The week across the frontier labs — new defaults, previews and open-weight releases — plus where the coding benchmarks stand.
Claude Fable 5 — Claude Fable 5 returned to global availability on 1 Jul after a 19-day US export-control suspension (12 Jun–1 Jul). As of today, 7 Jul, the six-day subscription-included window expires and all Pro/Max/Team/Enterprise access shifts to metered usage credits at $10/M input and $50/M output tokens — Anthropic's highest published rate. A retrained safety classifier now blocks the reported Amazon jailbreak in over 99% of cases; Mythos 5 remains restricted to approved US organizations via Project Glasswing. · 1 Jul / billing shift 7 Jul
Claude Sonnet 5 — Launched 30 Jun as the new default for Free and Pro plans, positioned as 'the most agentic Sonnet yet' with near-Opus-4.8 performance on agentic tasks. Priced at $2/M input and $10/M output through 31 Aug 2026 (introductory), rising to $3/$15 standard. Supports 1M token context and 128K max output; ships with real-time cybersecurity safeguards enabled by default. · 30 Jun
GPT-5.6 Sol / Terra / Luna — Launched 25–26 Jun as a three-tier family (Sol: flagship; Terra: balanced at ~GPT-5.5 perf for 2× lower cost; Luna: fastest/cheapest) in a limited preview restricted at US government request to ~20 trusted-partner organizations via API and Codex — not yet in ChatGPT. Pricing: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per 1M tokens. Sol introduces 'max reasoning' and 'ultra' subagent mode; Cerebras-hosted Sol targeting 750 tokens/sec planned for July. Broad GA expected 'in the coming weeks'; all three tiers rated High capability in Cybersecurity and Biological/Chemical risk under OpenAI's Preparedness Framework. · 26 Jun (preview ongoing)
gpt-realtime-2.1 / gpt-realtime-2.1-mini — Released as updated low-latency voice and multimodal models, delivering at least 25% lower p95 latency via improved caching. gpt-realtime-2.1 adds improved alphanumeric recognition, silence/noise handling, and interruption behavior, with support for speech-to-speech interactions with configurable reasoning effort and tool use. · ~7 Jul
Claude Fable 5 / Mythos 5 — Claude Fable 5 is Anthropic's most capable widely released model, built for demanding reasoning and long-horizon agentic work; scored 80.3% on SWE-bench Pro and 95% on SWE-bench Verified at launch. Claude Mythos 5 shares the same capability class but ships without the broad safety classifiers and is available only in limited release through Project Glasswing. Both carry 1M token context and 128K max output at $10/$50 per 1M tokens. · 9 Jun (returned 1 Jul)
Gemini 3.5 Flash — Launched in May 2026 and positioned as a fast, agentic production model rather than a chatbot; described as particularly strong for tool-calling and agentic workflows. Gemini 3.5 Pro cleared for July general-availability launch after slipping from June. · May 2026 (3.5 Flash); 3.5 Pro GA pending Jul
Kimi K2.7 Code / K2.7 Code Highspeed — Moonshot AI released Kimi K2.7 Code and K2.7 Code Highspeed on 12 Jun 2026, positioned for coding tasks. Kimi K2.7 was cited by Anthropic's own testing as capable of reproducing the same exploit that triggered the initial Fable 5 export-control ban, underscoring the competitive coding capability of the model. · 12 Jun
Grok 4.5 — Grok 4.5, described as xAI's 1.5T V9 model, entered private beta at SpaceX and Tesla on 28 Jun 2026. Based on xAI's monthly release cadence, a July public launch is considered the most likely window, though no public date has been set. · 28 Jun (private beta)
AI leaders and their capital cycle — capex commitments alongside the month’s market moves.
The four largest hyperscalers — Amazon, Microsoft, Alphabet, and Meta — plan to spend ~$725bn combined on capex in 2026, up ~77% from ~$410bn in 2025, with ~75% directed at AI infrastructure: GPU servers, data centers, and networking.
ValueAddVC · Fortune · TechCrunch · ReutersLow-power AI processor startup Syntiant filed for a US IPO, reporting $64.5M in Q1 revenue and a $20.9M net loss for the three months ended 31 Mar 2026.
ReutersNvidia authorized an $80bn share repurchase program and reported $50.3bn in operating cash flow, but its next-generation Kyber NVL144 rack-scale system faces delays pushing timelines to 2028, pressuring near-term supplier stocks.
Yahoo FinanceAs of 2 Jul, OpenAI is in preliminary discussions to grant the US government a ~5% stake worth ~$42.6bn at the $852bn post-money valuation, structured around a vehicle modeled on Alaska's Permanent Fund.
SacraOn 31 Mar 2026, OpenAI closed the largest private funding round in history at $122bn, pushing its post-money valuation to $852bn. Amazon anchored with ~$50bn ($15bn upfront, $35bn contingent on IPO by 2028 or AGI), Nvidia committed $30bn (largely in GPU compute), and SoftBank pledged $30bn in three equal tranches (Apr, Jul, Oct 2026). OpenAI reports ~$25bn annualized revenue and 900M+ weekly active users; it is not projected to be cash-flow positive until 2030.
OpenAI · Bloomberg · Forbes · CNBC · SacraAnthropic raised a $30bn Series G led by Singapore's GIC and Coatue Management, with participation from D.E. Shaw, Founders Fund, Microsoft, Nvidia, and the Qatar Investment Authority, pushing its valuation to ~$380bn. Anthropic also signed a deal with California making it a preferred AI partner for state government.
Tech-Insider · AIToolsRecapxAI closed a $20bn Series E led by Andreessen Horowitz alongside 8VC, Lightspeed, and Shield Capital, valuing the company at approximately $250bn inclusive of its SpaceX acquisition. Grok 5 remains in training on the 1.5GW Colossus 2 cluster with a Q3 2026 release probability near 3% per Polymarket.
Tech-Insider · AIToolsRecapOpenAI's Stargate joint venture with SoftBank and Oracle has expanded to nearly 7 GW of capacity across five-plus US data center sites, targeting 10 GW and $500bn in total investment. The flagship Abilene, TX site (co-owned by Crusoe and Oracle) is partially operational running early GPT-5 training and inference on Nvidia GB200 GPUs; final building at Abilene expected by end of 2026.
Sacra · TechCrunchTogether AI closed an $800M Series C on 1 Jul 2026, led by Aramco Ventures with Nvidia, Vista Equity Partners, and General Catalyst participating. Annual bookings crossed $1.15bn in Q2 2026; the company plans to use capital to grow public cloud capacity 50× over five years.
StartupHub.ai · SiliconAngleCrusoe was reported on 2 Jul 2026 to be in talks to raise $3bn in fresh capital, which would triple its valuation to ~$18bn. Crusoe co-owns the flagship Stargate Abilene data center site with Oracle.
StartupHub.aiChinese video AI company Kling AI closed a $2bn funding round backed by General Atlantic at an $18bn valuation, making it one of the largest generative media AI fundraises to date.
StartupHub.aiAbu Dhabi's MGX closed Fund I at $49bn (exceeding its $45bn target), with investments across 14 companies spanning semiconductors, AI infrastructure, and AI platforms. MGX is developing Europe's largest AI campus near Paris targeting 3 GW of compute capacity and acquired Aligned Data Centres as part of a $40bn consortium deal.
Crescendo.aiNot verified at send time: GPT-5.6 Sol/Terra/Luna SWE-bench Verified scores: all benchmarks are vendor-reported only; no independent verification published as of 7 Jul 2026 (model not yet broadly available) · Grok 4.5 public launch date and benchmark scores: private beta only as of 7 Jul, no published figures · Gemini 3.5 Pro official benchmark scores and GA date: cleared for July launch but not confirmed shipped as of 7 Jul · Claude Sonnet 5 SWE-bench Verified exact score: BenchLM lists it in tier below Opus 4.8 but does not publish a precise number · Claude Opus 4.7 Adaptive SWE-bench Verified exact score: tier listed but precise figure not confirmed in sources checked · Nvidia current market cap as of 7 Jul 2026 close: last confirmed figure is $4.85tn as of 1 Jul (Statista); intra-week moves unverified · Meta current market cap as of 7 Jul 2026: $1.92tn figure is approximate and predates the 2 Jul -4.9% selloff; no post-selloff close confirmed · DeepSeek V4 Pro and Mistral Large/Medium 3.5 benchmark updates for July 2026: no material new releases or scores confirmed in sources checked · OpenAI-Nvidia 10 GW / $100bn direct compute deal separate from Stargate: referenced in prior coverage but specific deal structure not independently confirmed in today's sources · Thrive Holdings $2bn raise status (SoftBank, Altimeter, D1): reported by The Information as sought but not confirmed closed
