No single blockbuster model today, but a clear direction: the GPT-5.6 family is quietly becoming the connective tissue of the week. OpenAI retuned its Sol tier inside ChatGPT and is about to give Luna away free to every user, while Perplexity wired its own Terra and Luna agent tiers into Personal Computer, so the same names keep turning up across platforms and price points. Against that closed-model momentum, DeepMind did something rarer, open-sourcing WeatherNext, a state-of-the-art cyclone-forecasting model, weights and all. Add a gold-level Olympiad showing from Meta and a responsible-AI toolkit from Microsoft, and the day's real story is cheaper, more delegatable AI rather than another jump in raw power.
GPT-5.6 Sol sharpens up and Luna goes free for everyone
If you live in ChatGPT day to day, this is the update that actually touches your workflow. OpenAI has retuned GPT-5.6 Sol so it now powers both the Instant and the deep-reasoning modes for Plus and Pro users, aiming for more factual, focused answers in everyday chats, and it has added a slider that lets you push reasoning effort up or down without hopping between models. The bigger shift is on the free tier, where free and Go users are getting unlimited text chats with GPT-5.6 Luna from tomorrow, removing the message caps that quietly nudged casual users towards rival apps.
For anyone building on the API this confirms the shape of the family, with Sol positioned for reasoning and Luna as the cheap, high-volume workhorse. If you still have automations wired to older GPT-5.5 endpoints, it is worth checking whether the Sol and Luna split changes your cost and latency maths before your next billing cycle.
DeepMind open-sources WeatherNext, its cyclone-forecasting model
This one matters even if you never touch a weather API, because it is a rare case of a frontier lab handing over usable weights for anyone to run. DeepMind published WeatherNext in Nature and released the code and model weights on GitHub across three variants, WeatherNext Cyclones, WeatherNext 2 and WeatherNext 2-mini, so researchers and smaller forecasters can build localised systems without Google in the loop. The headline capability is roughly a full extra day of accurate warning on a storm's track and intensity, which in plain terms means a three-day forecast that matches what older systems only managed at two days.
For developers, an openly licensed, state-of-the-art scientific model is a strong base to fine-tune, and it continues DeepMind's habit of shipping weights rather than papers alone. If you work anywhere near climate, logistics or insurance tooling, this is worth pulling down and poking at today.
Perplexity wires Terra and Luna agents into Personal Computer
Perplexity used its own channel to slot two new agent tiers into its Personal Computer product, and the specifics surfaced there before the mainstream press caught up. Terra is pitched as the heavy-lifting subagent for complex, goal-oriented work, with Perplexity's own WANDR benchmark putting it around eleven points above Sonnet while claiming order-of-magnitude cost reductions on multi-step tasks. Luna sits underneath it for recurring, speed-and-cost-sensitive jobs like summarising and routine automation.
If you already lean on Perplexity for research and light building, this turns it into more of a full agent harness that can orchestrate across your files, connected apps and the web at a lower running cost. Treat the benchmark figures as vendor-supplied for now, but the direction, cheaper agents you can hand longer jobs to, is the pattern worth watching.
Meta's models take gold-level scores across five STEM Olympiads
Meta used its own channel to share that its AI models entered five international STEM Olympiad competitions this year and came away with gold-level, sometimes perfect, theory-exam scores in physics events including the Asian and International Physics Olympiads. This is a reasoning-progress signal rather than something you can go and use, so treat it as a benchmark milestone rather than a launch.
It does, however, set the scene for Meta's product push around the Muse family. Worth noting for context: Meta's Muse Code terminal coding agent and its Muse Spark 1.2 model landed on 5 August, just outside this 24-hour window, so they are not covered in full here.
Microsoft Research ships Atlas, a toolkit for human-centred AI
Microsoft Research released Atlas, a set of frameworks and tools for designing, deploying and evaluating AI systems across different languages, cultures and communities. If you build products that ship internationally, the pitch is that inclusivity gets baked into the design and evaluation stage rather than bolted on after launch.
It is a research toolkit rather than a headline product, so the impact is practical rather than dramatic, but teams doing responsible-AI work now have a concrete set of playbooks to lift from.
Industry themes
The GPT-5.6 family is quietly becoming the connective tissue of the week. OpenAI is retuning Sol and giving Luna away free inside ChatGPT, while Perplexity is wiring its own Terra and Luna tiers into its agent product, so the story is less about one shiny new model and more about the same family fanning out across platforms and price tiers.
The sharper theme underneath is cheap, delegatable agents. Both OpenAI's free-Luna push and Perplexity's Terra and Luna tiers are aimed at making capable models cheap enough to run at high volume or hand a long, multi-step job to. That is the same cost pressure driving the week's model releases, and it keeps rewarding whoever can serve intelligence most cheaply rather than whoever tops a benchmark.
Running against that closed-model momentum, DeepMind open-sourcing WeatherNext alongside a Nature paper is a reminder that frontier labs still see strategic value in releasing usable weights for scientific domains, much as the Asian open-weight releases did earlier in the week. One practical note: the Perplexity Terra and Luna details surfaced on the company's own X channel before mainstream coverage, so the social scan is what put it in front of you today.