🧭 Agent SDK Credit Pool Is Live: Your Day-One Monitoring Checklist

As of midnight tonight, all Claude Agent SDK workloads, claude -p commands, Claude Code GitHub Actions runs, and third-party agent applications now draw from a separate monthly credit pool — not your subscription's interactive usage limit. The billing architecture described in yesterday's countdown article is now a runtime reality. If you enabled overflow billing and updated your model IDs, you're in good shape; if not, this is the guide for your next 30 minutes.

Where to find your credit balance

Navigate to Claude.ai → Account → Billing → Agent SDK Credit. You'll see:

• Credit used this period — displayed in dollars at standard API list rates
• Credit remaining — the gap between your plan's monthly allowance and consumed usage
• Overflow billing status — ON or OFF, with a toggle and a spend cap field
• Usage breakdown — a per-day chart and per-workload table, updated within ~5 minutes of each run

The credit resets on your billing cycle anniversary, not on the calendar month. If you subscribed on the 20th, your Agent SDK credit resets on the 20th — not on July 1.

What to configure now

The minimal safe configuration for any team running automated workloads:

# 1. Enable overflow billing via the Billing UI, then set a hard cap.
#    A cap of 2–3× your typical monthly run cost is a reasonable starting point.

# 2. Add a usage check to your CI pipeline wrapper:
#    (Example using the Anthropic management API — not the inference API)
curl -s -H "x-api-key: $ANTHROPIC_ADMIN_KEY" \
  "https://api.anthropic.com/v1/billing/agent-sdk-credit" \
  | jq '.remaining_usd'

# 3. If you run Python agents via the Agent SDK, check the credit header on each response:
response.headers.get("x-agent-sdk-credit-remaining-usd")
# Returns a float string like "17.42" — log it; alert if below your threshold.

Credits are strictly per-user

The credit pool is tied to the individual subscription seat. On a Team or Enterprise plan, each seat has its own $20–$200 monthly credit — there is no team-level pool to draw from. A power user who runs heavy overnight evaluations cannot "borrow" from a colleague who has barely touched the API. If your team's Agent SDK usage is concentrated in a few engineers, consider whether those individuals should be on a higher plan tier, or whether centralising agent workloads through a shared service account (billed via the API directly) makes more economic sense.

🧭 Anthropic's Economic Policy Framework: Three Unemployment Tiers, $350M Commitment, and a Proposed AI Jobs Tax

Published alongside Dario Amodei's "Policy on the AI Exponential" essay last week, Anthropic's standalone Economic Policy Framework (EPF) is a detailed US policy document calibrated to three possible levels of AI-driven labour-market disruption. It is the most concrete thing Anthropic has published about how it thinks the economy should respond if its own models cause mass unemployment — and it comes with a $350 million funding commitment to back it up.

The three-tier structure

The EPF does not propose a single policy response. Instead it defines three escalating tiers based on the headline US unemployment rate, with distinct recommended interventions at each level:

• Tier 1 — ~5% unemployment (near current levels): Universal capital accounts seeded with equity stakes in AI companies; expanded worker retraining programmes; portable benefit portability reforms so workers can move between gigs and employers without losing healthcare or pension continuity.
• Tier 2 — ~10% unemployment (severe but precedented): Automatic extensions to unemployment insurance triggered without Congressional action; direct basic-needs relief for displaced workers; temporary payroll-tax relief for sectors absorbing displaced workers.
• Tier 3 — unemployment at "unprecedented" levels: A common income floor (structurally similar to UBI but explicitly positioned as temporary and conditional); levies on AI infrastructure use — essentially a tax on compute or API calls paid by AI providers including Anthropic — directed into a displacement fund; exploration of AI sovereign wealth funds to distribute AI productivity gains broadly.

The $350M commitment

Anthropic is not just lobbying for these frameworks — it is funding the evidence base to make them credible:

• $200M Economic Futures Research Fund: Grants to economists, labour researchers, and policy institutes studying AI's actual impact on employment, sector by sector, and running controlled trials of the Tier 1 and 2 policy responses.
• $150M Claude Corps Fellowship: The programme announced on June 11 (placing 1,000 early-career fellows inside US nonprofits) is formally part of the EPF funding package — framed as a Tier 1–adjacent initiative that builds human institutional capacity alongside AI adoption.

🧭 The Advanced AI Framework: Anthropic Calls for FAA-Style Government Authority to Block Dangerous Model Deployments

The second governance document released alongside Dario Amodei's "Policy on the AI Exponential" is Anthropic's Advanced AI Framework (AAF) — a proposed regulatory structure for the most capable frontier models. Where the Economic Policy Framework addresses economic consequences, the AAF addresses catastrophic deployment risks: bioweapon uplift, CBRN capability transfer, autonomous cyberattack infrastructure, and loss of meaningful human oversight. The AAF is a direct answer to the question that the Fable 5 suspension has made unavoidably concrete this week: who should have the authority to take a dangerous model offline, and under what rules?

The FAA analogy

Anthropic's framework explicitly draws on the Federal Aviation Administration model: an independent technical agency with the authority to ground aircraft that pose safety risks, staffed by engineers who understand the systems they regulate, operating under clear pre-defined criteria rather than political discretion. Dario Amodei's VentureBeat interview articulated the core argument — the US currently has no body capable of making an informed, technically grounded decision about whether a specific model's capabilities cross a safety threshold that warrants deployment restriction. The AAF proposes creating one.

Key proposals

• Pre-deployment capability evaluations: Frontier models above a defined compute threshold (roughly 10²⁶ FLOPs training compute as a starting trigger) would require independent capability and safety evaluations before public release — run by a government-designated body or accredited third-party labs, not self-assessed by the developer.
• Deployment authority: The regulatory body would have statutory power to require phased deployment, impose capability restrictions, or suspend access — with defined criteria for each action and a right of expedited appeal. This is explicitly distinct from ad-hoc national security directives of the kind that triggered the Fable 5 suspension.
• International coordination: Anthropic proposes that the US framework be designed from the outset for multilateral adoption — coordinated with the UK AI Safety Institute, the EU AI Office, and treaty partners — so that unilateral US deployment blocks do not simply redirect dangerous model access to less regulated jurisdictions.
• Developer transparency obligations: Model cards expanded to include structured capability assessments in categories relevant to biosecurity, cybersecurity, and persuasion; mandatory incident reporting for any exploitation of safety restrictions at scale; and algorithmic auditability rights for the regulatory body.