Comment by jodacola

Deepseek cost/performance is incredible. That said, I still feel like for agentic coding we haven't plateaued (I slightly prefer GPT 5.5 to Claude for complex stuff, to be honest), and so the extra price is absolutely worth it to push you over the 'impossible' to 'feasible' bar on complex tasks. Once you're in a domain that Deepseek can handle though that requires volume, I would almost always default to it now.

For evals in particular (tuning workflows that agents are using), effectively not having to worry about price is an incredible multiplier - getting statistical significant signal is not cheap otherwise.

Depends on what you mean by 'bang for buck'. The open weights aren't better than openai/claude. But they are much cheaper and the limits are much higher, so you get more work out of it for less money. Every subscription provider out there provides better money-per-limit value than Anthropic (other than GitHub, who are by far the most embarrassingly overpriced and limited provider). (https://codeberg.org/mutablecc/calculate-ai-cost/src/branch/...)

> I’m still apprehensive about shipping all my stuff off to a lab under an adversarial government (to the US)

Do you mean you don't want to use the models created by a non-US lab? In that case, yes you're stuck with US models, but there's a half dozen big labs in the US. If you meant just where your inference is done, there are providers in 12 different countries through OpenRouter, including the US. Several subscription providers host in multiple countries. There's a lot of choices.

My advice -- give it a try. Chuck $5 into deepseek.com , and use this config (put it in a shell script, run ' . ./deepseek-claude.sh ', then just run claude as normal.

    export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
    export ANTHROPIC_AUTH_TOKEN= *** PUT YOUR DEEPSEEK KEY HERE ***
    export ANTHROPIC_MODEL=deepseek-v4-pro
    export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro
    export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro
    export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
    export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
    export CLAUDE_CODE_EFFORT_LEVEL=max

I started by using it for some bigger reading jobs, particularly when I was near limit. Honestly, it's not quite as good, but it's much cheaper, and means I can carry on working. I also find sometimes it's good to ask claude and deepseek to consider code, how to polish, it see what they both say.

I’m using Claude with a $100/month subscription. I’m playing around with using Opus as the Architect, Sonnet as the implementer/engineer and Deepseek-pro as the deep reviewer, and tester. It’s been quite good as I expected. If my usage pattern holds up, I would downgrade my subscription to the $20/month one and toss more money to Deepseek.

Repo reference here: https://github.com/aravindhsampath/agentic-template

Much more bang per dollar, yes. Somewhat less bang per hour.

As usual, different models get stuck on different things. I run DeepSeek v4 API for most of my Cursor experimentation / poking around / proof of concept stuff, but I trust it less than OpenAI/Claude for writing production code. Sometimes DeepSeek is great for debugging, planning, etc. Sometimes it gets stuck or outputs low quality. That's true of OpenAI and Anthropic models as well though.

Overall, DeepSeek seems serviceable but a rung below Opus 4.8 and GPT 5.5. I run them all on maximum thinking settings.

I've found myself liking opencode for workflows because i can plug GPT models into it, so i tossed 5$ at deepseek api and just toggle back and forth what my opencode.jsonc file is running model wise for my agents. I havent tried anything crazy yet with it, but its nailed all the tasks i felt were overall too simple to waste gpt usage on.

Hardest stuff i threw at it... i did like a set of 3 each for claude/gpt/ds, it was all pretty steady across all providers. I think claude won but it could have just been it rng'd into the 3 easier tasks, they are all similar tasks but not identical, these aren't like benchmark tasks just a steady flow of annoying html/json/regex type stuff. Almost always they need a second pass regardless of what model i throw at it, just to tighten up some loose ends, and it fit right into what my current expectation was of gpt 5.5 and opus 4.6.

I'd recommend carefully looking at a few benchmarks (even though generally relying on benchmarks is problematic)

https://artificialanalysis.ai/evaluations/omniscience

Esp check the Hallucination rate for Deepseek - it's not good.

I used ~16,000,000 input tokens yesterday on v4 pro, ~15,000,000 were cache hits, and I spent $0.47. Output tokens were negligible. However that's with Zed's harness, I'm not sure what you would get with Claude Code.

It's maybe not quite as knowledgeable as the most expensive American models and maybe makes more mistakes (just a feeling based off of vibes, don't take my word for it), so you need to constrain its scope more. That suits my workflow, half the time I have it generate code in the chat window and then write it myself, and I'm mostly using it at the level of generating function bodies and stuff, not entire features. Although it is writing a lot of SwiftUI without me really knowing the language and doing a fine job as far as I can tell (which isn't much admittedly).

One benefit I don't see talked about is it's speed - it's really quick, doesn't spend too much time reasoning even on "max", and the flash model is pretty dang good too. This lets me get into "flow state" when I'm writing code, compared to my experiences with Codex and Opus which would take minutes to complete even basic tasks and kind of ruined my focus.

It's so cheap though, you could download a different harness (Crush, OpenCode, Pi etc) and load $5 in credits and test it for yourself.

also curious. On the claude code $200 plan, get close to weekly limits but don't usually hit it. to me just about any small reduction in performance would not be acceptable, the cost of redirecting and getting stuck during long runs without me are too big (like when I tried gemini cli for a few days).

if it's 99.9% comparable performance for less money I'm interested, but I'm skeptical it's there

Yeah, the discounted deepseek inference is subsidized by the CCP for a reason, and it's one that might well come back to bite.