OpenAI's GPT-6 Recurrent Depth and Mostik's Token-Free Bridges Challenge AI Pricing

The Information reported that OpenAI's new technique has not been officially confirmed but represents an architectural shift. Instead of outputting a readable chain of thought, the model repeatedly computes within internal loop blocks before generating an answer. Lei Feng Network notes that this puts the reasoning process beyond human inspection at the architectural level. OpenAI chief scientist Jakub Pachocki is concerned that the public disclosure could push the industry toward an unmonitorable arms race.

In a conventional transformer, every generated token passes through a fixed number of layers. Recurrent depth reuses a set of layers in a loop, allowing the model to refine an intermediate state without expanding the network. A 2025 paper titled Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach found that a 3.5-billion-parameter model iterated 32 times could match or approach the reasoning performance of much larger dense models on math and coding tasks, with bounded memory usage. Similar ideas are being explored by Meta in projects such as continuous thought chains and COCONUT, Lei Feng Network reported.

According to the Lei Feng Network article, the Russian startup Mostik has taken the trend further. Led by a team of mathematicians including founder Sasha Malysheva and 2010 Fields Medalist Stanislav Smirnov, Mostik trains a lightweight bridge matrix that projects one model's latent semantic state directly into another model's input space. No tokens are generated and no pre-trained weights are modified. In a test described in the article, a bridge was built between GLM-5.2 (753B parameters) and Qwen-3.5 (4B parameters). The combined system consumed one-twentieth of the inference cost of GLM-5.2 alone while achieving roughly 80 percent of its accuracy, and it topped the ARC-AGI benchmark. A former Google DeepMind scientist praised the bridge as releasing new efficiency in model collaboration.

The central issue is the token economy. Today's AI business model equates computation with tokens: APIs bill per token, and major technology companies have guided up to $750 billion in AI infrastructure capital spending based on projected token consumption. Recurrent depth can produce very few tokens, even when computation is heavy, severing the link between output tokens and actual computing cost. Cross-model bridges remove the intermediate text that was pure overhead in agent workflows. Lei Feng Network argues that these developments are slicing the token economy from both the thinking and collaboration sides. If AI is no longer billed by the token, the article asks, how will AI companies charge for their services?

As models retreat from human-readable language into abstract vector spaces, the article observes a dual consequence: users may receive correct answers while losing the ability to verify why those answers are correct. Proposed alternatives for pricing might shift to real computation, number of loops, or outcome-based fees, but no consensus has emerged.