The AI Dispatch

Colorado Signs SB 26-189; Bartz v. Anthropic $1.5B Settlement Heads to Approval

By AI Dispatch Desk · May 14, 2026 · Sources: Holland & Knight, Words and Money, Courthouse News

Two consequential legal milestones landed on the same day, both reshaping how AI law looks heading into the back half of the year. Governor Jared Polis signed SB 26-189 into law in Denver, completing the rapid reversal of what had been America’s first comprehensive AI regulation; the same morning, Judge Araceli Martínez-Olguín held the final fairness hearing for Bartz v. Anthropic, the largest AI copyright settlement in United States history, in the Northern District of California — and signaled imminent approval. Taken together, the two rulings draw the clearest picture yet of where the AI legal landscape is settling: state-level prescriptive regulation is in retreat, while the courts are willing to enforce massive copyright recoveries at scale.

SB 26-189 is the replacement bill for the Colorado AI Act, the 2024 statute that had positioned the state as the U.S. analog to the EU AI Act and that had been scheduled for a phased rollout in early 2027. The original act’s obligations — algorithmic-impact assessments, risk-management programs, deployer notification requirements — had drawn sustained pushback from industry through the 2025 legislative session, and Governor Polis publicly signaled discomfort with the implementation timeline as early as October. The signed replacement narrows the statute to a disclosure regime: deployers of high-risk AI systems must notify consumers when an AI system is used to make a consequential decision, and consumers gain limited rights to obtain an explanation. The compliance date is January 1, 2027 — the same date the original Act would have taken effect, allowing affected businesses to plan against a single deadline. The constitutional challenge brought by xAI and the U.S. Department of Justice against the original Colorado AI Act, filed in February of this year, remains pending; the new statute’s narrower scope is widely expected to render most of the constitutional claims moot, though the litigation has not yet been formally withdrawn.

The Bartz v. Anthropic hearing in San Francisco took roughly two hours. Judge Martínez-Olguín had previously granted preliminary approval; Thursday’s hearing was the final fairness review under Federal Rule of Civil Procedure 23(e). According to courtroom reporting from Words and Money and Courthouse News, the judge expressed few concerns about the substance of the deal itself. The $1.5 billion settlement fund — the largest in any AI copyright matter to date — covers an estimated 448,000 books across the certified class. Counsel reported that 92.77 percent of covered books had been claimed by the close of the claims window, an unusually high participation rate for a class action of this size, which class counsel attributed to outreach through publisher associations and the relatively unambiguous criteria for inclusion. Eligible authors and publishers will receive approximately $3,000 to $3,100 per work, with variations depending on whether the work was used for training, evaluation, or both.

The judge’s scrutiny instead focused on attorney fees. Class counsel had requested a fee award representing roughly twenty-five percent of the settlement fund — the standard ceiling for class actions in the circuit, but unusually large in absolute terms given the size of the recovery. Martínez-Olguín questioned the lodestar cross-check and asked counsel to file supplemental briefing on the hours-versus-recovery ratio before she issues the final approval order. She nonetheless signaled, in remarks from the bench, that she expected to approve the settlement itself promptly, with the fee question handled in a separate order. Final approval — closing the door on the largest single AI copyright matter to date — is widely expected within weeks.

The pairing of these two rulings on the same day is the kind of coincidence the calendar throws up only rarely, and it produces a clarifying contrast. Where state legislatures are deciding that prescriptive AI regulation is more cost than benefit, federal courts are demonstrating that ordinary intellectual-property doctrine, applied at scale, can extract recoveries from frontier labs measured in the billions. The legal architecture that constrains AI in 2026 is not, in the end, the architecture that policymakers spent 2024 and 2025 drafting. It is the architecture the courts and the markets are settling into.

Claude Code Ships Twice in One Day — v2.1.141 and v2.1.142

Source: github.com/anthropics/claude-code releases

Anthropic pushed two Claude Code releases on Thursday within hours of each other — an unusual cadence for the tool. The morning release (v2.1.141) carried a batch of hook and plugin improvements that had been queued in the changelog for several weeks; the afternoon release (v2.1.142) bumped Fast Mode to a new default model and added a fleet of dispatch flags to the claude agents command that materially expand how power users compose subagents. Both are recommended upgrades; the second is the more consequential of the two.

v2.1.141 — Hooks, Plugins, and Skills Cleanup

• Hook JSON output gains a terminalSequence field, allowing hooks to emit desktop notifications, window-title updates, and terminal bells without a controlling terminal — previously a sharp edge for headless/CI usage.
• New CLAUDE_CODE_PLUGIN_PREFER_HTTPS environment variable forces HTTPS plugin URLs for environments where no GitHub SSH key is configured. Smoothest path yet for short-lived sandboxes and ephemeral cloud builders.
• Plugins that ship a root-level SKILL.md with no skills/ subdirectory are now surfaced automatically as a skill. Removes a long-standing packaging quirk that forced authors to either nest or duplicate.
• Fixed a class of stop hooks that caused an infinite loop on block. Esc and Ctrl+C now reliably cancel a pending /loop wakeup — closing a footgun that had bitten several reporters of the issue.

v2.1.142 — Opus 4.7 Default, Agents Dispatch Flags

• Fast Mode now defaults to Opus 4.7 (previously Opus 4.6). Users who want to pin the previous model can set CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1. The change improves the speed/quality envelope of the lighter-weight reasoning path without surfacing as a model selector to the average user.
• The claude agents dispatch command gains a suite of new flags: --add-dir, --settings, --mcp-config, --plugin-dir, --permission-mode, --model, --effort, and --dangerously-skip-permissions. Each flag is also accepted via the parent claude invocation, but exposing them on agents lets a parent agent dispatch subagents with materially different configurations, MCP servers, and plugin sets — a primitive that had been missing for orchestration patterns at any meaningful scale.

The combination is the most workflow-affecting Claude Code release pair of the past month. The Opus 4.7 Fast Mode default lifts the floor on what an average user gets without any configuration; the agents-dispatch flags lift the ceiling on what an orchestration-savvy user can express. Coming on the same day as the launch of Grok Build, the timing reads as deliberate — Anthropic moving to close the orchestration-flexibility gap before xAI’s entrant can establish a beachhead.