- Tencent in talks to become Manus' largest shareholder after Meta's $2B deal was blocked
Founder Xiao Hong's agent company may be rerouted from a Silicon Valley exit into a Chinese-backed ownership structure.
- Peter Fenton signals open-weight dominance within two years, via Aligned News
The Benchmark partner's view challenges closed labs' premium, per-token pricing as usage shifts toward cheaper, open-weight inference.
- Head to head: Bagel vs Nano Banana Lite Edit
This wasn’t a squeaker. Across eight image-editing and generation tasks, one model consistently hit the brief while the other too often missed core details or collapsed on execution.
- AI model costs could fall 3-4x in six months, Perplexity CEO predicts
Garry Tan amplified Aravind Srinivas's forecast that local devices could run Opus 4.8-grade models within 12 months.
- Collectly puts its Billie billing agent inside Epic's orbit
The YC company says its Epic bidirectional integration is live in Connection Hub, giving RCM teams a new route to patient billing automation.
- Head to head: Kimi-K2.7-Code vs mistral-medium-3-5
This wasn’t a contest so much as a sweep. Kimi-K2.7-Code dominated the matchup on both score and consistency, beating mistral-medium-3-5 across nearly every task type with a decisive 100% confidence verdict.
- Bun's Rust rewrite puts Jarred Sumner's biggest Zig bet on a new footing
Sumner says Claude helped port 535,496 lines of Zig after memory bugs made Bun's original engineering model harder to defend.
- Apple sues OpenAI, accusing its hardware group of stealing trade secrets
The lawsuit targets OpenAI's Jony Ive-led device push and names io Products, Tang Tan and Chang Liu as defendants.
- Head to head: AnimateDiff vs Gemini Omni Flash
One model showed up as a complete video generator; the other mostly showed flashes of promise without converting them into task wins. Across four prompts, Gemini Omni Flash separated itself on prompt fidelity and scene construction, and the margin wasn’t subtle.
- Insilico's AI-discovered lung fibrosis drug enters Phase III
Rentosertib will test whether Insilico's end-to-end AI drug discovery thesis can survive late-stage clinical scrutiny.
- ZML's Steeve Morin releases free LLMD to loosen AI inference from Nvidia
The Paris startup's closed-source inference server runs Llama, Gemma, Qwen and Mistral models across Nvidia, AMD, TPU, Intel and Apple backends.
- Head to head: xAI: Grok 4.5 vs OpenAI: GPT-5.6 Sol Pro
This was a close, format-sensitive matchup decided less by raw capability than by execution discipline. OpenAI’s GPT-5.6 Sol Pro finishes ahead on the aggregate and task count, but the margin is a lean one rather than a rout.
- OpenAI says GPT-5.6 Sol Ultra produced a proof of a 50-year graph theory problem
Ethan Knight said the run used 64 subagents in under an hour, pushing OpenAI's new ultra mode into peer-review territory.
- Head to head: AuraFlow vs V4.0q [instant]
This matchup wasn’t a rout, but it wasn’t a coin flip either. V4.0q [instant] put more points on the board and took more of the prompt-sensitive tests, giving it a real, statistically supported edge over AuraFlow.
- Melius wants agents, not chatbots, to run creative production
Joowon Kim's startup is betting that designers and marketers will manage agent work on a canvas instead of inside a chat window.
- Drafted V2 turns room placement into AI-generated home plans
Nick Donahue's AI home-design startup added partial regeneration, exterior options, saved projects, and faster floor-plan generation.
- Head to head: Muse Spark 1.1 vs DeepSeek-V4-Pro
This wasn’t a competitive split decision; it was a rout driven by instruction-following, robustness, and cleaner execution across a wide mix of real tasks. Muse Spark 1.1 consistently delivered answers that were tighter, more compliant, and less error-prone than DeepSeek-V4-Pro.
- Head to head: GPT Image 2 API vs Seedream 5.0 Pro Image Editing
This one wasn’t especially close on the numbers, even if several individual prompts were. GPT Image 2 API separated itself by being the more dependable model on prompt-critical composition and spatial fidelity, while Seedream 5.0 Pro Image Editing often looked nicer than it obeyed.
- GPT-5.6 Sol Turns Blender Into an AI Speedrun
A viral X clip showed OpenAI's flagship model driving Blender at a claimed 750 tokens per second in Very Fast mode.
- Mistral moves into robotics with a single-camera navigation model
Robostral Navigate is an 8B embodied AI model that Mistral says can follow language instructions without LiDAR or depth sensors.
- Head to head: Kimi-K2.7-Code vs grok-4.3
This one is as close as the aggregate score suggests, but the task sheet tells a more useful story: Kimi-K2.7-Code was the steadier model across the set, while grok-4.3 won fewer categories and mostly on narrower formatting-faithfulness calls. The statistical edge belongs to Kimi-K2.7-Code, and the reasons are concrete
- Head to head: Kimi-K2.7-Code vs mistral-medium-3-5
This matchup turned on a basic but decisive failure: Kimi-K2.7-Code didn’t get its scene on screen, while mistral-medium-3-5 did. With only one scored task, the edge is still a lean rather than a rout, but the result is straightforward.
- Stan's 14-day AI build turned customer knowledge into $3 million ARR
John Hu says Stanley began with manual interviews, public posts and AI-written cold emails before expanding from LinkedIn to Instagram.
- GPT-5.6, OpenAI's flagship model, helps build itself
Shruti Gandhi's July 9th post spotlights a deeper shift behind Sol, Terra and Luna: the frontier model is becoming production infrastructure.