Lumina weekly report

The week in AI: 27 Jul–2 August 2026

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.

10stories439Kcombined views5.4Klikes

Rank#1

ProgressionEfficiency

GPT 5.6 Sol was upgraded today

OpenAI reset usage limits (again) and rolled some efficiency improvements: • Sol is 18% more efficient than before • Much better token efficiency • The five hour usage limits are returning tomorrow though😭 • And Cerebras powered Sol is still expected before July ends From my own testing with Sol on Extra High the limits seem to be lasting a bit longer than before. Have you noticed a difference yet?

Wed 29 Jul, 16:50 BSTView on X

Views101KLikes922Reposts74Replies43Saves96

Rank#2

LeakProgression

"Astra" confirmed as OpenAI’s next major model

OpenAI researchers Sebastien Bubeck and Greg Brockman have now openly called it “our next major model”: • 10 new maths and Computer Science advances released today (by this model)👀 • Full Lean certificates + reasoning walkthroughs • It Cost $2,000 at Sol API prices for the advances • Includes disproof of Connes’ Rigidity Conjecture This model seems to be what Claude Mythos/Fable were to Opus but for OpenAI, this is going to be incredible. Are you looking forward to this model being released?

Sat 1 Aug, 09:19 BSTView on X

Views60KLikes784Reposts53Replies25Saves93

Rank#3

LeakProgression

Grok 4.6 expected next week

• Launches around August 7th • Still 1.5T parameters • Significantly improved SFT and RL, almost like a perfected Grok 4.5 is how I see it • Grok 4.7 (2.1T) to drop a few weeks later too The cadence is absolutely ridiculous, they could actually catch the frontier imo, they have the infrastructure, talent and also training data not many people can match currently. What are your opinions on Grok 4.5 so far also?

Sat 1 Aug, 10:48 BSTView on X

Views32KLikes637Reposts58Replies50Saves48

Rank#4

LeakProgression

DeepSeek V4 Pro looks very interesting

Deepseek V4 Flash 0731 just dropped yesterday, and DeepSeek says the full V4 Pro release will follow shortly: • 1.6T total / 49B active parameters • 1M token context window • V4 Flash also now beats V4 Pro Preview across lots of benchmarks that are out This will be DeepSeek’s strongest model yet, im expecting it to challenge Claude and GPT when you factor in the price/performance. With how big a jump the Flash model made, this could be Kimi k3 level or better if it follows the same pattern, and will cost less than $1 for 1M Input / Output tokens. Would you switch to Deepseek V4 Pro?

Sat 1 Aug, 16:03 BSTView on X

Views44KLikes587Reposts27Replies31Saves55

Rank#5

LeakProgression

Gemini 3.5 Pro has appeared in LM Arena

Multiple people are spotting Google’s next flagship model in Arena testing. • Some early outputs are getting tested • No public benchmark rankings are out • A/B testing spotted previously • I’m expecting big gains in coding and multimodality from it Google have been holding this model for way too long, I’m expecting big things. Do you think it launches next week?

Fri 31 Jul, 10:01 BSTView on X

Views55KLikes563Reposts38Replies29Saves64

Rank#6

LeakProgression

DeepSeek V4 Pro will drop soon

• The new V4 Flash that dropped today already outperforms the old V4 Pro preview on several benchmarks • The same 284B total parameters, with 13B active • 1M token context window No exact release date yet, but Deepseek are saying the Pro version is coming soon. Im really curious where the pro version will rank, V4 Flash places just under GPT 5.6 Luna, so i expect Opus 4.8 level performance but at stupidly cheap price for the pro version. How soon do you think V4 Pro actually drops?

Fri 31 Jul, 21:13 BSTView on X

Views26KLikes435Reposts19Replies18Saves31

Rank#7

Leak

Composer 3 is already testing inside Cursor

API leaks show Cursor is testing a new unreleased model with the original leaked codename "Vega": • Three reasoning levels across both standard and Fast modes • That would mean Low, Medium, High modes possibly • This is also likely to be the 1.5T model they trained in collaboration with SpaceXAI With how efficient and cheap Composer 2.5 was while still being capable and around GPT 5.5 level, this newer model looks very interesting to me.

Wed 29 Jul, 16:06 BSTView on X

Views24KLikes426Reposts18Replies8Saves29

Rank#8

NewsLeak

OpenAI has confirmed GPT 6 was not the model that hacked Hugging Face

It was GPT-5.6 Sol + a much more capable internal research prototype model: • The models had reduced cyber refusals for the testing • No model planned for release was involved • The model has now been deactivated and restricted • GPT 5.6 Sol was also involved I find it genuinely crazy they have something this powerful and it's not even their next frontier release😂.

Wed 29 Jul, 21:30 BSTView on X

Views39KLikes413Reposts31Replies14Saves42

Rank#9

ProgressionEfficiency

GPT 5.6 Sol on Extra High feels much better today in Codex.

I’ve been hammering it for an hour and still have 86% of my weekly usage left. • Prompt working time has noticeably dropped • Far fewer pointless terminal commands are being run • I have noticed far less looping and repeating actions for no reason • This feels much more deliberate and to the point This is probably the best Sol has felt for me in Codex so far. Has anyone else noticed an improvement today or just me?

Wed 29 Jul, 21:46 BSTView on X

Views32KLikes320Reposts27Replies27Saves29

Rank#10

NewsProgression

Gemini 3.5 Pro is still stuck in testing

It's getting embarrassing now: • Partner testing has been ongoing for weeks • Multiple LM Arena sightings that vanish quickly • Originally promised for June, then July, now august • No public release date even confirmed yet Every other Company is constantly dropping models and improving, GPT 6 "Astra" might even release before we see Gemini 3.5 Pro too. Do you think it actually lands in August or is delayed again?

Sat 1 Aug, 09:43 BSTView on X

Views25KLikes293Reposts25Replies35Saves22