Google DeepMind has launched Gemini 3.7 Flash, with enhanced capabilities in programming, agents, web development, and knowledge work. Its DeepSWE score improved from 49% in version 3.6 to 65.3%, FrontierCode rose from 34.4% to 43.6%, and GDP.pdf increased from 22.0% to 34.0%. The launch pricing is set at $0.75 and $3.75 per million input and output tokens respectively—approximately half the original price of the previous generation. The model is now available via Gemini API, AI Studio, Cursor, and Gemini Spark, though weights are not open-sourced.

OpenAI has introduced Computer History in the ChatGPT macOS desktop app, which builds a timeline of user activity based on interactions such as clicks, typing, and application switching—without recording screen or audio. ChatGPT and Codex can leverage this timeline to identify recently edited documents, summarize work sessions, and bridge tasks across applications. The feature uses an opt-in mechanism: users can exclude specific apps and websites, review or delete records, and access local memory files. It is now rolling out globally to Pro, Business, and Enterprise users, with availability in the European Economic Area, UK, and Switzerland coming later.

OpenAI has previewed the Ultrafast mode for GPT-5.6 Sol, leveraging Cerebras infrastructure to achieve up to 750 tokens/second in generation speed—up to 14 times faster than the standard mode. This mode is initially available through the API to select customers, targeting time-sensitive applications such as real-time customer support, financial research, incident investigation, and code refactoring. A demonstration showed that data collection, organization, and context augmentation, which previously took 1–2 hours, can now be completed in 10–15 minutes. OpenAI is currently optimizing the experience with early customers and accepting sign-ups for future capacity expansion.

DeepSeek has open-sourced Harness, which adopts a plugin-first architecture called Cordis, allowing runtime components to be designed as replaceable plugins. It offers four preset agent types: Standard, PTC, Minimal, and Creative. The PTC mode supports a TypeScript SDK, and users can combine custom presets with Skills. Its Trajectory window logs event-level raw data, enabling playback of model behavior, identification of execution bottlenecks, and tracking of token usage—improving debugging and cost control for long-running tasks. Tests reported in coverage show it capable of continuous reasoning for durations ranging from 20 minutes to even 10 hours, though comparisons are primarily based on experiential demonstrations.

Google has launched Gemini-powered Sheets canvas for Google Sheets, enabling users to convert spreadsheet data into interactive, visual mini-applications using natural language—no formulas or coding required. Canvas acts as a read-write layer atop the original sheet, generating interfaces such as learning trackers, sports dashboards, or wedding seating charts. Users can iteratively refine layout and functionality via prompts, with changes synchronized bidirectionally between Canvas and underlying cells in real time. Completed apps can be shared as Sheets tabs and support multi-user collaborative editing while preserving data consistency.

Cursor has updated its cloud Agent Builds mechanism by continuously pre-provisioning development environments, reducing startup time from minutes to seconds—a 3x improvement in launch speed. This optimization targets asynchronous delivery of long, complex tasks, eliminating repeated environment setup delays at the start of each job. Enterprise clients including Faire, Headway, and Descript have reportedly observed these performance gains. Additionally, Builds now incorporate validation into the deployment pipeline, preventing failed builds from entering production and maintaining clearer debugging paths—enhancing resilience and debuggability of cloud Agent execution.

A multi-agent engineering experiment used OKF to share precomputed token IDs across different-sized Qwen2.5-Coder models, eliminating redundant tokenization of identical text across agents. The approach adds a token_pointer field in OKF, stores the .npy array in /dev/shm shared memory, and performs full consistency checks across a 151,936-item vocabulary to prevent silent errors from mapping discrepancies. For the 3B model, first-token latency dropped from 69.3 ms to 49.9 ms (28% reduction); for the 1.5B model, it decreased from 49.6 ms to 30.9 ms—a 37.8% improvement. However, differing hidden dimensions still prevent KV cache sharing.

An engineering benchmark deployed Gemma 4 E2B on AWS EC2 G5g instances, combining Graviton2 ARM64 processors with NVIDIA T4G GPUs, achieving 43.1 tokens/second using a patched version of vLLM. This hardware combination lacks public precompiled support, and Gemma 4 uses attention heads of 256 and 512 dimensions across layers, making the Turing architecture’s 64 KiB shared memory insufficient for default Triton kernels. The solution leveraged AWS ARM64 GPU DLAMI and vLLM version 0.27.2rc0 or higher, reducing KV tile size and pipeline stages. Final performance was constrained by 277 GB/s GPU memory bandwidth.