A Chinese AI lab just released a model so cheap and open, it might have pushed OpenAI to start testing GPT 5.6 earlier than planned. While many watched the usual AI race, Deepseek arrived with V4, slashing API prices by up to 90%. This isn't just about price; it’s about a new multimodal system and a shift in the entire industry.

At the same time, OpenAI has been dealing with an odd bug in GPT 5.5, where the model became obsessed with goblins and other creatures. Then, developers spotted 'GPT 5.6' in Codex backend logs. The timing makes you wonder: did this Chinese lab just put OpenAI into fast-forward?

Key Takeaways

  • Deepseek V4 is an open-source model with significantly lower API prices, validated on both Nvidia and Huawei chips.
  • Its new 'Thinking with Visual Primitives' approach improves multimodal AI by using visual markers as reasoning tools.
  • OpenAI's GPT 5.5 faced a 'goblin bug' while its Codex app expanded, and 'GPT 5.6' appeared in backend logs.
  • The global AI market is seeing a split between closed US models and cheaper, more open Chinese alternatives.
  • Increased enterprise AI usage means cost efficiency can drive widespread adoption, even if a model isn't the absolute top performer in every benchmark.

Deepseek V4: Open, Cheap, and Everywhere

Deepseek V4 entered a crowded market but made an immediate impact. It's open source, so users can download and modify it. It's also incredibly cheap to run. For V4 Pro, the cost per million input tokens reportedly dropped from around 14.5 cents to just 3.6 cents. In China, input costs for V4 Flash fell to 0.02 yuan per million tokens.

This model also works on both Nvidia and Huawei Ascend processors. Chinese chip companies like MetaX and Cambricon now support it. The China Academy of Information and Communications Technology is testing it, a sign that this model is part of a larger national push to build an AI ecosystem less dependent on advanced Nvidia chips. If Huawei's Ascend 950 super nodes become widely available, V4 Pro could get even cheaper.

A Smarter Way for AI to See

Deepseek also released a new technical report, 'Thinking with Visual Primitives,' developed with Peking and Tsinghua Universities. This work addresses a weakness in multimodal AI: models often struggle to keep track of objects they're talking about in an image, a problem the report calls the 'reference gap.'

Instead of just trying to 'see' more pixels, this research gives the model a 'cyber finger.' The system uses points and bounding boxes as reasoning tools. When it discusses an object, it can anchor that object to coordinates. For example, instead of just saying 'the bear on the left,' it can attach a box around the bear and keep referring to that exact location as it thinks. This makes visual tasks like counting in a crowd or solving a maze much more accurate. The system does this while using far less visual memory than rivals, meaning faster answers and lower costs, especially for real-time systems like robots.

OpenAI's Week: Goblins and GPT 5.6

OpenAI had a strange week. Users noticed GPT 5.5 kept mentioning goblins, gremlins, and other creatures out of context. OpenAI reportedly banned these words in its Codex system prompt, which only made users try harder to trigger them.

Meanwhile, the Codex app is rapidly expanding, summarizing changes, analyzing data across Slack and Gmail, and assisting with decisions. Greg Brockman, OpenAI's President, said he loved the Codex app. Sam Altman joked about the 'Goblin moment' but highlighted Codex's 'ChatGPT moment.' Right in the middle of all this, 'GPT 5.6' appeared in backend logs. This likely points to early internal testing, but the timing suggests a response to increasing market pressure.

What This Means for the AI Market

IDC's Chang Mang notes the global AI market is splitting into two camps: US closed models and Chinese open-source models. Deepseek V4's cost advantage is significant. As enterprise AI usage explodes – with companies like Visa spending trillions of tokens in a month – cheaper, capable models change behavior. Teams use more AI, automate more workflows, and question premium prices. This aligns with Jevons' paradox: when something becomes cheaper and more useful, people consume more of it.

A cheaper model doesn't need to win every benchmark. It just needs to be good enough for daily tasks. That cost advantage then does the rest. Deepseek V4 isn't just a new model; it's a strategic move that could reshape the AI landscape. OpenAI might still lead at the very top, but the pressure from below is growing fast.