Out July 6 on the Realtime API. OpenAI reports at least 25 percent lower p95 latency versus Realtime v2 (per OpenAI's community post), with configurable reasoning effort (low, medium, high) that runs chain-of-thought inside the audio connection with no separate text-model hop. The mini variant drops reasoning for high-frequency short turns. If you build voice agents, this removes the second hop that was adding lag.

The model behind "Owl Alpha," which led OpenRouter's coding leaderboard anonymously for two months. Trained on 50,000 domestic ASICs with zero Nvidia GPUs. Vendor benchmarks: 59.5 pass@1 on SWE-bench Pro, 1M-token context at $0.30/$1.20 per million tokens (per VentureBeat). Weights "coming soon" on Hugging Face. A near-frontier agentic coder off Nvidia silicon changes the supply-chain math.

Google internally confirmed July 17 after missing a June target, a June 30 date, and entering July in Vertex AI enterprise preview only. Delays traced to heavy token consumption on agentic tasks, coding below Gemini 3.5 Flash, and long-horizon reasoning under Google's own I/O bar. Sonnet 5, Fable 5, and GPT-5.6 all shipped in the gap. Pro lands into a crowded field.

Previewed June 26 to roughly 20 vetted government partners, still no public access on July 7. The commercially interesting tier is Terra, which OpenAI positions as GPT-5.5-competitive at half the cost. Altman's "coming weeks" language from June 26 expires this week; market prediction markets place July 10-17 as the most probable window.

Arthur Mensch told TechCrunch on July 4 that an open-weight model is coming this summer with early access in July. No name, params, or license yet. With ARR past $400M and a EUR 1.7B Series C closed, this is Mistral's answer to Llama 5 and LongCat-2.0 in the open-weight frontier.

Ticker SKHY targets up to $29.4 billion (per US News), surpassing Saudi Aramco's $25.6B 2019 raise. Books run July 6-9, pricing July 9, trading July 10. SK Hynix holds roughly 60 percent of the HBM market with Nvidia, Google, and Microsoft as primary customers; three investors have already signaled $7B in combined interest. All proceeds go to AI memory capacity. The memory that trains your models is now a public-market bet.

The White House framework (FT-reported, expected as soon as this week) would set benchmark thresholds, 30-day pre-release government review windows, and domestic-versus-international access rules (per Towards AI citing the FT). Meanwhile the UN's inaugural two-day AI Governance Dialogue closed in Geneva with 169 countries and no binding framework. Its scientific panel, co-chaired by Yoshua Bengio and Maria Ressa, found capability advancing faster than any government can regulate. For operators: the near-term rule set will be US-domestic, not multilateral.

Claude Science is an AI workbench for literature review, hypothesis generation, and data analysis, now in beta for all paid Anthropic subscribers. Institutional grant applications, which unlock full access and higher rate limits, close July 15. If you run research workflows, the window to the free-tier is open now.

Per Cursor's pricing page, Standard runs $32/mo annual, Premium $96/mo annual. Picking wrong either overpays or throttles your power users. We also cover both CVSS 9.8 DuneSlide RCEs, closed in Cursor 3.0. Try Cursor →

Make when the job is a structured process with branching. Lindy when the job is acting like a person: inbox triage, phone calls, meeting prep. Includes current July 2026 pricing. Try Make →

Make charges a credit multiplier on AI steps, the number that flips the math for AI-heavy teams. The July 2026 decision per team, with the n8n MCP server and SAP deal factored in. Try Make →

WebMCP for cooperative sites in Chrome 149 plus origin trial, Playwright MCP when you need real browser control, Firecrawl for scale. A decision table and two API patterns. Try Firecrawl →

Sonnet 5 leads in-repo editing on SWE-bench Pro (63.2 percent vs 58.6, per SWE-bench Pro) at $2 per million input tokens through Aug 31. GPT-5.5 owns terminal agents. Gemini 3.1 Pro owns front-end. A per-persona verdict with cited pricing, not a leaderboard dump.