Know
What changedClover brings block-diffusion to coding agents
DiffuSpace says its released Apache-licensed Clover-35B-A3B-Preview uses blockwise diffusion, with 35B total and about 3B active parameters, for coding and tool-using agents. The company reports scores of 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1, plus 680 tokens per second on one GB300; these results are vendor-reported, not independently verified here.
Consider testing it against your agent workload; compare quality and speed with your current model.
Claude Ultra removes one usage cap
Anthropic renamed Max 20x to Ultra, moved existing subscribers without changing the price, and removed the rolling five-hour limit for verified accounts; a weekly limit remains. Separately, reports say changing temperature, top_p or top_k disables text watermarking on Opus 4.6 and Sonnet 4.6 without warning; Anthropic reportedly advises leaving those settings at default.
Check account verification and sampling settings if you rely on Ultra limits or Claude watermarking.
Four providers back a shared AI training opt-out
OpenAI, Google, Anthropic and Microsoft committed to honor a shared robots.txt signal for website operators who want to opt out of foundational-model training across participating providers. The approach draws on work by the IETF AI Preferences working group and could replace separate crawler rules; the captured announcement does not give a common rollout date.
If you manage a site, track implementation details before replacing existing crawler controls.
OpenAI limits Rosalind Research to approved organizations
OpenAI says GPT-6 Rosalind Research is a biology, drug-discovery and translational-medicine reasoning model available only through its trusted-access program. It reportedly uses the same weights as GPT-6 Astra, refuses fewer dual-use biology queries while still refusing bioweapons requests, and requires separate approval for API access.
Research teams should check trusted-access eligibility before planning around the model.
Grok Bot limits draw cost concerns
Users complained that xAI’s always-on Grok Bot agent hit usage limits sooner than expected. Elon Musk attributed its running cost to being “unsustainably high relative to revenue,” according to the captured report, but did not explain what changed. The report does not specify revised limits or say whether throttling is temporary.
Avoid depending on continuous Grok Bot runs until limits and availability are clearer.