Big week, so a consolidated rundown for anyone catching up.
OpenAI released the GPT-5.6 family publicly on July 9 after a limited partner preview — Sol (frontier reasoning), Terra (previous-flagship performance at ~2x lower cost), Luna (fast/cheap). They also shipped GPT-Live-1, a full-duplex voice model that handles simultaneous listening/speaking, plus gpt-realtime-2.1 with ~25% lower p95 latency.
xAI launched Grok 4.5 (trained alongside Cursor) at $2/M input and $6/M output, claiming Opus-class performance on coding/legal/finance tasks. Independent evals aren't in yet, so treat the claims accordingly.
Google delayed Gemini 3.5 Pro to July 17 — full architectural rebuild, 2M context. Separately, four senior DeepMind researchers departed in one week (Shazeer to OpenAI; Jumper, Adler, Pritzel to Anthropic), and Alphabet dropped ~$225B in market cap.
Microsoft is merging its Copilot apps into one by August. The notable disclosure: fewer than 4.5% of 450M M365 seats have converted to paid Copilot.
Meta launched Muse Image, its first Superintelligence Labs model — agentic image gen that invokes search/code tools and self-refines. Trains on public Instagram photos by default (opt-out).
Open source: Ollama raised $65M Series B (8.9M monthly devs). Gemma 4 got ~90% faster on Apple Silicon in Ollama via multi-token prediction. And a PSA — DeepSeek retires deepseek-chat and deepseek-reasoner on July 24. One-line migration, but note deepseek-reasoner maps to v4-flash thinking mode, not v4-pro, so heavy reasoning workloads should evaluate v4-pro explicitly rather than trusting the alias.
My take as someone building on top of these APIs: the simultaneous price drops (Terra, Grok 4.5, Sonnet 5's intro pricing) matter more than any single benchmark. Near-frontier inference costs fell across four vendors in one week, which changes what's economically viable to automate. Meanwhile Microsoft's 4.5% suggests horizontal assistants aren't converting even with unlimited distribution — the demand seems to be for task-specific automation, which matches what I see with SMB clients. And the DeepSeek cutoff is a good reminder to abstract your model layer.
Sources: OpenAI/xAI/Meta blogs, Euronews, Bloomberg, TechCrunch, CNBC, TechTimes coverage this week.
[link] [comments]