OpenAI’s GPT-5.3 Leak Hints at a Smarter, Faster, Cheaper AI

While the AI world has been buzzing about the power of Google’s Gemini 3 and the coding finesse of Claude 4.5, leaks suggest OpenAI is preparing a “tactical strike” with GPT-5.3. 

This isn’t just a minor patch; it’s an attempt to reclaim the crown by focusing on what insiders call “cognitive density” — making the AI smarter and faster without simply increasing its size.

According to leaks shared by multiple sources, the new version is codenamed “Garlic.” The name is a metaphor for the model’s design: just as a single clove of garlic can flavor an entire dish, this model is built to provide concentrated intelligence without the massive computational weight of its predecessors.

Rather than chasing “ever-larger parameter counts,” OpenAI is reportedly betting on a technique called Enhanced Pre-Training Efficiency (EPTE). This process “prunes” redundant data during training, resulting in a model that is physically smaller but retains the world knowledge of a much larger system. This shift is expected to lead to faster response times and significantly lower costs for developers.

New features: ‘Salute’ and smart maps

Technical researcher Tibor Blaho recently spotted several major changes hidden in the ChatGPT web app. The most intriguing feature is codenamed “Salute.” According to the leaked code, this feature will allow users to “create tasks with file uploads and track their progress in Salute.” 

The updates also point toward a better experience for local searches.

A new “is model preferred” flag suggests that ChatGPT will soon be able to pick specific models optimized for local businesses, restaurants, and hotels within its map widgets. Additionally, the user interface is getting an upgrade with “inline editable code blocks and math blocks,” which will allow for more interactive editing directly in the chat.

Under the hood: Secure tunnels and massive context

For the power users and developers, leaks indicate support for a new secure tunnel for Model Context Protocol (MCP) servers. 

According to the reference found by Blaho, this “Secure Tunnel connects your internal MCP server to OpenAI via a customer-hosted tunnel client over outbound-only HTTPS, so no inbound firewall changes are needed.”

Beyond security, the model’s “memory” is getting a boost. While it may not match Gemini’s massive window, GPT-5.3 is rumored to feature a 400,000-token context window with “Perfect Recall.” Perhaps even more impressive is the rumored 128,000-token output limit, which could allow the AI to write entire software libraries or full-length books in a single go.

The ‘Code Red’ strategy

This sudden push reportedly stems from an internal “Code Red” declared by Sam Altman in December 2025.

After Anthropic’s Claude 4.5 became the favorite among developers and Gemini 3 took the lead in multimodal benchmarks, OpenAI reportedly paused several side projects to focus entirely on this release. Then, in December, OpenAI released GPT 5.2, the company’s most capable model for coding and agentic workflows.

But OpenAI is looking to strike gold with the 5.3 model… and internal benchmarks shared by CometAPI suggest the effort is paying off.

GPT-5.3 is allegedly hitting 94.2% on coding benchmarks (HumanEval+), outperforming both Gemini 3 (89.1%) and Claude 4.5 (91.5%). Reliability is also a key focus; the model is being trained with “epistemic humility” to ensure it knows when to say “I don’t know,” drastically reducing hallucinations.

When can we expect it?

The wait shouldn’t be long. Reliable leaker Dan McAteer suggested the model will feature “Stronger pretraining” and “IMO Gold winning reasoning techniques.” When asked about a timeline on X, McAteer noted there is “no ETA but when I’ve gotten these tips before it’s been like 1 week, 2 max.”

Industry experts expect a staggered rollout starting in late January 2026, beginning with a “Preview” for ChatGPT Pro users and enterprise partners, followed by a full API release in February.

Also read: Google’s Gemini 3 Flash rollout shows how speed-and-cost positioning is becoming a central battleground in the AI race.