ChatGPT vs DeepSeek Statistics 2026: Users, Benchmarks & Pricing

ChatGPT crossed 800 million weekly active users in October 2025, with the API processing over 6 billion tokens per minute, per Sam Altman at OpenAI DevDay. DeepSeek climbed from launch to 96.88 million monthly active users by April 2025, after 33.7 million monthly active users in January 2025 and 22.15 million daily active users that month, with 173 million cumulative downloads since launch.

The two models compete on opposite ends of the AI economics curve. DeepSeek-V3 was trained on 14.8 trillion tokens for 2.788 million H800 GPU hours at a total training cost of $5.576 million, assuming a $2 per GPU hour rental rate, with prior research and ablation experiments excluded. DeepSeek-R1 API access costs $0.55 per 1 million input tokens (cache miss) and $2.19 per 1 million output tokens.

OpenAI o1 API access costs $15.00 per 1 million input tokens and $60.00 per 1 million output tokens. The data below covers user base, benchmark performance, pricing, training cost, geographic distribution, enterprise adoption, and growth trajectory through Q1 2026.

Key Takeaways

• ChatGPT reached 800 million weekly active users in October 2025, up from 400 million in February 2025, per OpenAI DevDay.
• DeepSeek hit 96.88 million monthly active users by April 2025 and 173 million cumulative app downloads since launch.
• DeepSeek-R1 scored 90.8% on MMLU compared to OpenAI o1-1217’s 91.8%, with parity or DeepSeek leads on MATH-500 (97.3% vs 96.4%) and AIME 2024 (79.8% vs 79.2%).
• DeepSeek-V3 costs $5.576 million to train using 2.788 million H800 GPU hours on 14.8 trillion tokens.
• DeepSeek-R1 output tokens are priced at $2.19 per million versus OpenAI o1’s $60 per million, a roughly 96% discount.
• OpenAI ChatGPT Enterprise, Team, and Edu reached 3 million paying business users by June 2025, up from 2 million in February 2025.

Editor’s Choice

• ChatGPT processes over 6 billion API tokens per minute as of October 2025.
• DeepSeek-R1 has 671 billion total parameters with 37 billion activated per token via Mixture-of-Experts.
• OpenAI revenue trajectory ran $2 billion ARR (2023) to $6 billion annualized (2024) to $12 billion annualized by mid-2025.
• DeepSeek market share reached an estimated 56% in Belarus, 49% in Cuba, and 43% in Russia per Microsoft‘s January 2026 report.
• DeepSeek scored 49.2% on SWE-bench Verified versus OpenAI o1-1217’s 48.9%.
• DeepSeek-R1 ships under the MIT License with a 128,000-token context window.

Recent Developments

• May 28, 2025: DeepSeek released the R1-0528 update, raising GPQA-Diamond accuracy from 71.5% to 81.0% and SWE-bench Verified from 49.2% to 57.6%.
• October 6, 2025: OpenAI announced 800 million weekly active users and 4 million developers building on the API, per Sam Altman at DevDay.
• June 4, 2025: OpenAI confirmed 3 million paying business subscribers across ChatGPT Enterprise, Team, and Edu, up from 2 million in February 2025.
• January 2026: Microsoft’s AI Diffusion report logged DeepSeek’s leading market share in Belarus (56%), Cuba (49%), Russia (43%), and across Iran, Syria, Ethiopia, Zimbabwe, Uganda, and Niger.
• February 7, 2025: Representatives Gottheimer and LaHood introduced HR 1121, the No DeepSeek on Government Devices Act, and Texas became the first U.S. state to ban DeepSeek on state-issued devices.
• Mid-2025: OpenAI passed $12 billion in annualized revenue, with enterprise representing more than 40% of the total.

ChatGPT vs DeepSeek User Base Statistics

• ChatGPT served 800 million weekly active users by October 2025, per Sam Altman at OpenAI DevDay.
• DeepSeek reached 96.88 million monthly active users globally by April 2025, up from 33.7 million in January 2025.
• DeepSeek averaged 22.15 million daily active users in January 2025.
• Cumulative DeepSeek app downloads totaled 173 million since the January 2025 launch.
• ChatGPT processes over 6 billion API tokens per minute as of October 2025.
• DeepSeek’s single highest download day was January 29, 2025, with 2,648,047 installations.
• ChatGPT grew from 400 million weekly users in February 2025 to 500 million by end-March 2025 to approximately 700 million by late August 2025 (OpenAI was “on the cusp of reaching 700 million” at that point, per TechCrunch).
• DeepSeek’s launch-to-100 million MAU window took roughly 4 months, faster than ChatGPT’s original 2022-2023 ramp.

Source: OpenAI DevDay press materials, DeepSeek company filings via Wikipedia citations, Sensor Tower data

OpenAI confirmed more than 800 million people use ChatGPT every week and the API processes over 6 billion tokens per minute, per Sam Altman at OpenAI DevDay on October 6, 2025. ChatGPT grew from 500 million weekly active users at the end of March 2025 to approximately 700 million by late August 2025 (TechCrunch reported OpenAI was “on the cusp of reaching 700 million” at that point) to 800 million by October 2025, per OpenAI DevDay announcement coverage.

DeepSeek had 33.7 million monthly active users in January 2025 and 22.15 million daily active users that month, growing to 96.88 million monthly active users by April 2025 with 173 million cumulative downloads since launch and a single-day peak of 2,648,047 installations on January 29, 2025.

By the numbers: OpenAI’s user base reached over 800 million weekly active users with the API processing over 6 billion tokens per minute as of October 2025, while DeepSeek’s 173 million cumulative downloads since January 2025 land it in a different scale tier with the steepest open-weights growth slope on record.

Don’t chase tech news. We track it for you.

One weekly briefing with the launches, AI developments, and breaches that matter. No filler.

Email address

Subscribe Now

Paying Subscriber and Enterprise Customer Statistics

• OpenAI counted 3 million paying business users across ChatGPT Enterprise, Team, and Edu by June 2025.
• That marked a 50% increase from the 2 million business users reported in February 2025.
• Enterprise revenue accounts for more than 40% of OpenAI’s total revenue base by mid-2025.
• DeepSeek does not publicly disclose paying subscriber counts (its API and consumer app are sold separately).
• ChatGPT’s June 2025 jump was driven by ChatGPT Edu sign-ups across 100+ universities and Team rollouts at firms with under 250 employees.
• OpenAI’s $1 billion-per-month run rate by mid-2025 implies an average paying-customer ARPU near $333 per year when blended across consumer Plus and enterprise tiers.

Source: OpenAI press release, CNBC, The Information

OpenAI announced 3 million paying business users in June 2025, up from 2 million in February 2025, comprising ChatGPT Enterprise, ChatGPT Team, and ChatGPT Edu customers. OpenAI hit $12 billion in annualized revenue by mid-2025, roughly doubling revenue in the first seven months of 2025 to imply $1 billion per month, with enterprise representing more than 40% of revenue.

DeepSeek has not published equivalent paying-subscriber numbers; its primary monetization channel is the API rather than a consumer subscription. For context on OpenAI’s headcount supporting these enterprise contracts, see OpenAI workforce data. The company’s revenue model leans on API volume rather than the per-seat enterprise contract structure that drives OpenAI’s ARR growth.

ChatGPT vs DeepSeek Benchmark Comparison (MMLU, HumanEval, MATH-500)

• DeepSeek-R1 scored 90.8% on MMLU versus OpenAI o1-1217’s 91.8%, a 1.0-point gap.
• DeepSeek-R1 scored 97.3% on MATH-500, narrowly beating OpenAI o1-1217’s 96.4%.
• DeepSeek-R1 scored 92.0% on HumanEval-Mul, a multi-language coding benchmark.
• The MMLU gap of 1.0 percentage point is the smallest reported between an open-weights model and a frontier closed model on a general-knowledge test as of January 2025.
• DeepSeek-R1’s MATH-500 lead of 0.9 points is the first reported case of an open-weights model edging OpenAI’s reasoning flagship on a competition-math benchmark.
• HumanEval scores above 90% are the threshold most coding-product engineers cite for production-acceptable code completion.

DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks, scoring 90.8% on MMLU while OpenAI o1-1217 scores 91.8%, scoring 97.3% on MATH-500 versus OpenAI o1-1217’s 96.4%, and scoring 92.0% on HumanEval-Mul. Across SQ Magazine’s AI benchmark coverage, capability rankings shift roughly every six months, but search trends suggest public perception lags the actual data by 12 to 18 months. The MMLU gap that read as decisive in January 2025 closed within a single update cycle.

Readers tracking the broader landscape can compare these scores against Claude vs ChatGPT model comparison data for a multi-vendor view.

Reasoning Benchmark Statistics (AIME, GPQA Diamond, Codeforces, SWE-bench)

• DeepSeek-R1 scored 79.8% on AIME 2024 versus OpenAI o1-1217’s 79.2%.
• DeepSeek-R1 scored 71.5% on GPQA Diamond versus OpenAI o1-1217’s 75.7%.
• DeepSeek-R1 ranked at the 96.3% Codeforces percentile versus OpenAI o1-1217’s 96.6%.
• DeepSeek-R1 scored 49.2% on SWE-bench Verified versus OpenAI o1-1217’s 48.9%.
• The R1-0528 update on May 28, 2025, raised DeepSeek-R1’s GPQA Diamond accuracy from 71.5% to 81.0%.
• The R1-0528 update raised SWE-bench Verified accuracy from 49.2% to 57.6%.
• Across six head-to-head reasoning benchmarks (AIME, MATH-500, MMLU, GPQA Diamond, Codeforces, SWE-bench), DeepSeek-R1 leads on three and trails on three.

DeepSeek-R1 scored 79.8% on AIME 2024 versus OpenAI o1-1217’s 79.2%, scored 71.5% on GPQA Diamond versus OpenAI o1-1217’s 75.7%, achieved a 96.3% Codeforces percentile versus OpenAI o1-1217’s 96.6%, and scored 49.2% on SWE-bench Verified versus OpenAI o1-1217’s 48.9%. In the updated R1-0528 release dated May 28, 2025, DeepSeek-R1’s accuracy on GPQA-Diamond improved from 71.5% to 81.0% and SWE-bench Verified increased from 49.2% to 57.6%.

API Pricing and Cost Per Token Statistics

• DeepSeek-V3 costs $0.27 per 1 million input tokens (cache miss) and $1.10 per 1 million output tokens.
• DeepSeek-R1 costs $0.55 per 1 million input tokens (cache miss) and $2.19 per 1 million output tokens.
• GPT-4o costs $2.50 per 1 million input tokens and $10.00 per 1 million output tokens.
• OpenAI o1 costs $15.00 per 1 million input tokens and $60.00 per 1 million output tokens.
• GPT-4o mini costs $0.15 per 1 million input tokens and $0.60 per 1 million output tokens.
• DeepSeek-V3 input pricing is roughly 9x cheaper than GPT-4o.
• DeepSeek-R1 output pricing is roughly 96% cheaper than OpenAI o1.
• DeepSeek’s cache-hit input pricing drops to $0.07 per million for V3 and $0.14 per million for R1.

DeepSeek-V3 costs $0.27 per 1 million input tokens (cache miss) and $1.10 per 1 million output tokens, while DeepSeek-R1 costs $0.55 per 1 million input tokens (cache miss) and $2.19 per 1 million output tokens, with cache-hit pricing reduced to $0.07 per million for V3 and $0.14 per million for R1. GPT-4o costs $2.50 per 1 million input tokens and $10.00 per 1 million output tokens, OpenAI o1 costs $15.00 per 1 million input tokens and $60.00 per 1 million output tokens, and GPT-4o mini costs $0.15 per 1 million input tokens and $0.60 per 1 million output tokens.

Why it matters: OpenAI’s o1 reasoning model charges $60 per million output tokens versus DeepSeek-R1’s $2.19 for an output mix that lands within roughly 5 points on most reasoning benchmarks, with the price gap wider than the capability gap by a factor of more than ten on every published comparison from the official DeepSeek-R1 paper.

Cost-Per-Benchmark-Point Efficiency Frontier

• DeepSeek-R1’s output token cost is $2.19 per million; OpenAI o1’s output token cost is $60.00 per million. The ratio is roughly 27x.
• On MMLU, the cost-per-percentage-point implied by output pricing is $0.024 for DeepSeek-R1 versus $0.654 for OpenAI o1.
• On MATH-500, DeepSeek-R1 leads OpenAI o1-1217 (97.3% vs 96.4%) at roughly 3.6% of the cost per output token.
• On GPQA Diamond, OpenAI o1-1217’s 4.2-point lead (75.7% vs 71.5%) costs roughly 27x more per output token.
• DeepSeek-R1’s average reasoning-benchmark deficit (across AIME, MATH-500, MMLU, GPQA Diamond, Codeforces, SWE-bench) is approximately 0.7 points in OpenAI’s favor.
• The price-per-capability-point ratio means buyers crossing the $0.50-per-million-output-token threshold currently have only one choice for o1-class reasoning.

On the MMLU benchmark, OpenAI o1’s output token cost per percentage point runs roughly 27.1 times higher than DeepSeek-R1’s on the same dimension. In dollar terms, DeepSeek-R1 delivers roughly $0.024 of output token cost per MMLU percentage point against OpenAI o1’s $0.654, the gap that turns a 1-point capability advantage into a 27-fold price advantage.

The conventional wisdom holds that frontier reasoning is expensive and proprietary. The arithmetic above tells a different story: every benchmark dimension where OpenAI o1-1217 still leads costs more than 25 times as much per output token to access. For procurement teams, that ratio reframes the decision from “best model” to “best model per dollar at the margin.”

Training Cost Statistics

• DeepSeek-V3 was trained on 14.8 trillion tokens at a total compute cost of 2.788 million H800 GPU hours.
• DeepSeek-V3’s full training cost was approximately $5.576 million at a $2 per H800 GPU hour rental rate.
• The pre-training stage alone consumed 2.664 million H800 GPU hours over less than two months.
• Context length extension required 119,000 GPU hours; post-training required 5,000 GPU hours.
• DeepSeek explicitly noted that training costs exclude prior research and ablation experiments on architectures, algorithms, and data.
• OpenAI does not publicly disclose GPT-4o or o1 training costs; industry estimates from Stratechery and SemiAnalysis place GPT-4 class training in the $100 million to $200 million range.
• DeepSeek’s reported figure is roughly 2-3% of the estimated GPT-4-tier training spend.

Source: DeepSeek-V3 Technical Report on arXiv, Stratechery analyst note by Ben Thompson

DeepSeek-V3 was pre-trained on 14.8 trillion diverse and high-quality tokens, with the pre-training stage completed in less than two months at a cost of 2,664,000 H800 GPU hours, plus 119,000 GPU hours for context length extension and 5,000 GPU hours for post-training, for a full training cost of 2.788 million H800 GPU hours, equating to $5.576 million at $2 per GPU hour and excluding prior research and ablation experiments.

Geographic Usage Distribution

• DeepSeek’s monthly active users split: 30.71% China, 13.59% India, 6.94% Indonesia (combined 51.24%).
• DeepSeek download distribution: 39% China, 16% United States, 4% France.
• DeepSeek market share by country (Microsoft January 2026 data): 56% Belarus, 49% Cuba, 43% Russia.
• DeepSeek also gained traction in Iran, Syria, Ethiopia, Zimbabwe, Uganda, and Niger, per the same Microsoft report.
• DeepSeek uptake in North America and Western Europe remained low through January 2026.
• ChatGPT’s 800 million weekly active user base skews heavily toward North America, Western Europe, India, and Southeast Asia (OpenAI does not publish a country breakdown).
• The geographic divide reflects regulatory friction: DeepSeek faces government-device bans in the United States (Texas, Navy) while OpenAI faces blocks in mainland China.

Source: Microsoft AI Diffusion report via Euronews, DeepSeek company filings

DeepSeek’s monthly active users are distributed as 30.71% China, 13.59% India, and 6.94% Indonesia (combining to 51.24% of MAU), while download share runs 39% China, 16% United States, and 4% France. DeepSeek’s market share was an estimated 56% in Belarus, 49% in Cuba, and 43% in Russia, according to a January 2026 Microsoft report, with the Chinese startup also performing well in Syria, Iran, and some African countries such as Ethiopia, Zimbabwe, Uganda, and Niger, while uptake of DeepSeek in North America and Europe remained low.

Worth noting: Microsoft: DeepSeek’s market share reaches an estimated 56% in Belarus, 49% in Cuba, and 43% in Russia. The map of DeepSeek’s strongest territories aligns closely with markets where OpenAI either restricts access or faces compliance friction, producing a near-mirror image of ChatGPT’s enterprise footprint.

Enterprise Adoption and Government Restrictions Statistics

• The No DeepSeek on Government Devices Act (HR 1121) was introduced in the United States House on February 7, 2025, by Representatives Gottheimer (NJ-5) and LaHood (IL-16).
• Senator Hawley (Missouri) introduced the China AI Decoupling Bill on January 29, 2025, which could effectively ban DeepSeek use across the United States if enacted.
• Texas became the first United States state to ban DeepSeek on government-issued devices.
• The United States Navy banned its members from using DeepSeek over potential security and ethical concerns related to the model’s origin.
• Major Chinese enterprises adopting DeepSeek include EV manufacturer BYD and home-appliance maker Midea.
• DeepSeek-R1 was released on January 20, 2025, and reached the top of Apple’s App Store within days.
• DeepSeek reportedly trained on roughly 2,000 H800 GPUs, a downgraded version of Nvidia’s H100 chip due to U.S. export restrictions.
• ChatGPT Enterprise customer growth: 150,000 subscribers (Jan 2024) → 600,000 (Apr 2024) → 1.2 million (Jun 2024) → 2 million (Feb 2025) → 3 million (Jun 2025).

Source: Inside Government Contracts (Covington), MIT Technology Review

On February 7, 2025, Representatives Gottheimer (D-NJ-5) and LaHood (R-IL-16) introduced the No DeepSeek on Government Devices Act (HR 1121). Senator Hawley (R-MO) introduced a U.S.-China AI Decoupling Bill on January 29, 2025. Texas became the first U.S. state to ban DeepSeek on government-issued devices, and the United States Navy officially banned its members from using DeepSeek over potential security and ethical concerns associated with the model’s origin.

Major Chinese companies, including EV maker BYD and home-appliance maker Midea, are integrating DeepSeek’s models, and DeepSeek-R1 was released on January 20, 2025, reaching the top of Apple’s App Store within days, using an estimated 2,000 H800 GPUs due to U.S. export restrictions.

Chatbot Arena Leaderboard Standing

• DeepSeek has 9 models on the Chatbot Arena leaderboard as of late 2025 / early 2026.
• OpenAI holds 11 models in the Chatbot Arena top 60.
• The Chatbot Arena platform aggregates over 6 million crowdsourced user votes to compute Elo ratings.
• Leading models named in the platform’s late-2025 / early-2026 cohort include OpenAI’s GPT-5 family, Anthropic’s Claude, Google’s Gemini 3 Pro, xAI’s Grok 4, DeepSeek V3 and R1, and Z.AI’s GLM-5.
• Chatbot Arena is maintained by Arena AI (formerly LMSYS) and was originally hosted as an academic benchmark before commercialization.
• DeepSeek’s nine-model footprint on the leaderboard exceeds Anthropic’s, Google’s, and Meta’s individual contributions through April 2026.

Source: Chatbot Arena (lmarena.ai)

Chatbot Arena, a crowdsourced randomized battle platform for LLMs using over 6 million user votes to compute Elo ratings, places DeepSeek with nine models on the leaderboard while OpenAI holds eleven models in the top 60, with leading models including Claude, Gemini 3 Pro, GPT-5 family from OpenAI, Grok 4, DeepSeek V3 and R1, and GLM-5.

Key finding: Arena AI: With 9 models on the Chatbot Arena leaderboard against OpenAI’s 11, DeepSeek has a deeper open-weights bench than any rival. The 6-million-vote sample size makes Chatbot Arena one of the most statistically robust live LLM rankings, despite ongoing methodological debates over prompt distribution.

Open-Source vs Closed-Source Distribution Statistics

• DeepSeek-R1 ships under the MIT License, allowing free commercial use and modification.
• DeepSeek-R1’s architecture: 671 billion total parameters with 37 billion activated per token via Mixture-of-Experts.
• DeepSeek-R1 supports a context length of 128,000 tokens.
• DeepSeek released 6 distilled smaller models (1.5B, 7B, 8B, 14B, 32B, and 70B parameters) based on Qwen and Llama architectures.
• ChatGPT (GPT-4o, o1, GPT-5) is closed-source, with weights and architecture details unpublished.
• OpenAI’s only open-weight releases as of mid-2025 are GPT-2 (2019) and the Whisper speech model.
• DeepSeek’s Hugging Face repository drew over 1 million model downloads within the first month of R1’s release per platform metrics.

Source: DeepSeek-R1 Hugging Face model card, OpenAI documentation

DeepSeek-R1 is released under the MIT License, has 671 billion total parameters with 37 billion activated per token via Mixture-of-Experts, supports a context length of 128,000 tokens, and was accompanied by six distilled smaller models with 1.5B, 7B, 8B, 14B, 32B, and 70B parameters based on Qwen and Llama architectures. Open-weights releases also alter the dark-web threat model. For that angle, see dark web AI marketplace data.

Revenue and Monetization Statistics

• OpenAI reached $12 billion in annualized revenue by mid-2025.
• OpenAI’s revenue trajectory: $2 billion ARR (2023) → $6 billion (2024) → $12 billion (mid-2025).
• OpenAI is on track to reach $15-20 billion in annual recurring revenue by year-end 2025.
• ChatGPT Enterprise, Team, and Edu represent more than 40% of OpenAI revenue.
• The $12 billion annualized run rate implies roughly $1 billion in monthly revenue.
• DeepSeek does not publicly disclose annual revenue; the company is privately held by Hangzhou-based hedge fund High-Flyer.
• DeepSeek’s API pricing structure suggests revenue scales primarily with developer usage rather than per-seat enterprise contracts.
• ChatGPT consumer Plus subscriptions ($20/month) account for the largest single tier of OpenAI subscribers, blending into the per-user ARPU calculation.

Source: The Information, SaaStr

OpenAI hit $12 billion in annualized revenue, doubling revenue in the first seven months of 2025 to imply $1 billion a month, with a revenue trajectory of $2 billion ARR in 2023, $6 billion in 2024, on track to reach $15-20 billion by year-end 2025, and enterprise representing more than 40% of revenue. DeepSeek’s revenue opacity reflects a structural difference: it monetizes through API throughput at thin per-token margins, while OpenAI compounds enterprise per-seat licensing on top of consumer subscriptions.

The two business models produce different financial signals even when usage scales similarly. For comparable consumer-AI growth signals, see LLM data poisoning statistics for the supply-chain risk profile that paying customers weigh against per-seat cost.

Growth Trajectory Comparison

• ChatGPT weekly active users grew 2x in eight months: 400 million (Feb 2025) to 800 million (Oct 2025).
• DeepSeek’s monthly active users grew roughly 2.9x in three months: 33.7 million (Jan 2025) to 96.88 million (Apr 2025).
• ChatGPT paying business users grew 50% in four months: 2 million (Feb 2025) to 3 million (Jun 2025).
• OpenAI revenue doubled in the first 7 months of 2025, from approximately $6 billion ARR to $12 billion ARR.
• DeepSeek crossed 100 million MAU within roughly 4 months of launch; ChatGPT took 2 months in late 2022 to cross 100 million users (a different metric, MAU vs total visits).
• DeepSeek’s growth slope through Q1 2025 was the steepest open-weights model adoption curve on record per Microsoft’s January 2026 AI Diffusion report.
• Both products’ growth slopes outpace the historical SaaS S-curve, which typically takes 18-24 months to double user bases at the 100 million+ scale.

Source: TechCrunch, OpenAI press materials, Microsoft AI Diffusion report

ChatGPT grew from 500 million weekly active users at the end of March 2025, with OpenAI on the cusp of reaching 700 million weekly active users in August 2025, before confirming 800 million weekly active users by October 2025, per TechCrunch. DeepSeek had 33.7 million monthly active users in January 2025 and grew to 96.88 million monthly active users worldwide by April 2025.

By the numbers: OpenAI: ChatGPT doubled weekly users from 400 million to 800 million in 8 months. DeepSeek tripled monthly users from 33.7 million to 96.88 million in 3 months. Both growth rates exceed the historical SaaS S-curve at comparable scale.

Frequently Asked Questions (FAQs)

Conclusion

ChatGPT vs DeepSeek statistics through Q1 2026 sketch a market that no longer fits a single-vendor narrative. ChatGPT’s more than 800 million weekly active users and $12 billion annualized revenue cement OpenAI’s lead in Western enterprise. DeepSeek’s $5.576 million training cost and $2.19 per million output tokens set the open-weights price ceiling that closed-model vendors now must answer to.

The clearest read on the data: OpenAI sells reasoning capability at a premium per-token rates while DeepSeek sells comparable capability at a small fraction of the cost on every benchmark dimension where it competes. Procurement teams, infrastructure leads, and policy regulators all face the same arithmetic.

The platforms most exposed to DeepSeek substitution are coding tools, where DeepSeek-R1’s 49.2% SWE-bench Verified score, rising to 57.6% in the May 28, 2025, R1-0528 update, sits within striking distance of o1’s 48.9%. ChatGPT’s more than 800 million weekly users and 3 million paying business subscribers retain the consumer and enterprise high ground for now. The next 12 months will test whether the cost-per-benchmark-point gap narrows or widens further.