AI GUIDEPartner content

OpenAI Bell: Scaling Compute and Reasoning Tests

Analysis of tests of OpenAI’s new flagship model: outperforming GPT-6 Astra in reasoning, agent experiments, and enhanced isolation protocols.

Affiliate link: your price stays the same and the project earns a commission.

A qualitative leap in reasoning scaling

OpenAI’s lab is conducting closed testing of an internal next-generation model that substantially surpasses the capabilities of its flagship GPT-6 Astra. In the analytics community, the system is known by the codename Bell. The key feature of the new architecture is a sharp increase in efficiency when using computational resources during reasoning generation (test-time compute).

Whereas previous generations of transformers scaled primarily by increasing the number of parameters and the volume of pretraining data, the new internal system demonstrates a qualitative leap specifically in the depth and reliability of multistep logical reasoning.

The main indicator of the new system’s capabilities was an experiment on a fundamental mathematical problem of the millennium: investigating the smoothness of solutions to the three-dimensional Navier–Stokes equations, which describe hydrodynamics and the motion of gases. The research pipeline deployed an infrastructure of approximately 10 000 interacting AI agents, which continuously performed computations for 88 hours.

In comparative measurements of reasoning reliability at comparable compute expenditure:

  • GPT-6 Astra: showed a probability of successfully completing critical mathematical proofs of about 0,1%.
  • New model (Bell): with similar compute costs, demonstrated a score of about 0,25%, showing more than double the efficiency in using allocated resources.

This confirms that the system does not simply try more options, but builds its solution search tree more selectively and accurately, pruning dead-end hypotheses.

Real-time training and quarantine protocols

According to official information, training of this internal architecture began on August 28 and is ongoing. With each optimization cycle, results on internal olympiad tests continue to improve, meaning that the published metrics reflect only an intermediate state of the network.

The increase in capabilities is accompanied by stricter precautionary measures. Because the emergent properties of such powerful networks are difficult to predict, OpenAI applies enhanced isolation protocols:

Control level Implementation Purpose
Inference sandbox Complete network isolation of test environments Preventing unauthorized model access to external infrastructure
Goal monitoring Continuous end-to-end auditing of hidden steps in reasoning chains Preventing agents from choosing undesirable or deceptive subgoals
Quarantine period Multiweek stress testing before a decision on commercial release Expert red teams testing the boundaries of the model’s controllability

ChatGPT Images 2.5: a breakthrough in editing consistency

Alongside its language research, OpenAI updated its generative graphics stack, introducing ChatGPT Images 2.5. A longstanding problem with diffusion systems was loss of composition: trying to change a small detail (for example, the background color or footwear) could cause accidental changes to facial features, clothing, and lighting.

The updated version includes the following improvements:

  • Spatial locking: precise masking and modification of the specified area while preserving the rest of the frame’s geometry completely.
  • Multistep sequence: stable preservation of the same character and style across a series of dozens of iterative edits.
  • Blind-test leadership: in the Text-to-Image Arena ranking, the Sunburst and Flare versions of the model took the top spots, displacing solutions from Google that had previously led the field.

Prospects for the development of reasoning architectures

The pace of OpenAI’s internal development sets the direction for the entire industry. The main takeaway for engineers and product teams: competitive advantage is shifting decisively away from simply calling base LLMs and toward agentic verification architectures and flexible management of the reasoning budget for each incoming request.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

OpenAI Bell GPT-6 Astra AI reasoning test-time compute AI Navier–Stokes ChatGPT Images 2.5 model scaling

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 4 October 2026

Related reading

All in this section →