First CAISI Eval Card Drops, TurboQuant Hits a Serving Path, Colorado Walks the Repeal Up: Sunday Digest, May 10, 2026

One of the open threads from Saturday’s briefing closed in 24 hours. OpenAIpublished a redacted pre-release evaluation card for a frontier model under its CAISIagreement — the first time any of the five U.S. frontier labs has voluntarily disclosed the shape of a federal pre-release eval. The card covers categories (CBRN uplift, cyber-offense, agentic autonomy, deception, model-weight exfil resistance) and pass/borderline/fail bands, but withholds specific prompts and numerical thresholds. It is the first concrete answer to the question every procurement office has been asking since CAISI extended to Google DeepMind, Microsoft, and xAI on May 5: what does CAISI conformance actually look like on paper?

The downstream effect is structural. The other four labs — Anthropic, Google DeepMind, Microsoft, xAI — are now under implicit pressure to either match the disclosure or explain why they will not, and any of them moving to publish a card with more detail (numerical thresholds, full category list, third-party attestation) instantly becomes the new baseline. Procurement teams that have been treating CAISI conformance as a checkbox can now treat it as a document — one that should sit alongside SOC 2 and ISO 42001 in vendor packets going forward.

Cloudflare’s Workers AI rolled out a production serving configuration that combines Google TurboQuant (PolarQuant rotation plus quantized Johnson–Lindenstrauss KV-cache projection, ICLR 2026) with its own Unweight15–22% weight-pruning pass and the previously announced separated-stage inference engine. Cloudflare reports a combined ~38% reduction in long-context unit cost relative to its April baseline at parity quality on internal eval suites. This is the first time the two techniques have appeared in a single production path from a Western inference provider, and it lands roughly two weeks earlier than the Q3 timeline industry analysts had penciled in.

The strategic reading is that long-context economics have stepped down ahead of schedule. Workloads that were marginal at 100K–300K tokens in April — long-document review, codebase-wide agent runs, multi-day research traces — are inside the budget envelope for a much wider set of tenants this week. Expect Bedrock, Vertex AI, and Azure AI Foundry to follow with comparable bundles before the end of Q2, and expect frontier-lab list pricing for long-context tiers to absorb at least part of the new cost curve in a refresh.

The Colorado AI Act repeal — the most-watched state-level rollback in the country — cleared its first committee vote on Friday 7–4 and is on the floor calendar this week. The replacement is a disclosure-only regime: covered deployers would have to notify users when an AI system makes a consequential decision, but the high-risk-system obligations, impact assessments, and AG enforcement teeth that defined the original Act would be stripped out. The Act’s previously scheduled June 30 effective date now sits inside the same window as a possible repeal vote.

If the repeal lands, it becomes the precedent any state legislature can point to when it wants disclosure without high-risk classification — the Connecticut Maroney bill, the Maryland AI price-setting ban, and the active tracks in Oklahoma, Hawaii, Michigan, and New York all sit in a different shape than they did a week ago. On the federal side, the White House National AI Policy Framework preemption push still has no Congressional vehicle attached, and the EU AI Act high-risk deferral remains in inter-institutional review.

The ten preconfigured financial-services agents Anthropic shipped on May 5 now have a first apples-to-apples benchmark, run by an industry consortium of buy-side and middle-office firms. The strongest result was compliance-escalation — flagging, triaging, and routing items to the right second-line owner — where the agent matched a senior human reviewer on a held-out set with materially lower latency. The weakest was pitch-deck drafting, which the consortium scored notably below an experienced associate on a structured rubric covering narrative coherence, data fidelity, and house-style adherence.

The middle of the table — financial-statement review, middle-office workflow automation, pre-trade compliance checks — is where most of the procurement decisions sit, and the consortium reported workable production-readiness on those five tasks with human-in-the-loop oversight. This is the first vertical-tailored scoreboard from a credible third party for a frontier-lab agent suite, and it sets the template for how agent procurement will work for the rest of 2026: per-task evaluation against a published rubric, not headline capability claims.

Four threads to track this week. First, which CAISI lab publishes second, and whether their card carries more category detail than OpenAI’s — the answer sets the ceiling on disclosure for the rest of 2026. Second, follow-on serving-path announcements from Bedrock, Vertex AI, and Azure AI Foundry on TurboQuant or comparable KV-cache compression — expect at least one before the end of May. Third, the Colorado floor vote — pass or fail, the result reshapes how every other state legislature drafts its 2026 AI bill. Fourth, the next vertical-agent scoreboard: IBM’s watsonx Orchestrate refresh and Microsoft’s Agent 365 are the most-likely next subjects, and the rubric Anthropic’s FS agents were graded on this week is a credible template for both.