Tech • AI • Robotics
OpenAI’s Secret New Model Makes Fable 5 Look Primitive
9.4/10TL;DR
A rumored OpenAI model codenamed Doug is being discussed as the company’s largest pre-training run yet, but its eventual release may depend less on raw scale than on government review, safety testing, and rising concerns over advanced cyber capabilities.
KEY POINTS
Rumor centers on timing and scale
Discussion around Doug points to a launch by November and describes it as OpenAI’s biggest pre-train so far. The claim remains unverified, but the cited rationale is notable: extended pre-training, possible White House involvement, and extensive cybersecurity testing. That combination suggests the bottleneck may be policy and risk review as much as engineering.
Pre-training is no longer the whole story
Recent AI progress has come less from larger base models than from post-training, especially reinforcement learning and other alignment steps. Hardware efficiency, cleaner datasets, and synthetic data have improved results, but scale alone no longer guarantees a breakthrough. A very large pre-train could still disappoint if the post-training stack fails to deliver.
Government demos would imply training is largely finished
One reason some observers treat the rumor as plausible is the idea that if a model is already being shown to government officials, its core pre-training is probably complete. At that stage, the remaining work would center on safety evaluation, red-teaming, and RL tuning. That would fit a delayed rollout without requiring ongoing giant training runs.
Model comparisons remain highly contested
Claims that a new system could make Fable look primitive triggered debate over how fast model quality is actually improving. Some developers argue the newest coding models have advanced dramatically in months, while others say practical gains are narrower than demos suggest. A recurring complaint is overwriting: models often generate large volumes of code or changes that users did not ask for.
Synthetic demos do not equal production reliability
Several developers drew a line between impressive one-shot outputs and real software work with fixed requirements, maintenance burdens, and edge cases. In controlled demos, a model can appear transformative; in production, teams still have to guide, constrain, and correct it. The gap between “looks solved” and “is deployable” remains substantial.
AI coding is spreading fast inside companies
Enterprise users report that business units increasingly want their own agentic coding environments, and some are already building internal tools without waiting for formal approval. That is fueling anxiety among software teams that coding may become a commodity within two to three years. Even so, subject-matter expertise still appears to determine who gets the best results from these systems.
The cyber policy shift is already documented
The broader backdrop is a move by Anthropic and OpenAI toward tiered access for high-risk capabilities. Anthropic launched Fable 5 for general users with protections that reroute dangerous cybersecurity and biology requests to Claude Opus 4.8, while more capable systems such as Mythos 5 are being limited through a trusted access program. More than 200 companies and governments have reportedly been included in expanded access discussions.
OpenAI is also separating public and restricted models
OpenAI has created a vetted structure that gives approved researchers and organizations access to less restricted versions of its systems for malware analysis, bug hunting, and reverse engineering. The newest example is GPT 5.6 Cyber, derived from GPT 5.6 Sol and available only through Daybreak Red, the higher tier of the company’s access program.
Cyber benchmarks show why restrictions are tightening
On an internal benchmark covering exploit chains, authentication bypass, and privilege escalation, GPT 5.6 Cyber reportedly completed 95% of requests, compared with 1.5% for the standard guarded GPT 5.6. On Exploit Gym, it outperformed both GPT 5.6 Sol and GPT 5.5 Cyber at turning known vulnerabilities into working exploits in controlled environments.
Real vulnerabilities have already been found
OpenAI researchers used the cyber model to identify two previously undocumented flaws in V8, the JavaScript engine used by Chrome, which were patched by Google as CVE-2026-1593. Researchers also reported finding high-severity issues in a mobile operating system, a widely used database, and an operating system kernel, though those have not been publicly named.
Astra’s delay changes how Doug is viewed
The rumor carries more weight because it follows the withdrawal of Astra after preliminary testing suggested it might reach the highest tier of hacking capability under OpenAI’s own preparedness framework. That means the question around Doug may not simply be whether it works well enough to ship. It may be whether it is judged safe enough to release at all.
CONCLUSION
The emerging divide in advanced AI is no longer just between stronger and weaker models, but between models that can be broadly released and models restricted for national-security and cyber-risk reasons. If Doug arrives this year, its significance may lie as much in access policy as in raw capability.
Explain this