DeepSeek V4 Preview is free on web, app, and API, open-sourced, with a 1M-token context. ChatGPT runs GPT-5.5, free at 10 messages every 5 hours or $20/mo for Plus. The split is open weights and zero cost versus a polished, multimodal, frontier product.
The Trade in One Line
ChatGPT is the finished product. It runs GPT-5.5, generates images and video with Sora, offers Custom GPTs, agents, Canvas, Advanced Data Analysis, and Deep Research, and ships a refined consumer app. You pay frontier prices and your data lives on OpenAI's servers.
DeepSeek is the open-weight bet. DeepSeek V4 Preview is open-sourced and free on web, app, and API. You can download the weights, inspect them, fine-tune, and self-host so data never leaves your infrastructure. It is text and reasoning focused, without ChatGPT's image, video, and agent ecosystem. See our take on DeepSeek V4.
Pricing: Every Tier
| Plan | ChatGPT | DeepSeek |
|---|---|---|
| Free | $0/mo, ad-supported (US), 10 GPT-5.5 msgs / 5h | $0, no ads, no in-app purchases |
| Entry paid | Go: $8/mo, ad-supported (98 countries) | Free (no paid consumer tier) |
| Standard | Plus: $20/mo | Free |
| Pro | Pro: $100/mo, Pro Max: $200/mo | Free |
| Business / team | $20/user/mo annual, $25 monthly, 2+ users | Self-host (open weights) |
| API access | Frontier tier, hosted | Free, open-sourced |
| Weights | Closed | Open |
ChatGPT Pro tiers
Pro at $100/mo gives 5x higher usage than Plus, elevated Codex limits, and 125 Deep Research queries per month. Pro Max at $200/mo gives 20x usage, GPT-5 Pro, and unlimited audio/video mode. ChatGPT Business dropped to $20/user/mo (annual) on April 2, 2026.
Models and Context Windows
| Spec | ChatGPT (GPT-5.5) | DeepSeek V4 |
|---|---|---|
| Flagship | GPT-5.5 (live in ChatGPT) | DeepSeek V4 Pro |
| Parameters | Not disclosed | Pro: 1.6T total / 49B activated; Flash: 285B / 13B activated |
| Context window | Not disclosed (consumer app) | 1M tokens (Pro and Flash) |
| Reasoning modes | GPT-5.5 Thinking (manual select) | Three reasoning modes per variant |
| Open source | No | Yes (V4 Preview) |
| Multimodal | Image gen + Sora video + Vision | Text and reasoning focus |
GPT-5.5 reaches 58.6% on SWE-Bench Pro (real-world GitHub issue resolution) and 82.7% on Terminal-Bench 2.0 (complex command-line workflows). GPT-5.5 and GPT-5.5 Pro became available in the API on April 24, 2026. DeepSeek V4 ships two variants tuned for cost versus capability: Pro for maximum quality, Flash for cheap high-throughput inference, both with the same 1M-token context and three reasoning modes.
Free Access Compared
DeepSeek's free tier is the whole product. The app and web are free with no ads and no in-app purchases, including web search, Deep-Think mode, and file upload with text extraction. The API is free and open-sourced.
ChatGPT Free is $0/month with ads in the US and caps GPT-5.5 at 10 messages every 5 hours. ChatGPT Go at $8/mo is ad-supported and available in 98 countries (including the EU) since January 2026. To remove the GPT-5.5 message cap and unlock agents, Sora, and Deep Research, you move to Plus at $20/mo.
Rate Limits
| Tier | ChatGPT | DeepSeek |
|---|---|---|
| Free | 10 GPT-5.5 messages every 5 hours | Free app/web usage |
| Plus / Go | 160 GPT-5.5 messages every 3 hours, then mini until reset | Free |
| Thinking model | Up to 3,000 GPT-5.5 Thinking messages per week (Plus/Business) | Three reasoning modes, no public cap |
Data and Privacy
On ChatGPT, model training (the "Improve the model for everyone" toggle) is on by default for consumer accounts. You can opt out via Settings, Data Controls, then switch off "Improve the model for everyone," after which new conversations are not used to train models. The setting applies to the entire account regardless of device and can be changed anytime; conversations still appear in chat history but are not used for training.
DeepSeek's training data is produced by its research team, with a small portion potentially based on user input. If user input is used to construct training data, it is encrypted, de-identified, and anonymized. Users can opt out of data usage for model training and delete historical data. Self-hosting the open weights removes the question entirely, since requests never reach DeepSeek's servers.
| Dimension | ChatGPT | DeepSeek |
|---|---|---|
| Training on by default | Yes (consumer) | Partial, encrypted and anonymized if used |
| Opt out of training | Yes, account-wide, anytime | Yes |
| Delete history | Yes | Yes |
| Self-host option | No (hosted only) | Yes (open weights) |
Run DeepSeek at Full Fidelity
DeepSeek V4 is open-sourced, but running 1.6T-parameter weights yourself means a GPU fleet. Most serverless providers serve DeepSeek with fp8 quantized activations to cut cost, which degrades output quality away from the reference weights. Morph serves DeepSeek with 16-bit (bf16) activations, no fp8 or int8 quantization, so responses match the reference model.
For coding agents specifically, Morph runs codegen-tuned speculative decoding (draft and ngram tuned on code) plus custom low-level inference kernels, which makes it the fastest and highest-quality option for code generation rather than a general-purpose menu. DeepSeek V4 Flash (morph-dsv4flash) is $0.139 per 1M input tokens and $0.278 per 1M output tokens.
16-bit activations
bf16, no fp8/int8 quantization. Output matches the reference weights.
Codegen-tuned
Speculative decoding and custom kernels tuned for code generation.
$0.139 / $0.278
morph-dsv4flash per 1M input / output tokens.
See Morph Open Source Models and pricing for the full lineup.
Pick by Priority
| Your priority | Best choice | Why |
|---|---|---|
| Zero cost | DeepSeek | Free app, web, and API. No ads, no purchases. |
| Open weights / self-host | DeepSeek | V4 is open-sourced; run it in your own infra. |
| Long context | DeepSeek | 1M-token context on both V4 variants. |
| Image or video generation | ChatGPT | Sora video and native image gen; DeepSeek is text-focused. |
| Agents and Custom GPTs | ChatGPT | ChatGPT Agent, Codex Agent, Custom GPTs, Canvas. |
| DeepSeek without GPUs | Morph | 16-bit DeepSeek V4 Flash at $0.139/$0.278 per 1M tokens. |
Frequently Asked Questions
Is DeepSeek free?
Yes. The app and web are free with no ads or in-app purchases, and DeepSeek V4 Preview is free on the API and open-sourced. ChatGPT Free is ad-supported in the US and caps GPT-5.5 at 10 messages every 5 hours.
What is DeepSeek V4's context window?
1M tokens on both variants. Pro has 1.6T total parameters (49B activated per token); Flash has 285B (13B activated). Both support three reasoning modes.
How many free messages does ChatGPT give?
Free: up to 10 GPT-5.5 messages every 5 hours. Plus and Go: up to 160 GPT-5.5 messages every 3 hours, then the mini model until reset. GPT-5.5 Thinking allows up to 3,000 messages per week on Plus or Business.
Can I self-host DeepSeek?
Yes, V4 is open-sourced. To skip the GPU fleet, Morph serves DeepSeek V4 Flash with full 16-bit activations at $0.139 per 1M input and $0.278 per 1M output tokens.
What does ChatGPT do that DeepSeek does not?
Image generation, Sora video, Vision, Custom GPTs, ChatGPT Agent, Codex Agent, Canvas, Advanced Data Analysis, and Deep Research. DeepSeek is text and reasoning focused.
Related comparisons
ChatGPT vs Gemini
GPT-5.5 vs Gemini 3.1 Pro: benchmarks, pricing, multimodal, and when to route to each.
ChatGPT vs Microsoft Copilot
Copilot runs on OpenAI models inside Microsoft 365. Where the wrapper helps and where it taxes you.
ChatGPT vs Grok
GPT-5.5 vs Grok: real-time X data and fewer guardrails vs ecosystem maturity.
ChatGPT vs Perplexity
General assistant vs answer engine: cited search vs open-ended generation.
Gemini vs DeepSeek
Multimodal Google model vs open-weight reasoning at a fraction of the API cost.
ChatGPT vs Claude vs Gemini
The three frontier assistants compared on coding, writing, multimodal, and price.
Run DeepSeek at Full 16-Bit Fidelity
Morph serves DeepSeek V4 Flash with bf16 activations (no fp8 quantization) plus codegen-tuned speculative decoding. morph-dsv4flash is $0.139 per 1M input tokens and $0.278 per 1M output tokens.