Blog

OpenAI shipped its GPT-5.6 model family — and picking an AI model now looks like picking an infrastructure tier. The days of one “best” model are over: OpenAI now sells three durable capability tiers — Sol (flagship, $5/$30 per million input/output tokens), Terra (balanced, $2.50/$15), and Luna (cheapest, $1/$6) — plus an “ultra” setting that coordinates four-plus agents in parallel on hard tasks, which means anyone buying AI for a business now has to think in portfolios, price-performance curves, and workload matching rather than a single vendor checkbox. GPT-5.6 went to general availability July 9 after a limited preview, claiming state-of-the-art results in coding and knowledge work (including presentations and financial models); a day earlier OpenAI released GPT-Live, a full-duplex voice model that listens and speaks simultaneously, replacing ChatGPT’s old turn-taking voice mode.

Apple sued OpenAI for trade secret theft — the AI talent war just became a courtroom war. This is the case to watch if you care about how companies protect IP when employees move between rivals at unprecedented pace: Apple alleges OpenAI “deliberately and systematically” solicited confidential information from Apple employees as it builds its first consumer hardware device, noting that more than 400 former Apple staff now work at OpenAI — including former iPhone design VP Tang Tan, now OpenAI’s chief hardware officer. The July 10 complaint alleges one engineer retained access to Apple’s cloud storage after joining OpenAI and downloaded dozens of confidential files; OpenAI says it has “no interest in other companies’ trade secrets.” The awkward twist: ChatGPT is still integrated into the iPhone under the companies’ 2024 partnership.

OpenAI says its model proved a 50-year-old math conjecture in under an hour — mathematicians say it’s real, but less magical than it sounds. If AI can close out problems that stumped humans for decades, the implications for R&D-heavy industries (pharma, materials, quant finance) are enormous — which is why the caveats matter as much as the claim. OpenAI reported that GPT-5.6 Sol Ultra, running 64 parallel subagents, produced a complete proof of the Cycle Double Cover Conjecture, open since the 1970s. Mathematician Thomas Bloom of the University of Manchester calls it “a very nice proof” but notes it is short, elementary, and “could have been discovered in the 1980s,” criticizes the paper for not citing the 1983 prior work its key ideas build on, and full verification by the mathematical community is still pending — so treat this as machine persistence at scale, not yet proven machine creativity.

Microsoft cut 4,800 more jobs as “AI-cited” layoffs keep piling up — but read the fine print on causation. For anyone recruiting into tech this fall, the pattern matters more than any single announcement: companies posting record revenues keep shrinking headcount while invoking AI, with roughly 120,000 tech roles cut so far in 2026 per Layoffs.fyi and AI the most-cited layoff reason in May’s record month, per Challenger, Gray & Christmas. Microsoft’s July 6 announcement eliminated about 2.1% of its workforce — though the company explicitly said the roles are “not being replaced by AI,” only that AI is “changing how work gets done.” That gap between the headline and the stated cause is the honest takeaway: AI is the era’s all-purpose restructuring rationale, and some analysts argue it’s partly cover for unwinding pandemic-era overhiring.

SK Hynix raised $26.5 billion on Nasdaq — the largest U.S. share sale ever by a foreign company — and jumped 17% on day one. The AI trade’s clearest winners right now are the picks-and-shovels suppliers: SK Hynix controls 56% of the global high-bandwidth memory market that Nvidia’s AI chips depend on, and investors oversubscribed its offering more than seven times. The South Korean chipmaker listed 177.9 million ADRs at $149 (ticker SKHYV) on July 10, pushing its market value past $1 trillion; proceeds fund fab expansion, including a $4 billion advanced-packaging plant in Indiana. For finance-bound students, it’s also a live case in where mega-cap listings are happening: this is the third-largest debut ever, behind Saudi Aramco and last month’s SpaceX IPO.

Meta started charging for its frontier models — a strategic reversal aimed straight at Anthropic and OpenAI. Meta built its AI reputation on giving models away; on July 9 it launched Muse Spark 1.1 alongside a public preview of the Meta Model API, the first time developers can pay to build on its frontier models — a bet that the agentic coding market is too lucrative to subsidize. The model is a multimodal agent with a 1-million-token context window, pitched at coding and computer-use workloads, with Replit, Cline, and Box among early partners. Watch whether a fourth serious paid competitor compresses API pricing across the board — good news for anyone building on these models, harder news for the labs’ margins.