DeepSeek has cut the price of 1 million tokens to $0.15

DeepSeek has rolled out DeepSeek-V4.1-Flash-1017, a build aimed at coding and terminal agents. The new API pricing the company announced noticeably cuts the item developers working with autonomous coding tools have complained about most in recent months: the token bill. During off-peak hours, the cost of 1 million input tokens drops to $0.15, and output tokens to $0.60.

What changed in the pricing

Western frontier models typically price 1 million input tokens between $2.50 and $3.00, while output tokens fall in the $10-15 range. It is normal for an autonomous agent that continuously reads dozens of files, inspects terminal output, and restarts from scratch to fix an error to spend 3-5 million tokens on a single task. That pushed monthly bills into the hundreds of dollars.

Here are the details of the Flash 1017 pricing: during off-peak hours, input is $0.15 and output is $0.60; on a cache hit, input can fall to as low as $0.003 per million; even during the busiest hours, input stays at $0.30 and output at $1.20. The model can also generate up to 384,000 tokens of uninterrupted output in a single response, which lets large refactoring jobs be completed without being split up.

Cheap, but powerful enough?

According to independent test results reported by the source, the model sits well above the lightweight class. In Terminal-Bench 4.0, under the Claude Code harness, the previous Flash build’s score of 31.2 rose to 42.3. On DeepSWE v1.1, a benchmark for solving real GitHub issues, it scored 77.0 under the Mini-SWE-V2 harness. In NL2Repo-Bench, a test of building repositories from scratch from natural language, the success rate was recorded as 68.2. The tendency of earlier versions to produce unnecessarily long code cluttered with comments has also been trimmed; the model delivers shorter blocks that just work.

Plugs into existing tools with one line

DeepSeek’s most practical move was standard compatibility. Developers can use the api.deepseek.com address or the api.deepseek.com/anthropic endpoint that speaks Anthropic format. In tools like Claude Code, Cline, or Cursor Composer, changing the model address is enough. OpenAI’s Python and TypeScript SDKs work by only updating the base_url parameter, without changing a line of code. When long project files are resent on every loop, the server automatically counts this as a cache hit and brings the bill down to the cent level.

What it means for teams in Turkey

Given the dollar exchange rate and subscription fees, the picture matters even more for local developers. According to the source’s calculation, instead of a single $20 monthly subscription, a $5 API balance can run uninterrupted autonomous coding loops for weeks. In token-heavy tasks such as customer support bots, CI/CD code review pipelines, and automatic documentation generation, cutting costs by about 90 percent expands small teams’ trial-and-error budget.

In short, price competition now runs not only on model quality but also on inference cost. DeepSeek’s pricing aims to turn coding agents from an experimental toy into part of the daily workflow.

Source: WebTuring