Four days before this daily, Moonshot AI shipped Kimi K3. 2.8 trillion parameters. Artificial Analysis scored it at 57 on its intelligence benchmark, against Grok 4.5's 54. On July 18 to 20, Elon Musk confirmed via a social media reply chain that Grok 4.6 is in development at 2 trillion parameters and will complete initial pre-training the week of July 21.

The mechanism: Grok 4.5, xAI's current flagship, uses a 1.5-trillion-parameter architecture internally called V9. Grok 4.6 at 2T is a 33-percent scale increase. The framing from xAI isn't "biggest model," since Kimi K3 is larger at 2.8T. It's efficiency per parameter. Grok 4.5 costs $0.31 per task against Kimi K3's $0.94. If Grok 4.6 closes the intelligence gap while staying near that cost neighborhood, that's the actual competitive claim. Whether the claim holds depends entirely on benchmarks that don't exist yet.

What doesn't exist as of July 20: no benchmarks, no pricing, no context window specification, no formal launch page, no release date. Community timeline estimates put public rollout four to six weeks after training completes, which puts the earliest window in late August. A social confirmation that a model is in training is not an announcement of anything a builder can integrate.

The blast radius for builders: if you were about to standardize on Kimi K3 or GPT-5.6 Sol for a new agent stack requiring frontier-tier intelligence, the comparison table is about to update. That's not a reason to wait indefinitely, but it is a reason to price in a re-evaluation once Grok 4.6 publishes actual benchmarks. The labs that announce training completions weeks ahead of launch have made "pre-announcement" a normal part of the calendar. The frontier now runs on two timelines: what shipped and what was announced. Both matter. Only one is testable.

Google DeepMind scrapped the Gemini 2.5 Pro base model entirely this summer after engineers found structural failures in recursive tool-calling and SVG generation. The ground-up rebuild targeted June, then July, then the specific date of July 17. That date passed without a launch. As of July 20, there is no gemini-3.5-pro in the public Gemini API model list and no officially rescheduled date from Google.

What the rebuild was supposed to produce: a 2-million-token context window, a Deep Think reasoning layer for complex problem-solving, and autonomous workflow capabilities. Rumored pricing around $15 per million input tokens. The target benchmarks were GPT-5.6 and Anthropic's Fable 5. Everything remains unconfirmed because the model isn't public.

The context that makes the delay sharper: in the weeks leading up to the missed July 17 date, four top DeepMind researchers left for competitors. Among them: Noam Shazeer, the Transformer co-inventor, and John Jumper, the Nobel laureate who built AlphaFold. Both reportedly moved to OpenAI and Anthropic. A lab rebuilding its flagship model from scratch after losing two of its most famous researchers is doing harder work than the benchmark will eventually show. The model that ships will be the product of a rebuild completed without the people who might have designed the rebuild differently.

The competitive window: GPT-5.6 Sol is actively rolling out to paid OpenAI plans. Kimi K3 is available. Grok 4.6 is weeks away. Every week Gemini 3.5 Pro remains unlaunched is a week when enterprise deals close elsewhere, against models that can actually be tested.

Anthropic opened an application round for rare genetic disease researchers: up to $50,000 in Claude API credits over six months. Application is at anthropic.com/news/rare-disease-research-grants.

The target audience is research groups that have the domain expertise to use Claude productively but not the budget for API calls at research scale. Rare genetic diseases are structurally under-studied: patient populations are small, commercial incentive is weak, and manual annotation is expensive. API credits at that volume translate to enough tokens for real research workloads: patient record analysis, literature synthesis, variant interpretation support, annotation at scale. The grant is credits, not cash, which means the money goes directly into compute rather than overhead.

This is part of Anthropic's AI for Science effort, which has included comparable grant windows for climate research, genomics, and public health. The program is easy to skip in a release log heavy with SDK patches and agent fixes. It is, however, the most direct expression of what Anthropic says it's building for. If the stated mission is beneficial AI, the grants program is a line item where the claim meets a verifiable action.

Both the main TypeScript SDK and the AWS wrapper dropped at 8:46 PM UTC on July 20 with the same core fix: AWS options and auth mode were not preserved across withOptions() calls. If your code used client.withOptions() to create a scoped client and relied on AWS authentication mode for subsequent requests, the options were silently dropped and you'd see auth failures that weren't straightforward to trace back to the withOptions boundary.

The blast radius: any TypeScript code using the Bedrock SDK with withOptions() for session-scoped configuration. The failure mode was silent credential loss, which tends to look like a permissions problem before it looks like an SDK bug. The aws-sdk package additionally removed a duplicate loggerFor import that was breaking downstream builds in strict module environments.