DeepSeek V4.1 Flash 1017: cheaper coding model, open weights and a search API at $0.75

DeepSeek, the Chinese AI company, released DeepSeek-V4.1-Flash-1017 on Saturday 17 October, an update to its V4.1 Flash model that it says is faster, gives shorter answers and costs far less per task. "deepseek v4" was among the rising searches in Pakistan this week. The model is on DeepSeek's website, app and API, and the weights are published on Hugging Face and ModelScope, so developers can run it on their own hardware.

The figures below are DeepSeek's own, as reported by The Decoder and Bloomberg. Independent tests were not yet out.

The numbers

• DeepSWE coding benchmark: 77.0% at its top setting, at an average of just under $0.20 a task. The earlier V4.1 Flash scored 74.2% at about $0.50 a task.
• Low setting: above 72% at roughly $0.10 a task.
• Rivals: on DeepSeek's cost chart, the larger models from OpenAI, Google and Anthropic cost between $2 and more than $10 a task. DeepSeek says its model beats several of them in coding and agent categories.

At Monday's interbank rate of about Rs 277 to the dollar, 20 cents is roughly Rs 55 a task. Check the live rate on our USD to PKR page or convert any amount with the currency converter.

How it was trained

DeepSeek says it used reinforcement learning with self-feedback (RLSF): the model is rewarded using its own confidence scores and its own critiques of its answers, rather than human raters or a separate reward model. That is one reason the company can update quickly and cheaply.

A search engine for developers

DeepSeek is also opening its internal search engine through two paid APIs, Search and Fetch, at $0.75 per 1,000 calls each. Use is free through DeepSeek Harness, Claude Code and Codex until 31 October (9pm PKT). The company says the engine already handles hundreds of millions of queries a day.

What it means for developers in Pakistan

For software houses and freelancers in Lahore, Karachi and Islamabad who bill clients in dollars, cheaper models lower the cost of building coding assistants and chatbots. Open weights also mean a model can be hosted locally, which matters for clients who do not want data leaving the country. Freelancers paying for API credits should remember the tax side of foreign income; our freelancer tax calculator shows what applies.

One caution: The Decoder notes reports that China may tighten rules on publishing the most capable open-weight models, which could affect future releases.