DeepSeek V3 vs GPT-4o: Benchmarks, Price and Verdict

DeepSeek V3 vs GPT-4o is a historical comparison between a December 2024 open-weight model from a Chinese lab and OpenAI’s May 2024 multimodal flagship. The deepseek v3 vs gpt-4o question was genuinely close on text benchmarks, and the price gap was large. Both models have since been superseded: DeepSeek’s API now serves deepseek-flash and deepseek-v4-pro, and OpenAI has deprecated GPT-4o, with shutdown scheduled for 14 January 2027. This page sets out the head-to-head numbers from primary sources, costs out a realistic workload, and explains what replaced each model.

Verdict up front: which model wins, and for whom

DeepSeek V3 won on text workloads at scale: RAG pipelines, code assistants, batch summarisation, internal tooling. In DeepSeek’s own comparison table it matched or beat GPT-4o-0513 on most knowledge, code and maths benchmarks, its weights were downloadable, and its standard API rates ($0.27 input on a cache miss / $1.10 output per 1M tokens) were roughly a ninth of GPT-4o’s $2.50 / $10.00.

GPT-4o won where V3 could not compete: image input and audio (through OpenAI’s separate GPT-4o audio and realtime models), integration with the OpenAI ecosystem, and procurement settings where a China-hosted endpoint was not acceptable. DeepSeek’s own table also shows GPT-4o-0513 ahead on factual-recall tests such as SimpleQA.

What this means in 2026: neither is a model to start a new integration on. On DeepSeek’s side, V3 was followed by V3.1 (August 2025), V3.2 (December 2025), the V4 preview (April 2026), and today’s deepseek-flash (DeepSeek-V4.1-Flash-1017, 17 October 2026) and deepseek-v4-pro (DeepSeek-V4-Pro-0813, 13 August 2026). On OpenAI’s side, the deprecations page lists GPT-4o (gpt-4o-2024-08-06 and gpt-4o-2024-11-20) for shutdown on 14 January 2027 with gpt-6-luna as the replacement, and the May 2024 snapshot used in the benchmark tables, gpt-4o-2024-05-13, was shut down on 23 October 2026. Read this comparison for the architectural and pricing lessons, then see the current-generation pages linked at the end.

At a glance: the spec sheet

Architecture: MoE vs dense, and why it matters for cost

DeepSeek V3 is a sparse Mixture-of-Experts model. Its model card describes 671B total parameters with 37B activated for each token, using the Multi-head Latent Attention (MLA) and DeepSeekMoE architectures validated in DeepSeek-V2. Only a fraction of the network runs for any given token, which is what let DeepSeek serve V3 at its prices.

OpenAI has not published GPT-4o’s architecture, parameter count or training compute. On the API, GPT-4o accepts text and image input and returns text; audio was offered through separate GPT-4o audio and realtime models, which OpenAI is retiring on 20 January 2027. V3 is text-only, so for vision or voice products the comparison ended there and GPT-4o won by default.

Benchmarks: the head-to-head numbers

The most-cited numbers come from DeepSeek’s V3 model card and the R1 technical report that followed it. Both compare against GPT-4o-0513 (the May 2024 snapshot) and Claude-3.5-Sonnet-1022; pin those versions when you read the numbers, because OpenAI updated GPT-4o several times afterwards. These are DeepSeek-reported results.

Knowledge and reasoning

V3 edges out GPT-4o-0513 on general knowledge (MMLU, MMLU-Pro) and clearly beats it on graduate-level science questions (GPQA-Diamond), while Claude-3.5-Sonnet-1022 leads on MMLU-Pro and GPQA-Diamond. GPT-4o-0513 leads on SimpleQA, a short-answer factual-recall test. For “explain this regulation” or “summarise this paper” workloads, the knowledge scores of V3 and GPT-4o-0513 sit within a few points of each other.

Code and math

Code and maths are where V3 looked strongest at launch. In the same table, V3 scores 82.6 on HumanEval-Mul against 80.5 for GPT-4o-0513 and 81.7 for Claude-3.5-Sonnet-1022. On maths the gap is wider: 39.2 vs 9.3 on AIME 2024 and 90.2 vs 74.6 on MATH-500. Always name the source table and the benchmark variant (HumanEval-Mul here, not plain HumanEval), because different evaluation harnesses give different numbers.

For deeper multi-step reasoning, neither model was the right tool; that is what DeepSeek R1 and OpenAI’s o1 were for at the time. See our DeepSeek for coding use-case guide for current options.

Pricing: the gap that defined the V3 launch

This is where the comparison was most uncomfortable for OpenAI.

V3’s rates are history. DeepSeek’s current rates (USD per 1M tokens, as of October 2026) are deepseek-flash: $0.006 cache hit / $0.30 cache miss / $1.20 output at peak and $0.003 / $0.15 / $0.60 off-peak; deepseek-v4-pro: $0.044 / $1.32 / $3.96 at peak and $0.022 / $0.66 / $1.98 off-peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Note that deepseek-flash costs more per token than V3 did on cache-miss input at peak ($0.30 vs $0.27) and on output ($1.20 vs $1.10); it is cheaper on cache hits and off-peak. Always confirm on the DeepSeek API pricing page before committing.

Worked example: 1M API calls per month

A representative chat workload: a 2,000-token system prompt (cached), a 200-token user message (uncached) and a 300-token response, repeated 1,000,000 times in a month. That is 2,000M cached input tokens, 200M uncached input tokens and 300M output tokens.

GPT-4o (gpt-4o) at its listed rates ($1.25 cached input / $2.50 input / $10.00 output):

• Cached input: 2,000M × $1.25/M = $2,500.00
• Uncached input: 200M × $2.50/M = $500.00
• Output: 300M × $10.00/M = $3,000.00
• Total: $6,000.00

DeepSeek V3 at its standard rates from 8 February 2025 ($0.07 cache hit / $0.27 cache miss / $1.10 output):

• Cached input: 2,000M × $0.07/M = $140.00
• Uncached input: 200M × $0.27/M = $54.00
• Output: 300M × $1.10/M = $330.00
• Total: $524.00, about 1/11 of the GPT-4o bill

Today’s deepseek-flash:

• Cached input: 2,000M × $0.006/M = $12.00 peak; × $0.003/M = $6.00 off-peak
• Uncached input: 200M × $0.30/M = $60.00 peak; × $0.15/M = $30.00 off-peak
• Output: 300M × $1.20/M = $360.00 peak; × $0.60/M = $180.00 off-peak
• Total: $432.00 at peak; $216.00 off-peak

GPT-4o’s named replacement, gpt-6-luna ($0.01 cached / $0.10 input / $0.50 output per 1M on OpenAI’s pricing page), would come to $20.00 + $20.00 + $150.00 = $190.00 before cache-write charges, below deepseek-flash on this workload. If you want to plug your own numbers in, use the DeepSeek pricing calculator. The DeepSeek API is stateless: chat requests go to POST /chat/completions, and the client must resend the conversation history with every call. The web chat keeps history for you; the API does not.

Coding head-to-head

Beyond HumanEval-Mul, the V3 model card’s comparison table gives a fuller picture of coding (DeepSeek-reported, December 2024):

• SWE-Bench Verified (resolved): V3 42.0, GPT-4o-0513 38.8, Claude-3.5-Sonnet-1022 50.8. On real repository issues, both trailed Claude.
• LiveCodeBench (Pass@1, chain of thought): V3 40.5, GPT-4o-0513 33.4.
• Codeforces (percentile): V3 51.6, GPT-4o-0513 23.6.
• Aider-Edit (accuracy): V3 79.7, GPT-4o-0513 72.9; Claude-3.5-Sonnet-1022 led at 84.2.

The table also notes that all models were evaluated with output limited to 8K tokens. The practical consequence was cost per resolved task: at roughly a ninth of GPT-4o’s token price, V3 made retry-heavy agent loops affordable.

Reasoning, writing, and the things benchmarks miss

Neither V3 nor GPT-4o is a “thinking” model; both answer without a separate reasoning phase. For genuinely hard reasoning (competition maths, multi-step proofs, planning under constraints) the choice at the time was DeepSeek R1 or OpenAI’s o-series. Today, both DeepSeek API models run in thinking mode by default and return reasoning_content alongside the final content.

For open-ended writing and conversation, DeepSeek’s table reports V3 ahead of GPT-4o-0513 on Arena-Hard (85.5 vs 80.4) and AlpacaEval 2.0 (70.0 vs 51.1), and close on instruction following (IF-Eval 86.1 vs 84.3), while GPT-4o-0513 led on FRAMES (80.5 vs 73.3). Benchmarks do not capture tone or house style, so test both against your own prompts. Our DeepSeek for writing guide has prompt patterns.

Privacy, jurisdiction, and the procurement conversation

Be clear about where the data goes. DeepSeek’s Privacy Policy (last updated 10 February 2026) says it collects, processes and stores personal data in the People’s Republic of China. OpenAI documents its data handling and regional-processing options in its “Your data” guide. For most internal-tool workloads this may not matter; for regulated workloads (health, finance, legal, government) it can, and DeepSeek’s hosted API may not be allowed at all.

The mitigation for DeepSeek is that the weights are published on Hugging Face, so you can self-host and remove the jurisdiction question. V3’s weights carry the DeepSeek Model License; DeepSeek’s current models, DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro-0813, are MIT-licensed. That is not an option with GPT-4o. See how to install DeepSeek locally for the hardware reality.

Ecosystem: where GPT-4o still pulls ahead

OpenAI’s developer ecosystem is denser: SDKs, vector-store integrations, agent frameworks, IDE plugins and third-party tools tend to assume OpenAI first. DeepSeek offsets this by being OpenAI SDK-compatible: point the OpenAI client at https://api.deepseek.com, change the model ID, and most code works unchanged. DeepSeek also accepts the Anthropic API format at https://api.deepseek.com/anthropic.

import os
from openai import OpenAI

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

resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Summarise this PR."}],
)
print(resp.choices[0].message.content)

V3 itself is no longer an API model. Integrations written for deepseek-chat, the ID that served V3, need the model string changed to deepseek-flash or deepseek-v4-pro; the base_url stays the same.

When to pick which: a decision table

Alternatives worth considering

If you are reading this in 2026 to make a current decision, neither V3 nor GPT-4o is the model to pick. Current comparisons:

• For DeepSeek’s current models: DeepSeek V4-Pro and our DeepSeek vs ChatGPT breakdown, which covers OpenAI’s GPT-6 lineup.
• For reasoning: DeepSeek R1 vs OpenAI o1 covers the thinking-model split of the same era.
• For Anthropic-curious teams: DeepSeek vs Claude.
• Browse the wider field: our AI comparison hub has the rest of the head-to-heads.