The week across the frontier labs — new defaults, previews and open-weight releases — plus where the coding benchmarks stand.
Grok 4.5 — SpaceXAI launched Grok 4.5 on 8 Jul, built on the 1.5-trillion-parameter V9 foundation and trained with real developer-session data from Cursor (acquired for ~$60 bn in June 2026). Priced at $2/M input and $6/M output tokens — roughly 2.5× cheaper than Anthropic Opus 4.8 — and claims ~4.2× better token efficiency on SWE-Bench Pro. xAI's own benchmark chart shows Grok 4.5 beats Opus 4.8 on DeepSWE 1.0 and Terminal-Bench 2.1 but trails on DeepSWE 1.1 and SWE-Bench Pro; EU availability expected mid-July. · 8 Jul
Read → x.ai · TechCrunch · AxiosGPT-5.6 (Sol / Terra / Luna) — OpenAI previewed GPT-5.6 on 26 Jun under a US-government-gated limited release (~20 partners); on 8 Jul it confirmed global public expansion launches today (9 Jul). Sol is the frontier tier tuned for agentic, biology, and cybersecurity work; Terra offers GPT-5.5-competitive performance at 2× lower cost; Luna is the fastest/cheapest tier. Sol Ultra reaches 91.9% on Terminal-Bench 2.1. OpenAI is also launching GPT-5.6 Sol on Cerebras hardware at up to 750 tokens/second for select customers. · 8 Jul (announce) / 9 Jul (GA)
Read → OpenAI · Axios · Nextgov · 9to5MacClaude Fable 5 (restored) + Claude Sonnet 5 — Anthropic restored Claude Fable 5 and Mythos 5 on 1 Jul 2026, after a 19-day global suspension triggered by a US Commerce Department export-control directive on 12 Jun following an Amazon-discovered jailbreak. A new safety classifier blocks the flagged technique in >99% of cases and reroutes flagged requests to Opus 4.8. Concurrently, Anthropic launched Claude Sonnet 5 (30 Jun/1 Jul) at introductory rates of $2/M input and $10/M output tokens through 31 Aug 2026, targeting agentic coding workloads. · 1 Jul
Read → Anthropic · The Hacker News · AI NewsClaude Fable 5 / Mythos 5 — Claude Fable 5 is Anthropic's most capable publicly available model, with a 1M-token context window and up to 128K output tokens per request. Priced at $10/M input and $50/M output tokens. It shares the same underlying model with Claude Mythos 5, which remains restricted to vetted Project Glasswing partners for defensive cybersecurity use. Both models were originally launched on 9 Jun 2026. · 9 Jun (launch) / 1 Jul (restored)
Read → Anthropic · Claude Platform DocsMuse Image (Muse Spark) — Meta's Muse Image model has rolled out to Instagram, enabling AI-generated content from public-account images unless users opt out. The model targets advertising, creator, and decorating use cases. CEO Zuckerberg offered no firm timeline for new frontier AI models to follow Muse Spark. · ~7 Jul
Read → Wired via LLM Stats · Tom's HardwareAI leaders and their capital cycle — capex commitments alongside the month’s market moves.
(8 Jul) Nvidia's forward price-to-earnings ratio has fallen to 22.22×, the lowest since 2019, despite record revenues of $215.9 billion — reflecting a broader rotation in AI investor positioning. Micron, Intel, and AMD added roughly $2 trillion in combined market cap in Q2 2026 as investors broadened AI chip exposure beyond Nvidia; Nvidia itself managed only a 5.6% YTD gain vs the S&P 500's 9.6%.
Read full article → Yahoo Finance(~3 Jun) Google, Amazon, Microsoft, and Meta collectively plan $725 billion in capital expenditures in 2026, up 77% from ~$410 billion in 2025, with the majority directed at AI data centres and Nvidia hardware. Goldman Sachs now projects $5.3 trillion in combined capex from the four largest hyperscalers from FY2025 to FY2030.
Read full article → Yahoo Finance · Tom's Hardware(8 Jul) SpaceXAI confirmed Grok 4.5 was trained on Nvidia GB300 (Blackwell) GPUs at the Colossus computing cluster in Memphis — the same compute infrastructure xAI is also leasing to Anthropic and Google, creating potential capacity tension as xAI's own model demand grows.
Read full article → x.ai · TechCrunchAnnounced Sep 2025, the OpenAI-Nvidia agreement targets deployment of 10 gigawatts of Nvidia systems; Nvidia committed up to $100 billion in investment as phases go live, starting with 1 GW on the Vera Rubin platform in H2 2026. OpenAI will purchase Nvidia chips using those funds; Nvidia's initial $10 bn deploys when the first gigawatt comes online. As of Feb 2026, no formal contract had been signed.
Read full article → NVIDIA Newsroom · AI News · CNBCUK-based Nscale contracted approximately 200,000 Nvidia GB300 GPUs with Microsoft across Texas, Portugal, the UK, and Norway. In March 2026 it raised a $2 billion Series C at a $14.6 billion valuation, led by Aker ASA and 8090 Industries, with Nvidia and Dell participating. Nscale is also a named partner on the Stargate Norway project with OpenAI, and in July was reported to have secured a $900 million debt facility.
Read full article → Spheron Blog / Nscale press release · Bloomberg · TechStartupsTogether AI closed an $800 million Series C on 1 Jul 2026, led by Aramco Ventures, with Nvidia, Vista Equity Partners, General Catalyst, Salesforce Ventures, and others participating. Annual bookings crossed $1.15 billion in Q2 2026. The company plans to grow its open-source inference cloud capacity by 50× over five years.
Read full article → TechStartups · Crunchbase · StartupHub.aiGlobal venture funding hit a record $510 billion in H1 2026, surpassing all of 2025's $440 billion. OpenAI and Anthropic alone accounted for $217 billion — 43% of all startup funding in the half. Q2 2026 was the second-largest quarter on record at $205 billion. SpaceX went public at a $1.77 trillion valuation raising $75 billion, and confirmed a $60 billion all-stock acquisition of Anysphere (Cursor).
Read full article → CrunchbaseGerman drone-intelligence company Quantum Systems raised $1.2 billion on 3 Jul 2026, doubling its valuation to $8 billion. The round marks one of the largest European defence-AI financings on record and signals that AI-enabled autonomous systems are attracting hyperscaler-adjacent capital sizes outside the US.
Read full article → StartupHub.ai (citing Bloomberg video coverage)Zeroth raised approximately $73.6 million in Series A funding for humanoid robotics, announced around 6 Jul 2026 and led by Ant Group. The deal reflects accelerating capital concentration at the intersection of physical AI and robotics, with multiple rounds closing the same week from Yingzhi XBOT ($56M Series B) and Luxonis ($14M Series A for AI perception hardware).
Read full article → TechStartupsThe Stargate joint venture (OpenAI / SoftBank / Oracle) targets $500 billion in US AI data-centre infrastructure through 2029. Across the major AI builders, capex is growing ~70% per year while operating cash flow grows ~23%; Epoch's model shows aggregate free cash flow crossing zero around Q3 2026. Oracle's FY2026 free cash flow was already negative $23.69 billion. Alphabet raised $84.75 billion in a June equity offering to fund its capex programme.
Read full article → 247WallSt · ValueAddVCAbu Dhabi's MGX closed Fund I at $49 billion, exceeding its $45 billion target, with investments in 14 companies spanning semiconductors, AI infrastructure, and AI platforms. MGX is developing Europe's largest AI campus near Paris with 3 GW compute capacity and acquired Aligned Data Centres as part of a $40 billion consortium deal.
Read full article → Crescendo.aiNot verified at send time: GPT-5.6 Sol/Terra/Luna independent SWE-bench Verified scores not yet published on official leaderboard — only Terminal-Bench 2.1 (Sol Ultra 91.9%) confirmed via OpenAI system card · Grok 4.5 score on SWE-bench Verified not yet independently verified — xAI published SWE-bench Pro (64.7%) and DeepSWE figures only · Nvidia NVDA exact market cap as of market close 9 Jul 2026 not yet available — last confirmed figure ~$4.72–4.74tn (6–7 Jul) · Precise Nvidia/OpenAI $100bn deal contract status as of July 9 — last confirmed report (Feb 2026) noted no contract had been signed · Microsoft FY2026 capex guidance figure ($120bn used here) is an analyst estimate — Microsoft confirmed capacity-constrained but did not issue a specific revised FY26 number · DeepSeek and Kimi/Moonshot AI: no new model release verified in the 7–9 Jul window; most recent tracked releases (DeepSeek V4 Pro Max, Kimi K2.6) date to April 2026 · Mistral: no new model release verified in the 7–9 Jul window · Google Gemini: Gemini 3.5 Flash referenced on third-party trackers but no official Google announcement URL verified in this session
