Issue #23: There's an AI model that costs 36x less than ChatGPT. Should you switch?

V4 Flash costs $0.14 per million input tokens. GPT-5.6 Sol costs $5.00. That's a 36x difference. And on most everyday tasks, the quality gap is smaller than you'd expect.

DeepSeek V4 Flash — the latest model from Chinese AI lab DeepSeek — has become the price floor of the AI model market. It launched in 2026 as part of a broader V4 family that's been systematically cutting costs and matching Western benchmarks. And it's creating a genuine decision for small businesses that use AI at any meaningful volume: should you switch, partially or fully, to save money?

This issue gives you the honest answer — not the hype version, not the fear version. Just the comparison that helps you decide.

First, let's understand what DeepSeek actually is. DeepSeek is a Chinese AI research lab that has consistently released powerful models at prices that dramatically undercut OpenAI, Anthropic, and Google. V4 Flash is their fastest, cheapest model. V4 Pro is their flagship reasoning model. Both are available via API and through the free web chat at chat.deepseek.com.

Here's the honest benchmark breakdown:

DeepSeek V4 Pro tops LiveCodeBench at 93.5 and sits at 80.6 on SWE-bench Verified — essentially tied with Claude Opus 4.7 at 80.8, at roughly 1/34th the input cost. On the Artificial Analysis Intelligence Index, V4 Pro and GPT-5.6 Sol score 58.9 each. Effectively a tie on general capability. Not a tie on cost.

DeepSeek V4 Flash is cheaper still — $0.14/$0.28 per million tokens versus GPT-5.6 Sol at $5/$30. On everyday tasks — writing, summarising, drafting — V4 Flash performs competitively with mid-tier Western models at a fraction of the price.

The honest comparison for small businesses:

At $0.14 input and $0.28 output, V4 Flash undercuts even OpenAI's cheapest nano tier on output. At production scale — say, 10 million tokens per day — V4 Flash costs roughly $51/month. The same volume on a comparable US frontier model runs into the hundreds.

The verdict: DeepSeek V4 Flash is genuinely impressive for its price point. For high-volume, text-heavy tasks where cost matters and quality can be slightly variable — content drafts, summarisation, data processing — it deserves serious consideration. For tasks requiring the highest reliability, nuanced judgment, or sensitive business data — the Western models still have the edge.

A content marketing agency producing 40+ blog posts per month for clients tested DeepSeek V4 Flash against their existing Claude Sonnet setup for first-draft generation. The brief: produce a 600-word first draft from a keyword and topic brief.

Results after 30 posts tested on each model: output quality was comparable on straightforward informational content. V4 Flash produced stronger output on technical topics. Claude Sonnet produced more natural, brand-voice-aware writing on lifestyle and consumer content.

They implemented a split workflow: V4 Flash for technical first drafts (roughly 60% of their volume), Claude Sonnet for consumer and lifestyle content. Monthly API costs dropped from $340 to $89. Output quality held. The model change was invisible to clients.

"Is it safe to use a Chinese AI model for my business?"

This is the right question and it deserves a straight answer. There are three legitimate concerns: data privacy (your prompts are processed on DeepSeek's servers, which are subject to Chinese law), censorship (DeepSeek will refuse certain politically sensitive topics, which matters for some use cases and not others), and geopolitical risk (if US-China tech relations deteriorate further, access to DeepSeek could be restricted). For most small business owners doing content drafts, summarisation, and data processing — none of these is a dealbreaker. If your work involves sensitive client data, legal or financial information, or anything that requires strict data residency — stick with Western models or access DeepSeek through AWS/Azure infrastructure. If your content occasionally touches on politically sensitive topics — test it first, because the model's refusals may frustrate you. For everything else: it's a capable tool at a significantly lower price, and the risk profile is manageable with sensible practices.

Important note for most readers: if you're using ChatGPT or Claude through a $20/month subscription — not the API — switching to DeepSeek won't save you money unless you also switch to their API. The savings are in API token costs, not subscription pricing. For small businesses not using AI at high volume via API, the cost difference is less significant than it sounds.

This week: go to chat.deepseek.com and run your five most common AI prompts. Same inputs you'd give ChatGPT. Compare the outputs. This takes 20 minutes and costs nothing. You'll know immediately whether the quality meets your standard for the tasks you're testing — and whether the switch is worth exploring further.

DeepSeek's pricing move is part of a broader pattern that's accelerating: every time a Western lab raises model capability, a Chinese competitor closes the gap at a fraction of the cost. V4 Pro now matches Claude Opus 4.7 on the SWE-bench coding benchmark at 1/34th the input price. The gap between "frontier Western model" and "frontier Chinese model" has narrowed to a point where price is now the more meaningful differentiator for most commercial use cases. OpenAI responded with aggressive pricing cuts on GPT-5.6 Terra specifically — which is now priced lower than before — as direct competitive pressure from DeepSeek pushes the whole market toward cheaper AI.

Switching your entire AI workflow to DeepSeek overnight based on the pricing headline. The right approach is selective routing: identify which of your current AI tasks are high-volume, low-risk, and output-consistent, and test DeepSeek specifically on those. Keep your primary model for sensitive, complex, or client-facing work until you've built genuine confidence in the alternative. Chasing the cheapest model everywhere is the AI equivalent of buying the cheapest tools for every job — some of the time it's fine, some of the time it costs you more than you saved.

Until next issue, Kris

The Layman's AI — The only AI updates your business actually needs.