DeepSeek just launched web search and fetch capabilities in its API, and the pricing is hard to ignore: $0.75 per 1000 successful requests.

That’s 0.075 cents per search. For context, Perplexity charges $5 per 1000 searches on its API. OpenAI doesn’t publish search pricing separately, but bundles it into usage on models like GPT-4o with search. Google’s grounding with web search runs about $35 per 1000 queries through Vertex AI.

DeepSeek’s implementation is straightforward. You add "search": true to your API request, and the model can query the web and fetch content from URLs during generation. The documentation shows it works with their V3 model. Successful requests get billed at the flat rate. Failed requests don’t count.

This fits DeepSeek’s pattern. The lab has consistently priced below Western competitors while publishing strong benchmarks. Their V3 model launched in January at $0.27 per million input tokens and $1.10 per million output tokens, roughly one-tenth of GPT-4’s pricing at the time.

Why this matters

Search-augmented generation has moved from research novelty to table stakes. Developers want models that can pull fresh data without manual RAG pipelines. But pricing has kept it out of reach for high-volume applications.

At $0.75 per thousand, you can afford to add search liberally. A customer support bot that searches documentation on every query costs $7.50 per 10,000 conversations. A research tool that runs 100,000 searches a month costs $75.

That changes what you can build. Suddenly search-augmented responses aren’t a premium feature you gate behind higher tiers. They’re baseline functionality.

The catch

DeepSeek is a Chinese company. That matters for some use cases and doesn’t for others. If you’re building internal tools or consumer apps where data locality isn’t regulated, it’s probably fine. If you’re handling HIPAA data or working with the federal government, you already know this isn’t an option.

The model quality is competitive but not cutting-edge. DeepSeek V3 performs well on benchmarks, but it’s not beating GPT-4o or Claude 3.5 Sonnet on complex reasoning tasks. For many search applications, that’s not the point. You need something good enough to parse query intent and synthesize results, not solve IMO problems.

Response times matter too. DeepSeek’s API runs on Chinese infrastructure. Latency from US or European servers will be higher than domestic providers. If you’re building a chatbot where every 100ms counts, test it first.

What else shipped this week

Simon Willison released web-edit 0.1a2, a minimal web app for browsing and editing files remotely. He built it to edit files on different machines via Tailscale and added clipboard support for text and images. It’s open source and deliberately tiny. If you’ve ever SSH’d into a box just to edit one config file, this is worth looking at.

Replit switched from Anthropic to OpenAI as its default model provider, according to CEO Amjad Masad. Claude will still be available on the platform, but new users will get OpenAI models by default. Masad didn’t elaborate on why, but timing suggests it’s related to OpenAI’s recently announced developer partnerships and pricing for high-volume customers. For Replit users, it mostly means different default behavior in the AI chat. You can still pick your model.

And someone on Hacker News built Sphinx, a hook that quizzes you on AI-generated code before you ship it. The idea: if you can’t explain what the code does, you probably shouldn’t merge it. It’s a good forcing function for the copy-paste problem, though whether developers will actually use a tool that makes shipping harder remains to be seen.