GPT-5.5 Instant, SubQ 12M Context, Gemini Flash Updates

Three major AI model updates landed in early May 2026, each targeting different segments of the generative AI market. OpenAI released GPT-5.5 Instant as its new default model for ChatGPT, Subquadratic announced a 12-million-token context window model, and Google updated its Gemini Flash offering with enhanced reasoning capabilities.

The OpenAI update replaces GPT-5.3 Instant as the default ChatGPT model. According to TechCrunch, the company emphasized improvements in factual accuracy and reduced hallucinations in sensitive domains like law, medicine, and finance. The model scored 81.2 on the AIME 2025 math benchmark, compared to 65.4 for its predecessor. On the MMMU-Pro multimodal reasoning benchmark, it achieved 76 versus 69.2 for the older version.

Context management is where the update gets interesting. GPT-5.5 Instant can now use its search tool to reference past conversations, files, and Gmail accounts for more personalized responses. This feature initially rolls out to Plus and Pro users on the web, with mobile deployment planned shortly. Free, Go Business, and enterprise users will gain access in coming weeks. ChatGPT will also display memory sources across all models, letting users delete outdated sources or correct them if answers were wrong.

For developers, the model is available through API as "chat-latest," with GPT-5.3 remaining as an option for paid users for only three months. This follows a pattern that has frustrated some users—when OpenAI deprecated GPT-4o in February 2026, significant backlash emerged from users who had formed attachments to that model's personality.

Subquadratic made arguably the more technically ambitious claim with its SubQ model. The Miami-based startup launched with $29 million in seed funding and announced a 12-million-token context window. That's roughly 9 million words, or about 120 books loaded into a single prompt. The company says it plans to offer a 50-million-token window model soon.

The architecture behind this is called Subquadratic Selective Attention (SSA). Traditional transformer models use dense attention, where every token compares with every other token—doubling input quadruples the work. Subquadratic claims SSA scales linearly in both compute and memory with respect to context length. The company reports 52 times faster performance than dense attention at 1 million tokens, 92.1% accuracy on needle-in-a-haystack retrieval at 12 million tokens, and an 83 score on MRCR v2, beating OpenAI by nine points.

On SWE-bench Verified, SubQ reports 82.4%, edging out Anthropic's Opus 4.6 at 81.4% and Google's Gemini 3.1 Pro at 80.6%. The company is launching the SubQ API with the full 12-million-token window, plus SubQ Code (a command-line coding agent) and SubQ Search (initially free). The model will not be open-weight or open-source in the near term, though it will be trainable for customer-specific use cases.

These are large claims, and the skepticism is warranted. The quadratic cost of attention has been a fundamental constraint since 2017. Previous attempts—sparse attention models like Longformer, state-space models like Mamba, hybrid architectures like Jamba—have all traded one necessary property for another. Subquadratic CTO Alex Whedon told The New Stack that SSA's selection mechanism itself does not go quadratic, which would be the breakthrough if verified.

Google's Gemini Flash updates take a more incremental approach. The company released Gemini 3 Flash as the new default model in the Gemini app, with improvements over Gemini 2.5 Flash in speed and reasoning capabilities. The model can reason at a PhD level, similar to larger models, according to Google's official release notes.

Gemini 3 Deep Think is available to Google AI Ultra subscribers—a specialized reasoning mode using iterative logical reasoning cycles to explore multiple hypotheses simultaneously. This mode typically takes several minutes to prepare answers for complex problems in mathematics, science, and logic. The update also includes a new Mac desktop app with keyboard shortcut access (Option + Space) and screen-sharing capabilities for contextual understanding.

User feedback on Reddit suggests the models themselves are competent, but the Gemini app's harnesses lag behind competitors. Users report that when given multiple long documents to read, Gemini often skims the first page of one file and hallucinates the rest, while ChatGPT launches agents that actually complete the requested tasks. The speed advantage of Flash is real for agentic workflows requiring many tokens for tool use and planning, but the app experience remains a pain point.

These three updates represent different strategies: OpenAI optimizing for accuracy and personalization, Subquadratic attempting to break fundamental scaling constraints, and Google balancing speed with reasoning depth. The next few months will show whether Subquadratic's claims hold up under real-world scrutiny, whether OpenAI's default model changes stick, and whether Google can improve its app harnesses to match its model capabilities.

For developers and enterprise users, the choice depends on use case. GPT-5.5 Instant offers the most polished experience for general tasks. SubQ offers unprecedented context length if the claims prove accurate. Gemini Flash provides speed advantages for agentic workflows, though the app experience needs work. All three are available now through their respective APIs and platforms.

What's clear is that the AI model race is intensifying. The context window wars continue, reasoning capabilities are improving, and the focus is shifting from raw model performance to how well these models integrate into real workflows. The next breakthrough may not come from bigger models, but from better architectures that scale efficiently.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn