- TechNode describes a cost chart in which DeepSeek's V4.1-Flash-1017 solves 77% of DeepSWE v1.1 tasks at about $0.20 each. The figures are mostly DeepSeek's own.
- The publication's argument: track the cost of one finished task on your own work, including retries and tool calls, not price per token.
- Search and Fetch calls through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing begins. TechNode lists direct API pricing at 5 yuan per 1,000 successful requests.
The number to track is the cost of a finished task
Most AI pricing talks about tokens, the small chunks of text a model reads and writes. An agent does not really buy tokens. It buys an outcome: a bug fixed, a report filed, a page checked.
The lesson here is our own argument, not DeepSeek's. When you fund AI agents, track the cost of one finished task, including retries and tool calls. Price per token is only an input. DeepSeek's October releases make a handy worked example.
What DeepSeek released
DeepSeek shipped V4.1-Flash-1017 on October 17, a refresh of its V4.1 Flash model. Its announcement calls the result a "Pareto improvement in performance and cost". Put simply, the company says the model posts stronger benchmark results, completes tasks faster and writes shorter outputs than its predecessor.
TechNode, which covered the release, lists gains in DeepSeek's published results. On Terminal-Bench 4.0, the score climbed from 31.2 to 42.3. On DeepSWE v1.1, it moved from 74.2 to 77.0.
TechNode also describes a cost-performance chart for DeepSWE v1.1. In it, the model solves 77% of tasks on the first attempt, at roughly $0.20 a task. That measure is called Pass@1. TechNode says several models that cost substantially more to run scored lower. At cheaper settings of $0.11 to $0.14 per task, the model still lands around 72% to 73%.
How much of this to believe
TechNode says these figures are mostly DeepSeek's own benchmark results. It adds that different tests use different frameworks and settings. The results do not show the model will work equally well in every application.
TechNode also cites a separate signal: ITHome's October 20 report on Arena's Agent Arena leaderboard. Across 7,498 real-use sessions, the model placed 11th overall and first among open-weight models. Its median cost was about $0.18 per task. Leaderboard pages showed figures from $0.18 to $0.19.
That is a different measure from the DeepSWE chart. One covers real-use sessions and the other covers benchmark tasks. The similar dollar figures are suggestive only.
Where an agent's bill comes from
An agent works in loops. It reads, calls a tool, reads the answer and writes again. Each loop uses tokens and computing time. TechNode says a model that does the same work with fewer tokens and less computing time can cut inference costs. That makes shorter outputs and faster completion cost claims, not style claims.
On October 20, DeepSeek also released DeepInfer. DeepSeek describes it as the framework behind its large-scale production services. An inference framework decides how a model uses computing resources to answer requests. TechNode says that design affects latency, throughput and the hardware needed to serve users, all of which shape serving cost.
The release has limits. It runs on Nvidia SM100 chips and Huawei Ascend 950 chips. Only one model, V4.1 Flash, is listed as supported. DeepSeek's documentation names three parts of a full service that the release covers only partly: a KV-cache system, separating the prefill and decode stages, and balancing load across copies of the model. The KV cache is the working memory a model keeps for the text so far.
TechNode adds some background from V4.1 Flash's September launch. DeepSeek then reported cutting KV-cache memory needs to a quarter and SSD storage needs to an eighth of the prior generation. TechNode also notes that DeepSeek charges half as much for API use at off-peak times. In TechNode's words, the benefits of DeepInfer will depend on hardware, workloads and extra engineering.
The free tools are about to be metered
DeepSeek says its Search and Fetch APIs are now open to everyone. It says they run on its own search engine, which handles hundreds of millions of requests a day worldwide. These are the calls an agent makes to look things up and read web pages.
DeepSeek's announcement says calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are not billed at present. Billing starts October 31, 2026, at 16:00 UTC. TechNode reports each API is priced at 5 yuan per 1,000 successful requests. It gives November 1 as the start of per-request billing. The sources give two dates, so confirm which one applies to your account.
So an agent task has two cost parts: the model and the tools it calls. A pilot that looks cheap while lookups carry no charge may look different once they are metered.
What to ask your team
First, ask for the cost of a completed task on your own work, with retries and tool calls included. A vendor chart on a public benchmark cannot answer that.
Second, ask which agent workflows call search or fetch tools, and how many calls a typical task makes. Then price those calls at the rates that apply after the free period.
Third, if someone proposes self-hosting, ask what is missing. DeepInfer lists one model and two chip families, and it leaves parts of a full service uncovered. Who will build those parts, and at what cost?
A cheaper model is only good news once it shows up on your invoice. The number worth tracking is the finished task.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: DeepSeek.
