Cloudflare released two decision models it trained itself, Clef and Clef-flash, both fully open-sourced on Hugging Face under an Apache 2.0 license. On its own numbers, median latency across the benchmark set is 209.3 ms for Clef and 38.8 ms for Clef-flash, against 524.1 ms for Jev. Clef is fully API-compatible with Jev, so Cloudflare says the swap is extremely easy. Post-training pairs label-smoothed cross-entropy for valid schema outputs with a Brier loss, which is how the models are pushed toward calibrated probabilities rather than raw labels.
If you already call Jev for a routing or classification step, the compatible API means you can point an existing integration at Clef without rewriting call sites, then time it yourself, since every figure here comes from the vendor. One thing to plan around: custom decision models are not self-serve yet, so you book time with Cloudflare's forward-deployed engineer team first.
Anthropic's IPO filing says Broadcom agreed to lend it up to $42 billion to finance infrastructure spending. The convertible note could cover about a third of the $125.2 billion Anthropic has committed for a five-year lease of TPU computing capacity. Broadcom could designate a financing partner, and the debt instruments could convert into Anthropic shares. Anthropic says it doesn't expect any notes to be sold before it completes the offering, which could value it at $2 trillion. It stands to become the largest customer in Broadcom's chip design business next year, and Broadcom's largest compute customer in 2027. Broadcom projects roughly $115 billion in AI semiconductor revenue in fiscal 2027 and $230 billion in fiscal 2028.
Why it matters for builders
The prospectus itself says Broadcom's double role as hardware supplier and financing partner creates potential conflicts of interest that might affect Anthropic's ability to access the computing power it needs. If your product runs on Claude, that is now a disclosed supply risk rather than a guess. Worth keeping a second model path tested this quarter, even if you never route traffic to it.
arXiv now limits each submitter to two submissions per calendar month, with a maximum of three active submissions at any time. The limit counts submissions, not announced papers, so a rejected paper still burns a slot and two tries in a month means waiting until the next one. arXiv says it received 40,363 submissions in September 2026, which generated almost 9,000 support tickets for staff and moderators, and that cs.AI submissions rose over 6X in the past two years. arXiv says AI tools make it easy for authors to flood arXiv and other repositories with low-value papers. Rate limits apply only to the submitter, not co-authors.
If you preprint your work, treat each slot as scarce: get formatting and category right before you hit submit, because a bounce costs you the month. Groups shipping several papers should decide now which co-author submits which paper, since the cap follows the account that submits, and keep no more than three in flight.
Claude Code v2.1.287 adds Claude Mods, letting plugins modify deeper behavior. It ships one built-in mod, "You should know": a side agent watches the session and flags things you or Claude might miss, enabled with /plugin enable cc-plugin-you-should-know@builtin for first-party sessions with telemetry on. Opus 4.7+ and Fable now default to a 1M context window on Bedrock, Vertex, Foundry and the Claude apps gateway, with no [1m] suffix, and CLAUDE_CODE_DISABLE_1M_CONTEXT=1 keeps 200K. MCP servers set to alwaysLoad: false now defer all of that server's tools behind tool search. VSCode adds "Run in background" for a running command or sub-agent.
Why it matters for builders
Two things to check before you upgrade. If an MCP server stops connecting afterwards, add "bareElicitationCapability": true to its config entry, since the 2025-11-25 protocol now carries URL prompts, for example to sign in. And because plugins can now reach deeper into the agent, read what a third-party plugin actually does before you install it on a work machine.
