GPT-5.3 'Garlic': What OpenAI's Next Model Is Rumoured to Include

What Is GPT-5.3 "Garlic"?

On February 23, 2026, users began reporting A/B test sightings of a model identified as GPT-5.3 in the ChatGPT interface. OpenAI has not officially announced it. But based on what has leaked — through user reports, API discoveries, and benchmarking accounts — here is what we currently know.

The internal codename is "Garlic." OpenAI has used vegetable codenames internally before (GPT-4o's codename was "Arrakis," and models in development have been called "Strawberry," "Orion," and others). Garlic appears to be the next iteration in the GPT-5 family, positioned between GPT-5 and whatever comes next in OpenAI's roadmap.

What Has Actually Been Leaked

400,000 token context window. This is the most significant claim and the number that matters most for developers. For comparison:

• GPT-4o: 128,000 tokens
• Claude 3.5 Sonnet: 200,000 tokens
• Gemini 1.5 Pro: 1,000,000 tokens (but with quality degradation at high context)

A genuine 400K context window that maintains quality throughout would mean GPT-5.3 can process approximately 300,000 words — roughly three full novels, or a very large codebase — in a single prompt. For developers building on the OpenAI API, this is a material capability change.

Benchmark improvements on non-coding tasks. Reports from users who accessed the model via A/B test describe noticeably better performance on tasks involving long documents, multi-step reasoning, and creative writing. One claim suggests benchmark scores that beat all current frontier models on knowledge-intensive tasks. This has not been independently verified.

Speed similar to GPT-4o. The A/B test reports suggest the model is not dramatically slower than GPT-4o despite the larger context, which would indicate significant architectural improvements in how long-context inference is handled.

Why the Context Window Number Is the Real Story

Most coverage of new model releases focuses on benchmark scores. The context window is more important than any benchmark.

Here is why: the value of a language model for real work scales with how much relevant context it can see simultaneously. A developer debugging a large codebase cannot fit it all into a 128K context. A lawyer reviewing a 200-page contract cannot get a complete analysis in one prompt with current models. A researcher synthesising multiple papers needs to compare them side-by-side.

400K tokens changes what is possible, not just what is slightly faster. Work that currently requires chunking, summarising, and manually managing context windows becomes possible in a single pass.

For Indian developers specifically: Many development teams in India work on large enterprise codebases — financial systems, government platforms, ERP implementations — that have been difficult to analyse with AI tools because of context limits. A 400K context model opens these use cases in a way that smaller context models have not.

What We Do Not Know Yet

Pricing. Longer context costs more. GPT-4 128K context was priced at a significant premium over GPT-4 8K context. Whether OpenAI prices GPT-5.3's 400K context at a rate that makes it practical for everyday API use is unknown.

Quality at maximum context. The notorious problem with long-context models is "lost in the middle" — the model pays less attention to content in the middle of a very long prompt than to content at the beginning and end. Whether GPT-5.3 has improved on this fundamental issue is not clear from leaked reports.

Release timeline. A/B test sightings two days ago suggest the release is imminent — OpenAI typically does not expose models in A/B tests more than a week or two before full launch. But "imminent" from OpenAI has historically meant anything from 48 hours to three months.

Multimodal capabilities. Whether GPT-5.3 extends the vision, audio, and image generation capabilities introduced in GPT-4o is not yet known.

How This Fits OpenAI's Model Roadmap

OpenAI has been releasing models at an accelerating pace in 2025-2026. The pattern has been:

• GPT-4o: multimodal, fast, cost-effective
• o1 / o3: reasoning-focused, slower, better at STEM problems
• GPT-5: the base next-generation model
• GPT-5.3 "Garlic": an iteration on GPT-5 with expanded context and capability improvements

The trend is toward models that are simultaneously more capable, cheaper to run per token, and available in more contexts. GPT-5.3 fits that trajectory: same speed as GPT-4o (per leaked reports), larger context, better quality.

What Developers Should Watch For at Launch

When OpenAI officially announces GPT-5.3, the key details to check:

• Context window pricing — input tokens vs output tokens, and whether long-context has a premium tier
• API availability — whether it launches simultaneously in the API and ChatGPT, or ChatGPT-first
• Rate limits — new models often launch with tighter rate limits that loosen over weeks
• Instruction following — whether it regressed (new models sometimes do) on precise instruction following compared to GPT-4o

OpenAI's track record is that major model improvements that look significant in A/B tests tend to be real. If the 400K context holds up in production, GPT-5.3 will be a meaningful upgrade for developers building on the OpenAI API.