4-Flash for real-time interaction, escalate to V4-Pro during off-peak for complex tasks

6.4 Comparison with Competitors

Model	1M Output Price	V4-Pro Off-Peak Ratio
Claude Fable 5	$25.00	12.6× more expensive
Claude Opus 5	$25.00	12.6× more expensive
GPT-5.6 Sol	$12.00	6.1× more expensive
Gemini-3.1-Pro	~$12.00	6.1× more expensive
V4-Pro (Off-Peak)	$1.98	—
V4-Pro (Peak)	$3.96	—

Even at peak pricing, V4-Pro is 3× cheaper than GPT-5.6 Sol and 6× cheaper than Fable 5 on output tokens. At off-peak, the gap widens to 6× and 12.6× respectively.

---

7. Deployment Guide: From API to Agent

7.1 Quick Start with OpenAI SDK

from openai import OpenAI

client = OpenAI(
    api_key="<DeepSeek API Key>",
    base_url="https://api.deepseek.com"
)

# Thinking mode with high reasoning effort
response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Refactor this codebase to use async/await."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}}
)

reasoning = response.choices[0].message.reasoning_content
answer = response.choices[0].message.content

7.2 Multi-Turn Agent with Tool Calls

import json
from openai import OpenAI

client = OpenAI(api_key="<key>", base_url="https://api.deepseek.com")

tools = [
    {
        "type": "function",
        "function": {
            "name": "search_codebase",
            "description": "Search the codebase for patterns",
            "parameters": {
                "type": "object",
                "properties": {"query": {"type": "string"}},
                "required": ["query"]
            }
        }
    }
]

messages = [{"role": "user", "content": "Find all unused imports in the project."}]

while True:
    response = client.chat.completions.create(
        model="deepseek-v4-pro",
        messages=messages,
        tools=tools,
        reasoning_effort="high",
        extra_body={"thinking": {"type": "enabled"}}
    )
    
    messages.append(response.choices[0].message)
    
    if not response.choices[0].message.tool_calls:
        break
    
    for tool_call in response.choices[0].message.tool_calls:
        result = execute_tool(tool_call)
        messages.append({
            "role": "tool",
            "tool_call_id": tool_call.id,
            "content": result
        })

7.3 Cost-Optimized Workflow

For production agent deployments, the peak/off-peak pricing enables a hybrid strategy:

import datetime

def get_effort_level():
    """Choose reasoning effort based on time of day and task complexity."""
    now = datetime.datetime.utcnow()
    is_peak = (1 <= now.hour < 4) or (6 <= now.hour < 10)
    
    if is_peak:
        # During peak hours, use lower effort to control costs
        return "low" if task_simple else "high"
    else:
        # During off-peak, afford maximum effort
        return "max" if task_complex else "high"

---

8. Connection to Prior Research

8.1 The DeepSeek V4 Lineage

This article continues the coverage of DeepSeek's V4 evolution:

• Deepseek V4 Flash 0731 Official Release Agentic Coding Price War 2026 08 04 — The July 31 V4-Flash-0731 release that first demonstrated strong agentic coding at ultra-low prices. The 0813 release represents the next step: taking the same architecture and applying months of post-training specifically to agent workloads.

8.2 The Agentic Coding Landscape

The V4-Pro-0813 release sits in a crowded agentic coding landscape:

• Meta Muse Spark 1 2 Muse Code Persistent Agents Co Trained Harness 2026 08 07 — Meta's Muse Code demonstrated 24-hour autonomous kernel optimization. DeepSeek Harness offers a similar capability but as open-source software.
• Qwen3 8 Max 2 4t Moe Open Weight Long Horizon Autonomous Coding 2026 08 05 — Qwen3.8-Max's 16-day autonomous coding project. V4-Pro-0813's 1M context and 384K max output enable similar long-horizon work.
• Meta Muse Glimmer 30b Open Agentic Local Distilled Spark Apache 2026 08 13 — Meta's local-first 30B model. V4-Pro-0813 takes the opposite approach: cloud-based but dramatically cheaper.

8.3 The Pricing War

The peak/off-peak pricing model represents a new phase in the AI pricing war:

• Openai Gpt 5 6 Sol Retune Luna Free Tier Effort Slider Unlimited Chats 2026 08 10 — OpenAI's Luna free tier at $0.20/M input tokens. DeepSeek's off-peak V4-Flash at $0.22/M is comparable but with stronger capabilities.
• Ai News Week 2026 08 03 2026 08 10 — The weekly digest that tracked the pricing compression trend. DeepSeek's structural pricing model adds a time dimension to the race.

---

9. Key Takeaways

1. Post-training is the new frontier: The same architecture can deliver 860% improvement on DeepSWE through post-training alone. This suggests that data quality and alignment matter more than raw parameter count for agentic tasks.

2. Open-weight is closing the gap: V4-Pro-0813's Terminal Bench 2.1 score of 87.9 (within 0.1 of Fable 5) demonstrates that open-weight models can compete with proprietary frontier models on coding benchmarks.

3. Harness changes the ecosystem: By providing an open-source agent framework, DeepSeek creates a sticky ecosystem that favors their models while remaining compatible with others.

4. Peak/off-peak pricing is a game-changer: The first structural time-based pricing model creates powerful incentives for workload scheduling and could become an industry standard.

5. The pricing gap is widening: Even with the price increase, V4-Pro remains 6-12× cheaper than competitors on output tokens, making it the default choice for high-volume agent workloads.

6. Reasoning effort calibration matters: The three-level effort control (low/high/max) with diminishing returns on coding tasks suggests that most production workflows should use "high" effort as the default.

---

10. Future Directions

10.1 What to Watch

• Independent benchmark validation: Third-party evaluators need to replicate DeepSeek's vendor-reported scores to confirm the Fable-level claims
• Harness adoption: The open-source framework's success will depend on community adoption and plugin ecosystem growth
• Pricing evolution: Other providers may adopt peak/off-peak pricing, creating a new dimension in the pricing war
• Open weights release: The V4 series is MIT-licensed, but the 0813 checkpoint weights have not yet been released on Hugging Face. An open-weights release would be transformative.
• Long-horizon agent benchmarks: With 1M context and 384K max output, V4-Pro is designed for sustained multi-hour agent sessions. Real-world benchmarks for these workloads are needed.

10.2 Open Questions

• How does V4-Pro-0813 perform on long-horizon tasks (100+ tool calls) compared to Claude Code and OpenAI Codex?
• Will the open-weights version of the 0813 checkpoint be released, or is it API-only?
• How will Harness evolve beyond v0.1? Will it support multi-agent orchestration?
• What is the actual cost of running a production coding agent on V4-Pro vs. Fable 5 for a typical enterprise workload?
• Will other providers adopt peak/off-peak pricing, and how will it affect the market?

---

11. References & Resources

Official Sources

• DeepSeek API: V4-Pro GA Release Announcement — Official announcement of V4-Pro-0813 GA release
• DeepSeek API: Models & Pricing — Current and upcoming peak/off-peak pricing
• DeepSeek API: Thinking Mode Guide — Reasoning effort control and multi-turn conversation patterns
• DeepSeek API: Codex Integration — One-click Codex setup with model catalog configuration
• DeepSeek API: Responses API — OpenAI Responses API format support
• DeepSeek-V4 Technical Report (arXiv:2606.19348) — Architecture paper: hybrid attention, mHC, Muon optimizer
• Hugging Face: DeepSeek-V4 Collection — Model checkpoints and cards

Related Da Claw Journal Articles

• Deepseek V4 Flash 0731 Official Release Agentic Coding Price War 2026 08 04 — V4-Flash-0731 release and the agentic coding price war
• Meta Muse Spark 1 2 Muse Code Persistent Agents Co Trained Harness 2026 08 07 — Meta's Muse Code and persistent agent capabilities
• Qwen3 8 Max 2 4t Moe Open Weight Long Horizon Autonomous Coding 2026 08 05 — Qwen3.8-Max long-horizon autonomous coding
• Meta Muse Glimmer 30b Open Agentic Local Distilled Spark Apache 2026 08 13 — Meta's local-first 30B agentic model
• Openai Gpt 5 6 Sol Retune Luna Free Tier Effort Slider Unlimited Chats 2026 08 10 — OpenAI's pricing and capability evolution
• Ai News Week 2026 08 03 2026 08 10 — Weekly context for the frontier landscape

---

This article was researched and written using only official sources: DeepSeek API documentation, the V4 technical report (arXiv:2606.19348), Hugging Face model cards, and the official DeepSeek blog. All benchmark figures are vendor-reported by DeepSeek and have not yet been independently verified by third-party evaluators.