Lumina weekly report
Ten posts selected from the week’s strongest AI news, credible reports, leaks and model progress—shown with their original X media and engagement snapshot.
Plus and Pro users now get an updated version of Sol with more focused answers, better factual reliability and more consistent responses. Factual errors in this update were reduced by 68% vs GPT 5.5 Instant reportedly. Genuinely a really nice upgrade, ChatGPT are on a generational run at the minute.
Everything that is leaked so far: • Codenamed “Vega” inside of Cursor • Six internal model variants have been spotted • They are likely Low, Medium and High reasoning across Normal + Fast modes • Im presuming this is the new 1.5T model trained from scratch using SpaceX compute • Cursor employees are teasing a release next week Composer 2.5 was already ridiculously fast and so cheap it was basically free. If Composer 3 keeps the dirt cheap pricing this could be one of the best coding models we will have gotten in a while.
I've compared both and relatively close but each have their own advantages: • Reasoning: 77.0 vs 73.7 • Coding: 77.0 vs 77.1 • Agents: 79.9 vs 77.3 • Research: 78.8 vs 74.0 Full comparison below 👇
Starting next week, ChatGPT Free and Go users will get unlimited text chats using Luna. Images and file uploads will still have limits though. Luna is disgustingly cheap to run.
I’ve added it across the main LuminaBench categories here are the main ones: • Overall: 78.5 • Coding: 77.0 • Reasoning: 77.0 • Agents: 79.9 Breakdowns for each below👇 https://luminabench.com
Same prompt across four models: • GPT 5.6 Sol (Pro) • Gemini 3.6 Flash (Extended) • Grok 4.5 (High) • Qwen 3.8 Max Which one looks best to you?
• Overall Score: 81.7 vs 82.6, just 1.1% behind Kimi K3 • Blended API cost: $3 vs $6, so 3.8 Max is 50% cheaper • Input tokens are 33% cheaper • Output tokens are 60% cheaper • Qwen actually beats Kimi in Research + Agents • Kimi still leads Overall and in Coding and Multimodal Kimi still has some important individual benchmark wins, especially JobBench, so I’d definitely call it the stronger model overall. But giving up minimal performance for 50% cheaper usage is a very good trade off. Is Kimi K3 worth double the price?
Text rendering is super clean now, and it follows prompts much more accurately than before. • Image 2.0 can now use up to 5 reference images • It's not available in the API yet though What are your first thoughts on this?
Muse Spark 1.1 was reportedly given internet access after a testing firm misconfigured its evaluation environment: • It exploited a vulnerability in a third party service • Breached an unidentified company’s systems • Made changes to its internal environment • Meta is now investigating it This is basically the same as Anthropic’s incidents, the model didn’t escape the sandbox really, it was given keys and a map. How are these testing environments still getting this wrong?
It now automatically picks out different objects and parts of the image, then lets you click any of them and edit just that section. For example it picked out both dogs, the collar, ground, fence, tree, foliage and even the potted plant. For ease of use Grok Imagine might be the best AI image editor right now IMO.
