🤖 AI Weekly Recap (Week 18)

DeepSeek-V4-Pro is priced at roughly 1/6th the cost of Claude Opus 4.7 and 1/7th the cost of GPT-5.5. The smaller V4-Flash variant is nearly 98% cheaper than the premium proprietary models.

Features a native 1-million-token context window, utilizing a radical new "Hybrid Attention Architecture" that requires only 10% of the memory footprint (KV cache) of its predecessor.

While GPT-5.5 and Opus 4.7 still hold slight leads on direct reasoning benchmarks, V4-Pro-Max gets unusually close, scoring 83.4% on the agentic BrowseComp benchmark to beat Opus 4.7 and nearly match GPT-5.5.

Introduces three reasoning modes ("Non-think", "Think High", and "Think Max") to let users efficiently match compute effort to the difficulty of the task.

Paid subscribers (Plus, Pro, Business, Enterprise) can use the new thinking model to generate up to 8 images at once while perfectly maintaining characters, objects, and styles across every scene.

The system can now pull real-time information from the web or use uploaded files to create highly accurate visual explainers.

Supports ultra-high resolution up to 2K and extreme aspect ratios ranging from wide 3:1 formats to tall 1:3 formats.

Massive improvements to in-image text generation, now reliably supporting Japanese, Korean, Chinese, Hindi, and Bengali alongside English.

Matches or beats top models like GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro on major coding and agent benchmarks (scoring 58.6 on SWE-Bench Pro and 83.2 on BrowseComp).

The "Agent Swarm" headline feature runs up to 300 specialized sub-agents at once, each capable of taking 4,000 steps to seamlessly combine skills like web research, data analysis, and writing.

Built for endurance, the model can chain together more than 4,000 tool calls and run continuously for over 12 hours in languages like Rust, Go, and Python.

Features "claw groups," allowing multiple AI agents and human developers to work together as a team, with K2.6 managing coordination and stepping in if an agent gets stuck.

Agents can independently gather product feedback from the web and drop a summarized report into Slack, or automatically draft follow-up sales emails in Gmail.

Teams can build an agent once, share it across the organization, and let it safely operate across integrated tools while asking for human approval when necessary.

This move directly answers the viral explosion of agentic frameworks like OpenClaw (whose founder now works at OpenAI) and escalating competition from Anthropic’s Claude Cowork platform.

OpenAI stated this is an "evolution" of the 2023 custom GPTs; while GPTs remain available for now, the company plans to seamlessly transition them into full workspace agents soon.

12-Hour Long Horizon Coding: The headline feature allows developers to point the CLI at a codebase, walk away, and let the agent code autonomously for up to 12 hours straight without timing out or losing context.

Agent Swarm Architecture: Instead of working sequentially, Kimi Code can spawn and coordinate up to 300 specialized sub-agents in parallel to execute massive batch refactors or build out full full-stack applications.

Deep Terminal Integration: It features a built-in shell mode (meaning you can run shell commands without leaving the agent), native Zsh integration, and full support for the Model Context Protocol (MCP) to add custom tools.

Open Weights & Rock-Bottom Pricing: The K2.6 model weights are open-sourced under a modified MIT license. For API users, it costs just $0.60 per million input tokens, roughly 4x to 8x cheaper than frontier models like GPT-5.4 and Claude Opus 4.6.

Introduces a novel "Thinking Preservation" feature that retains the AI's chain-of-thought reasoning across an entire conversation history, saving massive amounts of compute during iterative agent workflows.

Features a highly efficient hybrid architecture that blends Gated DeltaNet (linear attention) with traditional self-attention.

Matches Claude 4.5 Opus on Terminal-Bench 2.0 and dominates in repository-level code generation and frontend workflows.

Available right now on Hugging Face in both BF16 and a highly efficient FP8 quantized version, perfect for pulling down and running locally on your machine.