Subquadratic launches AI architecture with 12-million-token context window, outperforming GPT-5.5 on retrieval benchmarks
By
Frederic Lardinois
Pulled from the oven just right. Trustworthy, fact-dense, deeply satisfying.
Summary
Subquadratic has launched a new AI architecture featuring a 12-million-token context window, shattering the current million-token standard set by frontier models like GPT-5.5 and Claude Opus 4.7. While most models struggle to effectively utilize their advertised context windows (with GPT-5.5 scoring only 74.0% on the MRCR v2 retrieval benchmark), Subquadratic's architecture outperforms GPT-5.5 on retrieval benchmarks. The article explains that the million-token limit in current models stems from the computational constraints of attention mechanisms in transformer-based architectures, which Subquadratic's approach aims to overcome.
Key quotes
· 3 pulledEvery frontier model in 2026 advertises a context window of at least a million tokens, but almost none of them are actually great at making use of all of that information.
On MRCR v2, the multi-reference retrieval benchmark labs report, the best model is GPT-5.5, which scores 74.0%. Others like Claude Opus 4.7 at 32.2% are far behind.
One major reason for the million-token max is the same one that has shaped every transformer-based model since 2017: Attention co
You might also wanna read
Anthropic Expands AI Model's Context Window to 1 Million Tokens in Competitive Push
Anthropic has significantly increased the context window of its AI model Claude Sonnet 4 to 1 million tokens, marking a 5x improvement. This
The Verge·9mo ago
DeepSeek-V4: Hybrid Sparse-Attention Architecture Enables Efficient Million-Token Context Inference
DeepSeek-V4 introduces a hybrid sparse-attention architecture combined with on-policy distillation across domain specialists, enabling 1M-to
artgor.medium.com·1d ago
Datacurve's DeepSWE Benchmark Shows GPT-5.5 Leading AI Coding Models with 70% Pass Rate
A new benchmark called DeepSWE, released by startup Datacurve, reveals significant performance differences among AI coding models that were
share.transistor.fm·5d ago
OpenAI launches GPT-5.5 with improved coding and cross-tool capabilities
OpenAI has announced GPT-5.5, its latest AI model, just one month after releasing GPT-5.4. The company claims the new model excels at writin
The Verge·1mo ago
MiniCPM 4.0: Open-source 8B multimodal AI model outperforms GPT-4o and Gemini Pro on vision benchmarks
MiniCPM 4.0 is an ultra-efficient 8B open-source multimodal AI model designed for on-device use that outperforms larger models like GPT-4o a
Product Hunt·9mo ago