Tested and written up.
Added Jul 7, 2026
Claude Fable 5 is Anthropic's Mythos-class model released on June 9, 2026, focused on creative writing, worldbuilding, and narrative depth. It was suspended from distribution on June 12, 2026, under US export controls, making it a limited-availability release.
Why: Claude Fable 5 is notable as Anthropic's most experimental creative model. Even with limited availability, it represents an interesting direction for AI-assisted fiction and long-form creative work.
Added Jul 9, 2026
GPT-5.6 Sol is the highest-capability model in OpenAI's GPT-5.6 family, released July 9, 2026 after a limited preview starting June 26. It has a 1.05M-token context window with up to 128K output tokens, and is positioned for complex professional work: advanced coding, scientific and technical reasoning, long-document analysis, computer use, and multistep agent workflows. It sits alongside the cheaper Terra and Luna tiers in the same family: Sol is $5 input and $30 output per million tokens, Terra $2.50 and $15, Luna $1 and $6, and on July 30, 2026 OpenAI cut Luna's price by 80% and Terra's by 20%. Free and Go users of ChatGPT get Terra; paid users choose any of the three and set effort per model.
Why: Sol ranks third overall on the Artificial Analysis Intelligence Index at 58.9, behind Claude Opus 5 and Claude Fable 5, a genuine top-tier frontier model rather than an incremental update, and OpenAI's clear flagship pick for the hardest professional-grade tasks.
Added Jul 9, 2026
Grok 4.5 is xAI's flagship model, shipped July 8, 2026 with public rollout July 9. It has a 500K-token context window (with a high-context surcharge above 200K) and configurable reasoning effort (low/medium/high, defaulting to high). xAI trained it in partnership with Cursor specifically to handle long-running jobs across multiple repositories with minimal human intervention across hundreds of tool calls.
Why: xAI positions it as its flagship 'for code and everything else,' and real benchmark data backs that up, 53.8% on the Artificial Analysis Intelligence Index (rank 7 overall). The Cursor training partnership is a distinctive angle: it's specifically tuned for long-horizon, multi-repository coding agent work, not just general chat.
Added Jul 7, 2026
OpenCode is an open-source terminal-based coding agent that connects to multiple language models and performs autonomous software engineering tasks from the command line. It edits files, runs tests, manages git workflows, and iterates on code with minimal human intervention. The v1.17.8 release from June 2026 improves tool calling reliability, multi-file refactoring, and support for local and remote model backends.
Why: OpenCode is the best 'bring-your-own-model' coding agent for developers who want full control. Because it is open source and runs in the terminal, it fits naturally into existing CI/CD and shell-centric workflows without locking you into a specific vendor.
Added Jul 7, 2026
Kimi K2.7-Code is Moonshot AI's coding-specialized model released on June 12, 2026. It is tuned for software engineering tasks including code generation, debugging, refactoring, and technical reasoning in both English and Chinese contexts.
Why: Kimi K2.7-Code is one of the strongest coding models from a Chinese AI lab, with particular strength in long-context understanding and bilingual code tasks. It is a good addition for teams evaluating global coding models.
Added Jul 9, 2026
GPT-5.6 Terra is the mid-tier model in OpenAI's GPT-5.6 family, released alongside Sol and Luna in July 2026. It shares the same 1.05M-token context window and 128K max output as Sol but is optimized for workloads that balance capability, latency, and cost. It supports text and image input, function calling, web search, file search, computer use, image generation, and code interpreter, making it a practical default for general-purpose reasoning and agentic workflows.
Why: Terra is the sensible default for most GPT-5.6 work: it delivers the lion's share of Sol's capability at roughly 40% of the cost and is the default model for ChatGPT Free and Go users.
Added Jul 9, 2026
GPT-5.6 Luna is the smallest and cheapest model in OpenAI's GPT-5.6 family, released alongside Sol and Terra in July 2026. It is designed for cost-sensitive, high-volume workloads where latency and price matter more than absolute frontier performance. It shares the same 1.05M-token context window and multimodal input support as Sol and Terra, making it suitable for classification, summarization, light coding, chat, and high-throughput agent workflows.
Why: Luna brings GPT-5.6-scale capabilities to high-volume applications at roughly one-tenth of Sol's cost, with strong enough performance for everyday tasks and broad API availability.
Added Jul 7, 2026
TRELLIS 2 is an open-source image-to-3D generation model from Microsoft Research, released in 2026. It reconstructs 3D assets from single images or text prompts and is designed for research and experimentation.
Why: TRELLIS 2 is a valuable open research model for image-to-3D. It is ideal for academics, indie developers, and anyone who wants to run 3D generation locally or build on top of open weights.
Added Jul 7, 2026
Microsoft announced a family of MAI-branded models at Build 2026, including MAI-Thinking-1 for reasoning, MAI-Image-2.5 for image generation and editing, MAI-Voice-2 for expressive text-to-speech, MAI-Transcribe-1.5 for speech-to-text, MAI-Code-1-Flash for coding in GitHub Copilot, and Scout as a workplace personal agent. They integrate tightly with Microsoft 365, Azure, and GitHub.
Why: The MAI family gives Microsoft a cohesive, enterprise-ready AI stack. For organizations already using Microsoft services, these models reduce friction by running inside familiar tools rather than requiring separate platforms.
Added Jul 7, 2026
NVIDIA Cosmos 3 is an open physical-AI omnimodel released around May 31 to June 1, 2026. It generates video, 3D, and physical-world simulations to train and evaluate robotics and autonomous vehicle systems without expensive real-world data collection.
Why: Cosmos 3 is a major open contribution to physical AI. By simulating realistic worlds, it can accelerate training for robots and self-driving cars while reducing the need for dangerous or costly real-world trials.