First look – Deep Cogito: Cogito V2 Preview Deepseek 671B

Deep Cogito V2 Preview Deepseek 671B is an interesting LLM simply because it is as expensive as GPT5-Full for input token pricing

Created Sep 2, 2025
163,840 context
$1.25/M input tokens
$1.25/M output tokens

Current prices for GPT-5 Full:

Created Aug 7, 2025
400,000 context
$1.25/M input tokens
$10/M output tokens

However, when I tried to get it to extract structured data from 100 autism VAERS reports, it was able to provide structured JSON only for 65 of them.

These are the results from a little Python script I use to test how many times the LLM was able to extract structured data from 100 VAERS reports

Audit complete. Found 100 JSON files, 65 with valid JSON, updated 0 files.

While it is possible that things could improve over time, if an LLM is as expensive as GPT-5 Full, I expect it to be a lot more reliable when it comes to producing structured outputs.

Description from OpenRouter site:

Cogito v2 is a multilingual, instruction-tuned Mixture of Experts (MoE) large language model with 671 billion parameters. It supports both standard and reasoning-based generation modes. The model introduces hybrid reasoning via Iterated Distillation and Amplification (IDA)—an iterative self-improvement strategy designed to scale alignment with general intelligence. Cogito v2 has been optimized for STEM, programming, instruction following, and tool use. It supports 128k context length and offers strong performance in both multilingual and code-heavy environments. Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs