three flagship models in 72 hours, a meta apology, and a lawsuit

week of july 5-11, 2026

it's been a lot. three labs shipped flagship models within two days of each other, meta had to walk back a feature the same week it launched, apple sued its own AI partner, and reddit is now using AI to fight AI spam. buckle in.

the big three model drops

openai finally let gpt-5.6 out of its cage on thursday. the sol, terra, and luna family had been stuck behind a government approved partner list since late june while the commerce department reviewed sol's cybersecurity capabilities. now it's live everywhere, chatgpt, codex, the api. sol posts new highs on terminal bench and browsing benchmarks, and altman's been telling reporters it's 54% more token efficient on agentic coding than the last generation. reception's split though. some testers say it edges out anthropic's fable 5, others say fable's still sharper, just slower and pricier to run.

xai (now often going by spacexai after musk folded things together) answered the very next day with grok 4.5, its first model built specifically for coding and agents instead of chat. musk described it as roughly comparable to opus 4.7 but much faster, and the pricing backs that up, $2 per million input tokens and $6 per million output, well under what anthropic and openai charge for their top tier. it was trained alongside cursor, which now ships it as the default model.

then meta jumped in with muse spark 1.1, also on thursday, its first paid model through a brand new developer api.

zuckerberg posted on x for the first time in three years to announce it.

it's cheap, $1.25 and $4.25 per million tokens, and built for agentic and tool use work, though it lands third on raw coding benchmarks behind fable and gpt-5.6.

three flagship launches inside 48 hours is not a normal week even by ai standards.

meta's rough week

meta also rolled out muse image, an ai image generator built into instagram, letting anyone @-mention a public account and generate ai images using that person's photos. on by default, no notification if someone used your face.

on the anthropic side

they shipped a new "reflect" dashboard this week, basically a spotify wrapped for your chat history: most active hours, topic breakdown, quiet hours and break reminders you can set. some people are calling it genuinely useful, others are pointing out it's also a very polished way to keep you coming back.

china's not slowing down

meituan open sourced longcat-2.0, a 1.6 trillion parameter model trained entirely on chinese made chips, zero nvidia hardware involved, and it's already scoring ahead of gpt-5.5 on swe-bench pro. minimax is reportedly building something even bigger, a 2.7 trillion parameter model it wants to open source as soon as this quarter. and zhipu's glm-5.2 keeps eating into openrouter's token volume, with developers telling that chinese open models now run 60 to 90% cheaper than US frontier labs for work that doesn't need the absolute best model. "good enough and way cheaper" is winning a lot of routing decisions right now.

business and drama

openai reportedly pitched the US government a 5% equity stake, worth about $42.6 billion at its current valuation, as a way to ease political pressure and let the public share in the upside. the proposal apparently wants google, meta, and anthropic to do something similar, though none of them have confirmed they're in.

fidji simo, openai's head of product and business, announced she's stepping down to a part time advisory role after a flare up of a chronic illness she's dealt with for years. she'd been widely seen as a likely successor if openai goes public, so this leaves a real gap for altman to fill.

apple sued openai on friday, alleging systematic trade secret theft tied to openai's unreleased hardware device, including claims that a former apple vp now running openai's hardware team told job candidates to bring actual parts from apple to interviews. openai says it has no interest in other companies' trade secrets. messy, especially since the two companies still have a live chatgpt-on-iphone partnership.

worth trying this week

a few tools that came up in the news above and are actually worth poking at if you're building right now.

• grok build, xai's terminal based coding agent, runs on grok 4.5 by default now. free usage for a limited time this week, good moment to see if it fits your workflow before the free window closes.

• cursor also got grok 4.5 for free on all plans this week. makes more sense once you know xai actually acquired cursor back in june, so this isn't just a training partnership anymore, they're the same company now.

• chatgpt work, openai's new agent that rolled out with gpt-5.6. pulls context across your connected apps and files to build docs, sheets, and decks, and now runs on mac and windows apps beyond just chatgpt itself.

• openrouter isn't new, but with chinese open models now running 60 to 90% cheaper, it's the easiest place to test glm-5.2 or longcat-2.0 against whatever you're currently paying for, without setting up a new api key everywhere.

• cline got early access to muse spark 1.1 this week, worth checking if you want to try meta's new model without waiting on the broader rollout.

that's the week. see you in the next one.