June 2026 - Deepseek API For Performance & Cost

(following post has been updated 5th of july, 2026)

Intro

A shift in LLM/AI usage: GitHub Copilot changed its user pricing. Microsoft changed the pricing. Nvidia & AMD are coming up with new laptops with 128 or more GB of VRAM to run agents and LLMs locally! But the prices of the chips are not going down; at the time of writing, the Nvidia RTX 5070 16GB is 65k INR.

The company that I work for at the time of writing this broke the Microsoft GitHub Copilot tie-up and looked for other vendors. They did it so fast: they sent a company-wide email stating it would be terminated immediately for the FTE (own employees) developers but will be kept only for the vendors, and asked users to utilize and adopt Claude Code. So we started using that, but a lot of people just use it for the sake of it to finish the job. At this point, there is no auto mode in the Claude version they allow inside the organization; people just set the model to high and burn through the tokens. It’s been 2 days and people went ham on it, and already 25% of the team budget is gone. We can check out individual budgets and the team’s budget at the usage metrics.

For me, I got around 300 dollars of AI credits at my disposal: $100 for Claude and around 200 for Cursor. To think that is around 28k INR—the salary I got at the time of starting my career when I left college. Back then, was it considered good money? Things changed really fast and got pricey, huh? These token budgets are arbitrary values set by companies on their own terms and bullshit logic. We are paying money and allocating budget for thin air—tokens? A decade ago, people would laugh at the idea! That would be 2016. I remember 2016; it was a different time: peaceful, chaotic, it had its charms.

Deepseek api For The Win

I’ve been using the DeepSeek API directly for the entire month of June 2026. It’s been great; they offer cheap prices for the latest model they have (at the point of writing this in June/July 2026 (this line was edited on July 5th, 2026, 1:08 AM)). Here is a comparison of the API pricing per 1 million tokens (as of July 2026):

*Note: Claude Sonnet 5 features introductory pricing of $2.00/$10.00 per million tokens through August 31, 2026. For verification, check the official docs: DeepSeek Pricing, Google AI Studio Pricing, Anthropic Claude Pricing, and OpenAI API Pricing.

Deepseek api for the right reason.

All the frontier AI models at this point are capable of doing the job they are told to do. There is a saying that you just have to do 80%/90% of the work, and the remaining work I can do myself. A lot of people are feeling that way regarding the AI models. What about them? The AI models have been performing well over the last 6 months; we don’t want the last itty-bitty performance for the day-to-day tasks that we do. It is needed for hard problems, sure, but for daily things it’s definitely not needed. So, for the reason of being cheap, affordable, and open weights… I chose to use DeepSeek.

Purpose

I mainly started to use the deepseek api for the coding, they offer the model to be compatible with claude code and vscode (github copilot), they have instructions to integrate with every tool you can think of — see the DeepSeek API Documentation for details. Specifically, for VS Code you can use extensions like Continue or Cline configured to point to the base URL https://api.deepseek.com, or install the DeepSeek V4 for Copilot Chat extension. For Claude Code, you can use the Anthropic-compatible endpoint by exporting environment variables before starting the CLI:

export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="<your_deepseek_api_key>"
export ANTHROPIC_MODEL="deepseek-v4-pro[1m]"
export ANTHROPIC_DEFAULT_OPUS_MODEL="deepseek-v4-pro[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="deepseek-v4-pro[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-v4-flash"
export CLAUDE_CODE_SUBAGENT_MODEL="deepseek-v4-flash"

No matter how much I use, they didn’t utilize the money I topped up on their site. It was just $5, but the usage I was doing was hardly making a dent. Given that I was using the pro model for the tasks and the flash model for menial tasks, seeing that it made no difference on the token budget, I started to use the DeepSeek-V4-Pro model for all tasks. Even then… it made no dent in the billing. I still hadn’t utilized that first $5, but after starting to use agents ²⁴⁄₇ (which I talk about later in the post), I topped up another $5 in the middle of the month.

I was mostly using it for coding and simple tasks like automation to do mundane tasks, but just for the one instance, that’s it. Since I was doing coding mostly for experimental stuff and not code that I want to be private, I didn’t care about the Chinese company using my prompts or data. It’s been an agreed-upon truth (for anyone reading from far beyond~ maybe a different era), (everybody in the world agrees upon this one thing) that all the Chinese companies ever do is steal technology and steal data from people around the world. At the point of writing this, maybe things will have changed in the future. There are other service providers that advertise/say they don’t train on the API we use (like OpenCode (I’m going to try that next month, I’ll tell you why in a bit)); they also host the DeepSeek API inside the USA. Since the model is open-weights… they can just download and host them on their own hardware… or buy it from vendors inside the USA and serve it to customers, but internally they could be doing anything we don’t know.

Evolving

As I’ve stated before, since there is no cost involved for most of the time when I’m using the APIs, I started to use AI agents ²⁴⁄₇. Before, I just used them whenever necessary with different tools, but never ran Hermes or OpenClaw for a long time. I started experimenting with them on my laptop, but then I made the setup on my server—configuring 3 messaging apps and adding others so they can work independently. Now, both the Hermes agent and OpenClaw are running ²⁴⁄₇ on my server inside a container without sudo access. At first, it was just intrigue: why did others get obsessed with them? But since the cost is low, why not? They work nicely and actually save time by doing work independently.

Initially, I liked OpenClaw, but then when I used it more and more, I started to use Hermes more. It is important to note that OpenClaw is written in Node and uses Node for everything when doing tasks, but Hermes is written in Python, so it uses Python everywhere when it’s doing tasks.

Since there is no cost, I started to use AI agents ²⁴⁄₇ (Hermes and OpenClaw). Before, I just used them whenever required, so I started experimenting. Instead of just on my laptop, I hosted them on the server, configured 3 messaging apps, and added others. They work nicely and actually save time mostly by doing work independently.

How much AI tokens did I spend this month from DeepSeek directly? (updated after June)

So for the whole month, since I started using it, I’ve used 400M tokens and around $10. I’ve already shared how I’m running Hermes and OpenClaw ²⁴⁄₇ above, and I’ll cover more details about their implementation in my next post.

Personal updates

It’s 3:26 AM on a night. I’m pondering over things, thinking through things. Things are muddy during the day; I’m thinking how can I get this calmness during the day… silence? The comfort of knowing others are asleep, only I am awake, and no one… willingly at this hour contacts anyone. I’ve gotta find that cheat. Is it the ambiance? Or the feeling of knowing there is still 7 hours of time left… until I won’t be bothered by other things.

This month, I’m doing something that I wanted to do for a long time: moving out of Chennai and back to my hometown with all my things and stuff. I’ve been thinking… “what if” this and that for a long time… but the thing is… it’s good to be prepared. I’ve realized I can just do it again when it’s required; it’s not a big deal. Sure, it can be based on luck, but I’m sure I have talented people and smart SAs to sort it out. It’s been a long time… I have an attachment to this place now—my messy table, AC at 21 degrees C on a July night, lights on, curtains half open, fan on top of my head at 55% with rainy weather outside in Chennai and a cold night. Remember this: no glasses, closet behind, half asleep but yet wide awake, thoughts of wanting to sleep so I won’t be groggy for work next morning, reflecting, searing memory, what-ifs. Sorry, all over the place. Swear I’m not drunk.

Still double-minded. Maybe I should keep a small place for myself alone? On the fence on this. All the actions I take based on this… are fruitless. I have the weekend to disassemble things; I think it would be a different set of tasks than I usually do. Gosh, to think I was doing a lot of different things 2 and 4 years ago on the weekends than what I am doing currently.

What happened to EV S, H, M? What happened to those rules of L? Can’t control how and what you are already perceived as. Best to get rid of it, stepping stone, learn?

From me talking earlier about GPUs, I’m thinking of building a machine just for running desktop-level LLMs. For most not-so-significant knowledge work, we don’t need a state-of-the-art model with advanced reasoning. The things people require are context-aware agents/AI that follow the instructions and act accordingly… that’s all there is to it, at least at the current state of LLMs now. All the open-weights models have gotten a lot better over the months; we can host the flash version of those models easily and we are good. But I think I would need a desktop-grade/level GPU… multiple GPUs, to have that sort of resource available to me. In the future, there are probably cheaper ways to run LLMs, or even better, LLMs won’t require GPUs. There are already TPUs from Google and NPUs… but still the same as a GPU in the future… or better yet, there are no LLMs at all? AI is completely different.

After so long, I am planning to leave my Chennai apartment. I remember writing about it 2 years ago on a December, now on a June. I already informed them that I would be leaving, and I plan to vacate by the end of this month. Will update on this next month. Long story short, they asked me to buy it, but I said no.