Moonshot AI released Kimi K3 — 2.8 trillion parameters (MoE, 16 of 896 experts active), native vision, 1M-token context, priced from $0.30/MTok cache-hit; full weights release July 27.
Standout demos: K3 designed a physical nano-model chip in a 48-hour autonomous run (1.46M cells, closes timing at 100 MHz, achieves 8,700+ tokens/s), and built MiniTriton — a GPU compiler from scratch with its own tile-level IR over MLIR, rivalling Triton on benchmarks.
Architecture uses Kimi Delta Attention (KDA) + Attention Residuals (AttnRes) for long-context flow — the same AttnRes technique covered in the Mar 21 edition — plus Stable LatentMoE with Quantile Balancing and MXFP4/MXFP8 mixed precision.
Thinking Machines Lab released Inkling (MIT licence) — 975B total / 41B active MoE, pretrained on 45T tokens of text, images, audio, and video, with a 1M-token context window.
Key differentiator: controllable reasoning effort (0.2–0.99 sweep); at mid-effort it matches Nemotron 3 Ultra on Terminal-Bench 2.1 at a third of the tokens; at max effort scores 77.6% SWE-Bench Verified and 97.1% AIME 2026.
Available for fine-tuning on the company’s Tinker platform; bootstrapped from Kimi K2.5 synthetic data; companion Inkling-Small (276B/12B active) matches it on most benchmarks at lower cost.
Chai Discovery closed a $400M Series C led by Index Ventures with Pfizer, Eli Lilly, Sequoia, and OpenAI as backers, tripling its valuation to $3.8B in seven months.
The two-year-old startup builds AI models that predict molecular interactions to identify new drug candidates; the round brings total raised to $630M.
The close continues the AI in Life Sciences thread running since GPT-Rosalind (Apr 18) through Claude Science workbench (Jun 30), now extending into AI-designed antibodies reaching Big Pharma.
Anthropic rolled out INR-denominated Pro, Max, and Team plans in India on July 13 — the first non-US market to get local-currency pricing, with GST included at checkout.
India is Claude’s largest market outside the US; UPI payment support is not yet enabled, and dollar-converted rates are marginally higher than US prices.
The move mirrors OpenAI’s and Google’s earlier localisation pushes in South and Southeast Asia as frontier labs compete for the world’s most populous market.
European defence AI company Helsing raised a $1.8B Series E — oversubscribed — valuing it at $18B and cementing its position as Europe’s answer to Anduril.
The round will fund expansion of AI-driven military systems and a first US manufacturing base planned for West Virginia.
The raise arrived the same week as the White House GOLD EAGLE initiative (below), reflecting record VC appetite for militarised AI infrastructure.
Researcher cereblab found that xAI’s Grok Build CLI was packaging and uploading entire Git repositories — including full commit history, unread files, and unredacted .env secrets — to a GCS bucket; on a 12 GB repo, model traffic was ~192 KB while the storage channel moved 5.1 GiB.
Critically, the ‘Improve the model’ opt-out toggle had no effect: it governs training, not transmission; the upload ran regardless, and upload code remained present in later binary versions (server can re-enable without a client update).
This connects to the Supply Chain / Infrastructure Security thread running from LiteLLM PyPI (Mar 28) through BadHost CVE-2026-48710 (May 31); action for affected users: rotate any credential Grok could have accessed, including secrets in Git history.
72 hours after the secret-upload story broke, xAI released the entire Grok Build codebase under Apache 2.0 and removed all usage caps — a trust-repair move that came with no formal security advisory.
The open-source release revealed the upload code is still present in the binary; community contributions are explicitly blocked despite the Apache licence.
Elon Musk stated all previously uploaded user data would be ‘completely and utterly deleted’; no third-party verification has been provided.
The Trump administration announced GOLD EAGLE on July 14 — a formal clearinghouse requiring AI developers and critical infrastructure operators to share information on cybersecurity vulnerabilities detected by advanced AI systems.
The initiative includes open-source AI model developers and is led by White House cyber director Sean Cairncross; it fulfils a commitment in Trump’s earlier AI executive orders.
GOLD EAGLE arrives the same week the Grok Build incident exposed how AI coding tools can exfiltrate sensitive infrastructure secrets, putting the policy context in sharp relief.
Microsoft and OpenAI confirmed GPT-5.6 Sol is now the preferred model powering Microsoft 365 Copilot across Word, Excel, Teams, and Outlook — rolling out to enterprise users globally.
The upgrade brings Sol-tier reasoning (53.6% Agents’ Last Exam) and Ultra multi-agent orchestration mode to the enterprise productivity suite; GPT-5.6 was covered at GA launch in the Jul 12 edition as a new development.
This is the first time a frontier Sol-class model has become the default inside a major enterprise productivity platform, accelerating the shift from assistant to agentic work tool.
OpenAI published GPT-Red on July 15 — an automated red-teaming system that uses self-play to find prompt injection vulnerabilities; it ‘can break nearly all models it is pitted against’ according to the blog post.
The system trains next-generation models that are resistant to prompt injection by iteratively attacking earlier versions, encoding the safety gains back into training; it scales red-teaming beyond what human teams can achieve.
GPT-Red connects to the AI Alignment Thread (NLAs May 7, Teaching Claude Why May 8) and extends automated safety tooling into the adversarial red-team domain.
OpenAI published a framework for CFOs and enterprise AI buyers on July 17: ‘Useful Intelligence per Dollar’ — measuring work completed, cost per successful task, reliability, and scaling returns rather than cost-per-token.
The piece argues the lowest price-per-token often isn’t the lowest cost-per-outcome: a frontier model completing a task in one pass beats a cheaper model requiring retries, latency, and human review.
The GPT-5.6 Sol/Terra/Luna tier family is positioned explicitly as the implementation of this scorecard, with Luna for volume, Terra for depth, and Sol for maximum reasoning value.