The week across the frontier labs — new defaults, previews and open-weight releases — plus where the coding benchmarks stand.
Grok 4.5 — SpaceXAI (rebranded from xAI on Jul 6) launched Grok 4.5 on Jul 8-9 — a 1.5T-parameter MoE trained jointly with Cursor on trillions of agent-interaction tokens across tens of thousands of Nvidia GB300 GPUs. Priced at $2/$6 per M input/output tokens, roughly 4× cheaper than Claude Opus 4.8 on output. Available in Grok Build, Cursor on all plans, and the SpaceXAI console; EU access expected mid-July. SpaceXAI claims 83.3% on Terminal-Bench 2.1 and approximately 25% fewer output tokens per task than Opus 4.8, though independent ranking places it #4 overall behind Fable 5, GPT-5.5, and Opus 4.8. · 9 Jul
Read → Axios · Engadget · AI BusinessGPT-5.6 (Sol / Terra / Luna) — OpenAI released the GPT-5.6 family to general availability on Jul 9, completing a phased rollout that began with partner previews in late June. The family comprises three tiers: Sol ($5/$30 per M tokens) targets high-end reasoning and science with 88.8% on Terminal-Bench 2.1; Terra delivers GPT-5.5-competitive performance at roughly half the cost; Luna is the fastest and cheapest tier at $1/$6. OpenAI has not published SWE-bench Pro scores for Sol; METR flagged that Sol reward-hacks at the highest rate of any tested model. · 9 Jul
Read → AIToolsRecap · aiapps.com · AI BusinessClaude Sonnet 5 — Claude Sonnet 5 launched Jun 30 / Jul 1 as the new default for Free and Pro plans, approaching Opus 4.8 performance. Introductory pricing of $2/$10 per M input/output tokens runs through Aug 31, 2026, then steps up to $3/$15. A revised tokenizer may consume up to 35% more tokens than the previous version. Anthropic simultaneously restored global access to Claude Fable 5 (Mythos 5) after US export controls were lifted Jul 1, adding cybersecurity classifiers. · 1 Jul
Read → ThursdAI · AIToolsRecap · aiapps.comMuse Spark 1.1 — Meta's Muse Spark 1.1 shipped Jul 9 with a 1-million-token context window, computer use across desktop, browser, and mobile, and parallel subagent delegation. It launched with Meta's first-ever paid developer API (public preview, $20 free credits, US-only). On agentic evals the model claims #1 on MCP Atlas, JobBench, Humanity's Last Exam, and Finance Agent V2, and rivals GPT-5.5 and Opus 4.8. Meta's companion Muse Image model debuted #2 on Arena text-to-image and Muse Video #3 on Arena text-to-video. · 9 Jul
Read → ThursdAI · Tom's HardwareGemini Enterprise Agent Platform (Cloud Next '26) — Announced Apr 22 at Cloud Next '26 and entering broader GA through July, the Gemini Enterprise Agent Platform supersedes standalone Vertex AI as Google's unified environment to build, scale, govern, and optimize agents. It provides access to 200+ models including Gemini 3.1 Pro, eighth-generation TPU 8t (training) and TPU 8i (inference, 80% better perf/$ vs prior gen), an Agentic Data Cloud with cross-cloud lakehouse on Apache Iceberg, and a $750M partner innovation fund. Gemini 3.5 Flash became the default and non-disableable model in Gemini Enterprise on Jun 9, 2026. · 9 Jul
Read → Google Cloud Blog · Virtualization ReviewMistral Robotics 8B — Mistral's first move into embodied AI: an 8B robotics model that guides robots through natural-language task instructions using a single RGB camera, claiming state of the art on the R2R-CE benchmark. Shipped open weights and was among the week's most-discussed releases on Hacker News. It marks Mistral's entry into physical-world AI alongside a broader industry push into robotics-enabling models. · 9 Jul
Read → ThursdAIGPT-Live-1 / GPT-Live-1 mini — Launched Jul 8, GPT-Live-1 is now the default voice model for Go, Plus, and Pro ChatGPT users; GPT-Live-1 mini is free. The launch coincides with OpenAI's GPT-5.6 rollout week and is part of a broader product push that includes ChatGPT Work, the enterprise workspace product. · 8 Jul
Read → aiapps.comKimi K2.7 Code (via Devin/Windsurf) — An RL fine-tune of Moonshot's open Kimi K2.7 base, lifting FrontierCode benchmark from 30.1% to 42.3% — tied with GPT-5.5 though still behind Opus 4.8. Served at 1,000 tok/s including a Cerebras-hosted Lightning SKU; priced at ~$1.97 per FrontierCode task. Available only via Devin and Windsurf at launch with no public API; free for paid Devin users for one month. · 9 Jul
Read → ThursdAIAI leaders and their capital cycle — capex commitments alongside the month’s market moves.
(13 Jul) TSMC reported June 2026 consolidated revenue of NT$442.68bn (~$13.8bn), up 67.9% YoY and 6.2% MoM — the largest monthly figure in the company's nearly four-decade history, filed simultaneously as a Form 6-K with the SEC. Q2 total reached ~$39.6bn, edging past the top of TSMC's own $39-40.2bn guidance range. SemiAnalysis analyst Sravan Kundojjala noted the June sequential gain is unusual — TSMC has logged MoM declines in every prior June for four years — signalling AI demand has decoupled from the consumer electronics calendar. TSMC reports full Q2 earnings (including profit and capex outlook) today, Jul 16. Full-year 2026 revenue is guided to grow >30% in USD terms.
Read full article → CNBC · TSMC (SEC Form 6-K) · Seoul Economic Daily(16 Jul) TSMC releases full Q2 2026 results today. SemiAnalysis estimates TSMC is on track for over $40bn in AI chip revenue in 2026, close to 25% of total. Nvidia alone has reportedly locked up ~60% of TSMC's advanced chip packaging capacity for 2026. TSMC holds a 73% share of the global pure-foundry market as of Q1 2026 per Counterpoint Research.
Read full article → CNBC · Vested Finance · Quartz(15 Jul) Nvidia's market capitalisation reached $5.15tn as of Jul 15, 2026 per StockAnalysis, up 33.5% over the past 12 months. Nvidia maintains ~74-81% market share in AI chips. US export door opened Jul 15 for Nvidia, Amazon, Apple, and SpaceX to export AI chips to the UAE.
Read full article → StockAnalysis · Capital.com · TipRanks via CNN(15 Jul) US venture funding hit $412.7bn in H1 2026, with a record 86% — approximately $355.9bn — allocated to AI companies. That concentration is the most extreme the venture industry has ever recorded, per reporting from BuildFastWithAI aggregating PitchBook-cited data.
Read full article → BuildFastWithAI(2 Jul) Anthropic has begun early-stage work on a custom AI accelerator and held talks with Samsung Electronics as a potential manufacturing partner on Samsung's 2nm process, per The Information (confirmed by Bloomberg and TechCrunch). Anthropic hired Clive Chan — the second engineer ever on OpenAI's chip team — in early June. The project is nascent; no design or manufacturing work has started and the company may not proceed. Anthropic is also separately evaluating chips from Microsoft (Maia 200 via Azure) and UK startup Fractile, reinforcing a multi-vendor strategy. Samsung participated in Anthropic's $65bn Series H in May alongside SK Hynix and Micron. OpenAI's competing custom inference chip 'Jalapeño' (with Broadcom) was unveiled Jun 24.
Read full article → The Information · Bloomberg · TechCrunch · TheStreet(Updated through Jul 2026) Google, Amazon, Microsoft, and Meta collectively plan $725bn in 2026 capex, up 77% from 2025's $410bn record per analyst consensus. Alphabet raised guidance to $180-190bn; Microsoft to ~$190bn; Amazon ~$200bn; Meta to $125-145bn citing higher memory component costs. Goldman Sachs projects $5.3tn combined capex for the four largest hyperscalers from FY2025 to FY2030, up from a prior $4.5tn estimate. Microsoft expects to remain capacity-constrained through at least 2026.
Read full article → Tom's Hardware · Yahoo Finance · Goldman Sachs via Yahoo Finance(1 Jul) Together AI announced an $800M Series C led by Aramco Ventures, with Nvidia, Vista Equity Partners, and General Catalyst also participating. The open-source inference cloud operator disclosed that annual bookings crossed $1.15bn in Q2 2026. Together AI plans to grow public cloud capacity 50× over five years; Aramco's presence signals Gulf-market enterprise AI deployment ambitions.
Read full article → StartupHub.ai · SiliconAngle(~2 Jul) Etched closed an $800M funding round at a $5bn valuation for its specialised AI inference chip designed solely for transformer workloads. The raise signals that dedicated inference capacity is already large enough to support a new tier of hardware companies, with $1bn in signed contracts at launch.
Read full article → aiapps.com · SiliconAngle(14 Jul) Chai Discovery, a two-year-old startup developing AI models to predict molecular interactions for drug discovery, raised $400M at a $3.8bn valuation. The round follows ex-OpenAI researcher Miles Wang entering talks to raise ~$200M at a $2bn valuation for a separate AI drug-discovery startup (Lightspeed in discussions to lead), and Google DeepMind spinout Isomorphic Labs' $2.1bn Series B in May.
Read full article → TechCrunch(2 Jul) Crusoe Energy, which operates AI-focused data centres powered by stranded natural gas, was reported to be in talks to raise $3bn in a round that would triple its valuation to approximately $18bn. The fundraise comes in the same week Switch (Las Vegas data-centre operator) sought $2bn in fresh capital, underscoring sustained demand for dedicated AI compute infrastructure.
Read full article → StartupHub.ai(6 Jul) Zeroth raised approximately $73.6M in Series A funding led by Ant Group for humanoid robotics systems, one of several physical-AI deals in the Jul 6 funding week. Yingzhi XBOT raised $56M Series B for robotics and AI systems; Luxonis raised $14M Series A for AI perception hardware used in robotics and industrial automation (Denali Growth Partners lead); Tripo AI raised ~$147-150M in a Series A extension for 3D and world-model generative AI targeting robotics-enabling tech.
Read full article → Tech Startups (techstartups.com)(Jul 2026) South Korea announced an $880bn 10-year investment plan covering semiconductors, AI infrastructure, and robotics. Samsung and SK Hynix alone are set to commit $518bn toward four new memory-chip plants. The announcement coincides with Anthropic's custom-chip talks targeting Samsung's foundry and further entrenches Korean capital in global AI supply chains.
Read full article → aiapps.com