My ChatGPT Plus bill renews on the 14th of every month. Twenty dollars. I've paid it without thinking for almost two years. It's less than my Netflix subscription, and I use it far more than Netflix, so it's never felt like a hard decision.
Then in late April 2026, DeepSeek V4 launched — completely free, with benchmarks claiming it matched or beat ChatGPT on coding and reasoning — and my Twitter feed basically exploded. Developers I respect were posting screenshots of $0.19 API bills for work that would have cost $40+ on Claude. Someone called it "the most important AI release of the year." Someone else said it was "Chinese spyware dressed up as a chatbot."
I decided the only honest way to find out which crowd was right was to actually use it. So I cancelled ChatGPT Plus, set DeepSeek as my default AI for everything, and gave it 30 days.
What followed was more complicated than I expected. Here's the full, unfiltered version.
First: What Is DeepSeek V4 and Why Is Everyone Talking About It?
DeepSeek is a Chinese AI lab — a subsidiary of the hedge fund High-Flyer — that first shocked the AI world in January 2025 when its R1 reasoning model matched far more expensive American models for a tiny fraction of their training cost. The widely cited $6 million training figure was actually for the V3 base model; R1 itself reportedly cost just $294,000 to train. For context, OpenAI has spent billions on training runs for GPT-5.x.
DeepSeek V4 launched April 24, 2026 — exactly one day after OpenAI released GPT-5.5. The timing felt deliberate. V4 ships as two versions: V4-Pro (1.6 trillion parameters, 49 billion active per token via a mixture-of-experts architecture) and V4-Flash (284 billion parameters, faster and cheaper). Both support a 1 million token context window. Both are free in the chat app. Both are open-weight — meaning developers can download and run them on their own hardware.
The price comparison is what stops people mid-sentence:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Consumer Plan |
|---|---|---|---|
| DeepSeek V4-Flash | $0.14 | $0.28 | Free (no limits) |
| DeepSeek V4-Pro | $0.44 | $0.87 | Free (no limits) |
| ChatGPT (GPT-5.5) | $5.00 | $30.00 | $20/month (Plus) |
| Claude Opus 4.8 | $5.00 | $25.00 | $20/month (Pro) |
DeepSeek V4-Flash output tokens cost $0.28 per million. GPT-5.5 costs $30 per million. That's not a small difference. That's more than a 100x price gap. At API scale, that single number reshapes every build decision a developer makes.
That's what made me want to actually test it. Could something this cheap really be this good?
Week 1 — The Honeymoon Phase
Day one, I was impressed. Genuinely.
I threw DeepSeek my usual morning warm-up: summarize three long articles I'd bookmarked, draft a reply email to a client, and explain a Python error I'd been ignoring for two days. It handled all three cleanly. The summaries were accurate and concise. The email draft was professional and hit the right tone. The Python fix was correct on the first try.
I remember thinking: okay, this might actually work.
The coding was what surprised me most. V4-Pro has an 80.6% score on SWE-bench Verified — the real-world software engineering benchmark — which puts it close to Claude Opus 4.8's 69.2% on SWE-bench Pro. In practical terms, this means it handles real code problems well, not just toy examples. I used it to refactor a messy data pipeline, write a browser extension from scratch, and debug a race condition in an async function. All three: clean, fast, largely correct on first pass.
On math and structured reasoning, DeepSeek has always been strong. V4-Pro scores roughly 92% on competition math benchmarks versus GPT-5.5's 88%. If you do anything quantitative — financial modeling, statistics, data analysis — DeepSeek is genuinely competitive at the top tier.
The interface is simple. Maybe too simple — no image generation, no voice mode, no integrations. But for pure text and code work, week one felt like I'd found a sports car that someone left running with the keys in the ignition.
✅ Week 1 verdict: For coding, math, and structured reasoning tasks — DeepSeek V4 competes directly with ChatGPT and occasionally beats it. Free. No subscription. The performance-to-cost ratio is genuinely unlike anything else available right now.
Week 2 — The Cracks Start Showing
The writing quality gap showed up in week two, and it showed up clearly.
I write regularly — articles, client deliverables, email newsletters. When I run the same prompt through ChatGPT and DeepSeek on creative or persuasive writing tasks, the gap is noticeable. DeepSeek's prose is competent. It's accurate. But it reads slightly differently — a little more mechanical in its sentence variety, a little less attuned to tone when the instructions are subtle.
Reviewers who've tested this independently describe V4-Pro as occasionally "over-explaining" when a shorter, sharper answer would serve better. It also "misses the intent behind underspecified requests more often than Claude does." Both of those showed up in my experience. When I gave explicit, detailed instructions, DeepSeek performed well. When I expected it to read between the lines the way ChatGPT often does — less so.
The reliability issue also surfaced in week two. DeepSeek's servers occasionally struggle during peak hours. The "retry tax" is real: sometimes you submit a prompt, get a timeout or error, and have to resubmit. This happened to me maybe 8–10 times across the month. It's not constant, but it's consistent enough to notice. ChatGPT — especially on Azure infrastructure — almost never drops a request. That reliability difference matters more than it sounds when you're in the middle of something.
The third thing I noticed was what I started calling "the wall."
Ask DeepSeek about certain topics — anything related to Chinese government, Tiananmen Square, Tibet, Taiwan's independence, Uyghur detention, or political criticism of Chinese leadership — and it either refuses, gives a heavily neutral non-answer, or simply changes the subject. This isn't a keyword filter. Research has shown the censorship is baked into the model weights through supervised fine-tuning. It's not a bouncer at the door. It's in the model's actual learned behavior.
For most everyday use, this never comes up. But it came up twice for me during news research tasks, and both times it was jarring. You ask a question and instead of an answer, you get silence or deflection. It reminded me that this tool was built under a different set of constraints than the ones I'm used to.
⚠️ Week 2 reality check: The writing quality gap is real. The server reliability is noticeably worse than ChatGPT. And the political censorship — while irrelevant for most tasks — is a real limitation that's structurally built into the model, not something that can be prompted around.
Week 3 — The Privacy Question I Couldn't Ignore
I'd been putting off the privacy research. Week three, I couldn't anymore.
Here's what DeepSeek's own privacy policy says it collects: your chat history and all prompts, account information, device data including hardware model and OS, network information including IP address and carrier, keystroke patterns, location data, and cookies and usage analytics. All of this is stored on servers in the People's Republic of China.
Under China's National Intelligence Law and Cybersecurity Law, the Chinese government can legally demand access to data held by Chinese companies — including DeepSeek — at any time, for national security purposes. The company is not required to notify users when this happens, and cannot refuse.
The list of organizations that have banned DeepSeek is long and, for an American user, worth reading carefully:
- NASA — blocked from all agency systems
- US Navy — all use prohibited, including personal devices
- Pentagon — blocked after unauthorized staff access incidents
- US House of Representatives — functionality restricted on official devices
- US Department of Commerce — banned on all government-furnished equipment
- Microsoft — employees prohibited from using DeepSeek
- Countries with full or partial bans: Italy, Australia, Taiwan, South Korea
There was also a documented data breach shortly after DeepSeek's January 2025 launch — over one million sensitive records exposed. And independent security researchers found a 100% jailbreak success rate in early testing, meaning the model's safety filters could be bypassed completely with standard techniques.
I want to be fair here. American AI companies — OpenAI, Anthropic, Google — are not privacy saints either. ChatGPT uses your conversations to improve its models unless you opt out. These companies are subject to US legal demands including NSA data requests. The difference is one of legal framework, not purity: Chinese law gives the government broader, less judicially constrained access to data than US law. That's a real distinction for anyone who uses AI to process sensitive work.
During week three, I started being more careful about what I put into DeepSeek. Business strategy. Client information. Personal financial data. Anything I wouldn't want on a Chinese server — which, when I actually thought about it, was more of my daily work than I'd initially realized.
💡 The privacy reality: For casual questions, public information, and coding tasks that don't involve proprietary code — the privacy risk is relatively low. For business strategy, client work, personal financial data, medical information, or anything sensitive — you're sending that data to servers under Chinese government jurisdiction. That's a decision, not an oversight. Make it consciously.
Week 4 — Finding My Actual Answer
By week four I'd stopped trying to declare a winner and started figuring out where DeepSeek actually fit in my workflow.
The honest answer is: it fit well in some places and poorly in others, and the line between those two categories was clearer than I expected.
DeepSeek V4 was legitimately excellent for:
- Coding tasks where the logic is well-defined. Write this function. Debug this error. Explain this architecture. Refactor this script. On tasks like these, I genuinely could not tell the difference between V4-Pro and GPT-5.5 in most cases. The $19 cents vs $42 comparison from that developer's weekend project is not an exaggeration — the cost efficiency is real.
- Math, data analysis, and quantitative reasoning. DeepSeek showed its strongest benchmark scores here, and they held up in practice. If you work with numbers, formulas, or statistical reasoning, this model is competitive with anything.
- Long document processing. The 1 million token context window handled everything I threw at it — long PDFs, multi-file code reviews, extended research notes. No degradation in coherence at long context lengths that I could detect.
- Non-sensitive research and information retrieval. For factual questions where the topic isn't politically sensitive, DeepSeek is fast, accurate, and free.
DeepSeek V4 was clearly worse than ChatGPT for:
- Writing that needs voice and nuance. Anything going in front of real humans — articles, client emails, marketing copy, anything where "reads like a person wrote it" matters — ChatGPT (and especially Claude) produces more natural output.
- Multimodal tasks. DeepSeek is text-only. No image generation, no image understanding (in the chat interface), no voice mode. ChatGPT includes DALL-E 3, Advanced Voice Mode, and image analysis in one subscription. If your workflow uses images, there's no comparison.
- Reliability under deadline pressure. The occasional server timeouts are manageable when you're experimenting. When you need a response in the next five minutes for a client call, they become a real problem.
- Any task involving sensitive or proprietary information. This is a personal line, and it's different for everyone. But I drew mine clearly during week three and didn't cross it.
- Topics with political sensitivity. For journalism, research, or analysis touching on Chinese government or geopolitics — DeepSeek is the wrong tool.
The Real Benchmark: My 30-Day Scorecard
| Task Type | DeepSeek V4 | ChatGPT (GPT-5.5) | Winner |
|---|---|---|---|
| Coding & Debugging | Excellent — 80.6% SWE-bench | Excellent — 58.6% SWE-bench Pro | 🏆 DeepSeek (+ free) |
| Math & Reasoning | 92% benchmark score | 88% benchmark score | 🏆 DeepSeek (+ free) |
| Writing Quality | Competent, slightly mechanical | More natural, better tone | 🏆 ChatGPT |
| Long Document Analysis | 1M token context — excellent | ~1M token — excellent | 🤝 Tie |
| Multimodal (Images/Voice) | Text only | Images, voice, DALL-E 3 | 🏆 ChatGPT |
| Reliability / Uptime | Occasional timeouts at peak hours | Rock-solid (Azure infra) | 🏆 ChatGPT |
| Workflow Integrations | None | Zapier, Notion, Slack, 100+ | 🏆 ChatGPT |
| Privacy / Data Security | Data stored in China, gov access risk | US-based, opt-out training data | 🏆 ChatGPT |
| Political Sensitivity | Censored topics — built into weights | No political censorship | 🏆 ChatGPT |
| Cost | Free / 100x+ cheaper API | $20/month / expensive API | 🏆 DeepSeek |
| Open Source | ✅ MIT license, self-hostable | ❌ Closed source | 🏆 DeepSeek |
Did I Go Back to ChatGPT?
Yes. But not entirely.
On day 31, I reactivated ChatGPT Plus. But I kept DeepSeek installed and running alongside it. Here's the workflow I landed on after the experiment:
DeepSeek V4-Pro for: coding tasks that don't involve proprietary code, math and quantitative problems, long document analysis of public information, any high-volume API work where cost is a genuine constraint.
ChatGPT for: writing that will go in front of clients or readers, anything involving images or voice, tasks requiring third-party integrations, sensitive or proprietary information, anything related to current events or politically sensitive topics.
The $20/month I spend on ChatGPT Plus now feels more justified, not less — because I know exactly what I'm paying for. I'm paying for the writing quality, the reliability, the multimodal capability, the integrations, and the privacy architecture of a US-based company operating under US law. Those things have a real value. They're just not the only value in AI tools right now.
Who Should Actually Switch to DeepSeek?
Strong case for switching (or using both):
- Developers and engineers doing coding work that doesn't involve proprietary code
- Students doing math, science, or technical coursework
- Researchers working with public information and long documents
- Anyone building AI-powered products who needs API access at scale without enterprise budgets
- Anyone currently paying $20/month for ChatGPT but primarily using it for coding or math
Strong case against switching (or for serious caution):
- Anyone who handles sensitive client data, legal information, or medical records in their AI workflow
- Business owners who use AI for strategy, competitive analysis, or internal planning
- Writers and content creators where natural prose quality is the primary output
- Journalists and researchers who need to explore politically sensitive topics
- Anyone working in defense, government, healthcare, or finance where data jurisdiction matters
- Anyone who needs reliable uptime under deadline pressure
The Honest Bottom Line After 30 Days
DeepSeek V4 is the most interesting AI development of 2026. Not because it's the best model — it isn't, on most dimensions I care about for my actual work. But because it proves something important: frontier-class AI capability can be built and deployed at a fraction of what American labs charge for it. That's going to force prices down across the entire industry, which benefits everyone.
The privacy and censorship concerns are real. Don't let anyone tell you they're not. But they're also not disqualifying for every use case. They're a constraint you need to understand before deciding what to use the tool for.
If I had to give you one sentence: DeepSeek V4 is the best free AI for coding and math. ChatGPT is the better daily driver for most people. Claude is the better writer. And if you care about where your data lives — think carefully before you type your most sensitive work into any of them, free or paid.
The $20/month I was going to save? I'm spending it. But I'm using DeepSeek too — just for the right tasks.
Frequently Asked Questions
Is DeepSeek really free?
Yes. DeepSeek V4-Pro and V4-Flash are completely free to use in the chat app at chat.deepseek.com, with no subscription tiers and no meaningful query limits. The API has very low costs ($0.14–$0.44 per million input tokens), but the consumer chat interface is entirely free.
Is DeepSeek safe to use?
For casual, non-sensitive tasks — yes, with awareness. DeepSeek collects extensive user data (chat history, device data, keystroke patterns) stored on servers in China, subject to Chinese government data access laws. It's been banned from US government agencies, NASA, the Pentagon, and multiple countries' government devices. For sensitive business, legal, medical, or proprietary information — the risk is real and worth taking seriously.
Is DeepSeek better than ChatGPT?
On coding benchmarks and math, DeepSeek V4-Pro is competitive with or ahead of GPT-5.5. On writing quality, multimodal tasks, reliability, integrations, and privacy architecture — ChatGPT is better. There's no single winner; the right choice depends entirely on what you're using AI for.
Why is DeepSeek so much cheaper than ChatGPT?
DeepSeek uses a mixture-of-experts architecture that activates only about 49 billion of its 1.6 trillion parameters per token — making inference significantly more efficient. It was also trained on lower-cost hardware (Huawei Ascend chips, partly avoiding Nvidia GPU costs due to US export restrictions). The open-source MIT license also removes the need to recoup closed-model development costs through pricing.
Does DeepSeek censor topics?
Yes. DeepSeek refuses or deflects on politically sensitive topics related to the Chinese government — including Tiananmen Square, Taiwan's independence, Tibet, and Xinjiang. This censorship is built into the model weights through supervised fine-tuning, not just a keyword filter. It cannot be prompted around in the standard chat interface.
Can DeepSeek generate images?
No. DeepSeek V4 is text-only in its current form. It has no image generation, no image analysis in the chat interface, and no voice mode. If these capabilities matter to your workflow, ChatGPT (with DALL-E 3 and Advanced Voice Mode) or Gemini (with Omni) are the better choices.
Should I cancel ChatGPT Plus and switch to DeepSeek?
Probably not entirely. If your primary ChatGPT use is coding or math — DeepSeek is a legitimate free alternative worth trying. If you use ChatGPT for writing, images, voice, integrations, or sensitive work — keep the subscription. The smartest approach for most users is to use both, routing tasks based on what each does best.
Last updated: June 14, 2026. DeepSeek V4 is currently in Preview; behavior and pricing may change at general availability.
0 Comments