DeepSeek in Claude Code (and the Engine They Gave Away)

Author: Mervin Praison · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: DeepSeek in Claude Code (and the Engine They Gave Away) · Channel: Mervin Praison · 07:54 · Watch on YouTube

Topics: AI agents, Claude Code, Model inference and optimization, AI models, Local and open-source models

Summary

DeepSeek released an updated Flash model under the MIT license, followed three days later by DeepInfra, the Apache 2.0 serving engine used for its production services. The video covers the model’s benchmark improvements, the engine’s capabilities and hardware requirements, and how to connect the model to Claude Code through DeepSeek’s Anthropic-compatible API. A sample workload costs about $2.10 at off-peak prices, compared with $15 for GPT 6.1 Terra and $30 for Gemini 4 Argon, before caching. The speaker’s main takeaway is that the model is a capable, inexpensive workhorse, while the open-source inference engine is the more significant release.

Key points

DeepSeek V4.1 Flash 1017 has a 1-million-token context window, and its weights are available under the MIT license.
DeepInfra, released under Apache 2.0, is described as the framework powering DeepSeek’s large-scale production services.
DeepSeek reports benchmark gains over the previous Flash model, including Terminal Bench 4.0 rising from 31.2 to 42.3; cross-lab comparisons should be treated cautiously because benchmarks use different harnesses.
DeepInfra includes DeepSeek kernel libraries, expert parallelism with load balancing, speculative decoding, asynchronous batch scheduling, structured outputs, and deterministic sampling on Nvidia hardware.
The engine targets Nvidia SM 100 Blackwell-class GPUs and Huawei Ascend 950; it is intended for large-scale inference, not a MacBook.
DeepInfra supports offline inference but does not include a complete serving system, such as KV cache management, prefill-decode disaggregation, or load balancing across replicas.
Claude Code can use DeepSeek through its Anthropic-compatible endpoint by setting the base URL, API key, and model.
The demonstrated coding tasks cost about 22 cents at off-peak prices; the speaker says to check the dashboard for actual costs.
For a workload of 10 million input tokens and 1 million output tokens, the video estimates costs of about $2.10 for DeepSeek Flash, $15 for GPT 6.1 Terra, and $30 for Gemini 4 Argon at list prices before caching.
The speaker recommends testing on your own repository, considering API data-compliance needs, and using Flash for volume while reserving frontier models for harder tasks.

Chapters

Start	Chapter	Summary
00:00	Two releases and why the engine matters	The video introduces DeepSeek’s updated Flash model and the more unusual release of its production inference engine.
01:03	Benchmarks and what DeepInfra includes	The speaker compares reported benchmark results, cautions against direct cross-lab comparisons, and outlines the engine’s features.
03:10	Hardware limits and Claude Code setup	The video explains DeepInfra’s hardware requirements and missing serving components, then shows how to route Claude Code requests to DeepSeek.
04:42	Coding examples and API costs	The speaker discusses a multi-file refactor, estimates task and token-based costs, and mentions temporarily free search and fetch calls.
06:17	Market context and practical cautions	The video places DeepSeek in an open-model price war and closes with advice on benchmarking, data compliance, self-hosting risks, and model choice.

Quotes

Releasing the serving engine is much rarer, and that’s why developers paid attention. — 01:03

So, you get the engine but not the whole car. — 03:40

At list prices, before caching, DeepSeek V 4.1 flash, off-peak, about $2.10. — 05:13

the model is a solid, cheap workhorse, but the engine is the real story. — 07:21

Concepts and entities

DeepSeek V4.1 Flash 1017, DeepInfra, DeepSeek, Claude Code, Anthropic, Terminal-Bench 4.0, Flash MLA, DSpark, KV cache, Huawei Ascend 950, Codex, Cloudflare, SGLang

Full transcript (original language)

00:00 $30 or $2.10 [music] same day of AI coding. That’s DeepSeek’s new open model, and it’s not even the big news. Three days later, they open-source the engine that runs their own API. That’s rare. Number one AI story on Hacker News this month. >> [music] >> Here’s how to plug it into Claude code and what a session costs. But there’s a catch, the part they didn’t open-source.

00:32 Two releases, three days apart. On the 17th of October, DeepSeek released V4.1 Flash 1017, an update to its Flash model. The weights are on Hugging Face and ModelScope under MIT. The context window is 1 million tokens, and the safe tensors add up to about 763 billion parameters. >> [music] >> It’s already the model behind the DeepSeek Flash name in their API. So, if you were calling Flash last week, you’re

01:03 on the new one now. Then on the 20th of October, they put DeepInfra on GitHub under Apache 2.0. It’s described as the framework that powers DeepSeek’s large-scale production services, and it passed [music] 21,000 stars within days. Labs release weights all the time. Releasing the serving engine is much rarer, and that’s why developers paid attention. Let’s look at DeepSeek’s own numbers against the previous Flash. Terminal Bench 4.0 goes from 31.2 to

01:35 42.3, >> [music] >> an 11-point jump. Program Bench goes from 20.3 to 29.1. Deep SWE goes from 74.2 to 77. Their release notes sums it up as stronger scores, faster task completion, and shorter outputs. And shorter outputs directly lower your bill. Now the honest part. On Terminal Bench, this is not a frontier model. Claude Opus 5.5 is around 66. Kim is new K 3.1 Pro claims 65.7

02:06 and Open AI’s GPT 6.1 Terra, released on the 26th of October, claims about 50. Deep Seek is at 42. Also note that every lab runs these with its own harness and Deep Seek ran terminal bench inside Claude code. So, treat cross lab comparisons as rough. The pitch isn’t best model. The pitch is good enough for most coding tasks at a fraction of the price with weights you can own. Here’s what’s inside Deep Info. It uses

02:39 Deep Seek’s own kernel libraries, Flash MLA, Deep [music] Gem and Deep EP. It has expert parallelism with load balancing for the mixture of experts layers, speculative decoding they call DSpark, asynchronous batch scheduling, and structured outputs with JSON schema, rejects or grammar. On Nvidia, there’s batch invariant >> [music] >> deterministic sampling, so the same prompt gives the same answer. If you’ve ever debugged a flaky agent,

03:10 that last one is a big deal. Two catches. First, hardware. It targets Nvidia’s SM 100 Blackwell class GPUs >> [music] >> and Huawei’s Ascend 950. Their benchmarks used four GPUs for prefill and 64 for decode. This is not something you run on a MacBook. It’s a blueprint for teams running inference at scale and it’s notable that a Chinese lab shipped first class support for a Huawei chip. Second, it’s

03:40 offline inference only. The readme says a full service still needs a KV cache system, prefill decode disaggregation and load balancing across replicas, and those aren’t in this release. >> [music] >> So, you get the engine but not the whole car. Most of you won’t host this yourselves, so here’s the fastest way to try it. DeepSeek exposes an Anthropic compatible endpoint, so you point Claude code at it with three environment variables. The

04:12 base URL is api.deepseek.com/anthropic, then your DeepSeek key, then the model DeepSeek flash. Launch Claude code as normal, and every request now goes to DeepSeek. Here’s a recreation of a typical first task. Add a health endpoint, write a test, and run [music] it. It plans, edits the app, and writes the test. The first run fails, and it fixes the root on its own. When you try it, watch how many turns it takes and whether it recovers from a failing test.

04:42 That’s where the 11-point terminal bench gain should show up. Now a harder one, a multi-file refactor. It finds every import, writes one loader, and fixes a failing test on the second pass. Your own run will differ, so try it on your repo. In this example, both tasks come to about 22 cents at off-peak prices. [music] Your dashboard shows the real figure. One bonus, DeepSeek also opened its search and fetch APIs. They cost 75 cents per thousand calls,

05:13 but when Claude code or Codex calls them as server-side tools, they’re free until Saturday, the 31st of October, at 4:00 p.m. UTC. If you want to try web-grounded agents cheaply, this is the week. Let’s put numbers on it. Take a heavy day of agentic coding, 10 million input tokens, and 1 million output tokens. At list prices, before caching, >> [music] >> DeepSeek V 4.1 flash, off-peak, about

05:45 $2.10. GPT 6.1 tera, at $1 in and $5 out, $15. Gemini 4 Argon, at $2 in and $10 out, $30. Caching helps DeepSeek even more. Cash hit input is a third of a cent per million. One catch for UK viewers, DeepSeek has peak hours on weekdays, >> [music] >> 1:00 to 4:00 a.m. and 6:00 to 10:00 a.m. UK time when prices double. Run your batch jobs in the afternoon. Zoom out.

06:17 And this is a two-week open model price war. Meta shipped MuSpark open. Kimmy promised K 3.1 weights by the 4th of November. MiniMax matched the 0.15 price. And now DeepSeek has shipped the engine, too. Cheap intelligence is turning into a commodity. And the moat is moving to harnesses, tools, and workflows. Four things before you move your team over. One, benchmarks are self-reported and harness dependent. Test on your own

06:48 repo. Two, data. The DeepSeek API runs on DeepSeek servers. If that’s a compliance issue, the MIT weights mean you can use another host or run it yourself. Three, if you self-host, KV cache is where things break. On the 26th of October, Cloudflare published a postmortem and SG Lang bug in disaggregated serving leaked 155 response containing other customers’ data on very long prompts. They fixed it

07:21 and they recommend SG Lang 0.5.22 or later. That’s exactly the layer Deep Infra leaves out. Four, it’s a flash model. For your harness refactors, Frontier model still leads. Use this for volume and the expensive models for the hard parts. My take, the model is a solid, cheap workhorse, [music] but the engine is the real story. DeepSeek is publishing how a Frontier lab actually serves models. And every

07:54 inference team will read this code. If you build agents, try DeepSeek flash and Claude code this week while search and fetch are free and measure it against what do pay now. I’ve put the exact commands and every source in the description. Tell me in the comments what it cost you per task. Follow for more AI developer breakdowns like this one.