OpenAI raised prices - DeepSeek cut them by 5x

TL;DR: on OpenRouter, V4 Pro costs $1.74 / $3.48 per 1M input/output tokens - roughly 5x cheaper than Opus 4.6 ($5 / $25), while performing close to it on key benchmarks. For the first time, there's an open-weight model genuinely in Opus's league for agent coding, and it ships just as OpenAI raised GPT-5.5 pricing 2x for marginal gains.

DeepSeek just released DeepSeek-V4-Flash (284B parameters) and DeepSeek-V4-Pro (1.6T parameters).

Both models are open-weight, available to download on Hugging Face - Pro is here and Flash is here.

Why might this be a big deal? Several reasons.

First of all, it looks like we finally have an open model with capabilities close enough to Opus 4.6.

How does it compare to Opus 4.6

As you see, it's somewhat on par. Definitely better than any other open model out there.

Why am I comparing to Opus 4.6?

Because Opus 4.6 was the first model where agent coding clicked for a lot of people. That alone makes V4 Pro the first open-weight model genuinely suitable for agent coding.

What is the punch line?

Of course, it's pricing. On OpenRouter, DeepSeek V4 Pro costs $1.74 / $3.48 per 1M input/output tokens - roughly 5x cheaper than Opus 4.6 ($5 / $25).

Here's the bigger picture with some related models, sorted by output price:

DeepSeek V3.2             ($0.26 / $0.40)
xAI: Grok 4.1 Fast        ($0.20 / $0.50)
MiniMax M2.7              ($0.3  / $1.2)
Gemini 3 Flash Preview    ($0.50 / $3)
Xiaomi MiMo-V2-Pro        ($1    / $3)
DeepSeek V4 Pro           ($1.74 / $3.48)
Kimi K2.6                 ($0.8  / $3.5)
Sonnet 4.6                ($3    / $15)
Opus 4.6                  ($5    / $25)
Opus 4.7                  ($5    / $25)

Personally I'm fine with an Opus 4.6-level model that costs just a fraction of it. While companies can continue chasing marginal improvements, economically speaking the one who does the job well enough and cheaply enough wins.

Bigger context

Both Anthropic and OpenAI are fighting for users and at the same time frantically trying to optimize the costs.

Just a couple of days ago Anthropic was caught A/B testing removing Claude Code from the Pro plan, which obviously raised a backlash among users.

On the other hand, GPT-5.5 just shipped with marginal improvements at a steep price bump ($5 / $30) compared to $2.5 / $15 for GPT-5.4. That's a 2x increase.

Developers are divided. Some of us jumped from OpenAI to Anthropic and back multiple times.

Not sure how long the DeepSeek guys sat on the new model, but their timing was perfect.

How to try it out?

Theoretically Claude Code can be used with any model, but I've heard their prompts are really focused on their own models, so it's better to use a model-agnostic harness, for example PI or OpenCode.

As the model was just released, there are still open questions.

Will it perform well enough for daily coding tasks in the real world?

How are the bigger players like Anthropic, Google, and OpenAI going to react?

What will investors say? Will it affect the market and maybe even lead to bursting the AI bubble?

All of this remains to be seen.

#article #ai #public