In May 2026, gihyo.jp reported that Anthropic would introduce monthly credits dedicated to programmatic usage in Claude paid plans. Separate from the "conversational UI quota" in Claude Pro and Max plans, this is a new pricing model that sets aside monthly credits dedicated to programmatic usage (such as APIs, SDKs, and Claude Code).
This is the official answer to real-world incidents in custom development where "borrowing API keys from personal plans → heavy consumption on projects → depletion early in the month" kept occurring. This represents Anthropic independently structuring the model already covered in GitHub Copilot Pay-as-You-Go — Token Governance for Custom Development. In this article, we outline budget governance design to prevent Claude API cost explosions in custom development.
Why "programmatic-only credits" became essential
| Use case | Traditional | After introduction of monthly programmatic credits |
|---|---|---|
| Conversational UI usage | Pro / Max message allocation | Same as above |
| API / SDK usage | Consumes same allocation → Depletes conversational side | Measured in separate allocation |
| Claude Code batch execution | Heavily consumes personal credits | Consumes programmatic allocation only |
| CI integration | Explodes invisibly | Halts at limit |
| Multi-project apportionment | Impossible | Split by project |
In particular, incidents where "credits vanished the moment Claude Code ran in CI" were wreaking havoc on monthly cost accounting in custom development. Separating a dedicated programmatic tier is an inevitable evolution to "make programmatic usage measurable while safeguarding the interactive conversational experience."
Four layers of budget governance built for custom development
Layer 1: Three-axis quota across project × environment × agent
| Axis | Examples | Quota |
|---|---|---|
| Project | Client A Project | 500,000 JPY / month |
| Environment | Production / Staging / Development | 100,000 / 200,000 / 200,000 JPY each |
| Agent | Refactor / Test / Docs | 150,000 / 250,000 / 100,000 JPY each |
Relying solely on a "project-level ceiling" leads to accidents where a single agent consumes everything and starves the rest. We physically isolate them using three-axis quotas.
Layer 2: Three stages of consumption alerts
We send tiered notifications to Slack / email / PagerDuty at 80% / 95% / 100%. We incorporate automatic agent shutdown at 100% as a guardrail. This standardizes the "budget cutoff" addressed in Claude Code Auto Mode Approval Gates for Custom Development.
Layer 3: Dynamic optimization of model selection
We implement routing that automatically downgrades models based on remaining credits.
| Remaining credits | Recommended model | Application |
|---|---|---|
| 80% or more | Claude Opus 4.7 | Critical architecture and review |
| 50〜80% | Claude Sonnet 4.6 | Standard implementation |
| 30〜50% | Claude Haiku 4.5 | Lightweight tasks |
| Under 30% | Batch only + approval required | Production emergencies only |
We dynamically operate "model-by-model cost optimization," as covered in Optimizing Claude Code Operational Costs in 2026, under a monthly credit system.
Layer 4: Transparency in monthly reporting and client billing
We present project-by-project API consumption reports to clients on a monthly basis, visualizing "what was spent on what." This ensures we can demonstrate with figures the basis for invoice allocations.
Four phases to build in custom development
Phase 1: Current state audit and budget allocation definition (2 weeks)
We conduct an inventory of Claude API consumption over the past three months by project, environment, and agent, and agree with the client on the initial monthly programmatic credit allocation.
Phase 2: Quota gate implementation (3 weeks)
We implement three-axis quotas at the proxy layer (such as LiteLLM), incorporating 80/95/100% alerts and automatic shutdown upon overruns.
Phase 3: Model router implementation and optimization (4 weeks)
We build a reporting foundation that enables dynamic model switching based on remaining credits and forecasts next month's budget from historical consumption data.
Phase 4: Operationalizing monthly reporting (ongoing)
In monthly review meetings, we run a PDCA cycle for quota adjustments and model selection rule revisions.
Standard technology stack set for custom development
| Layer | Recommended technology | Alternative |
|---|---|---|
| API gateway | LiteLLM Proxy | OpenRouter |
| Quota management | Redis + Lua scripts | DynamoDB |
| Alerting | Slack Block Kit + PagerDuty | Microsoft Teams |
| Observability | Langfuse / Helicone | Datadog APM |
| Cost aggregation | BigQuery + dbt | Snowflake |
| Client reporting | Looker Studio | Metabase |
| Model router | LiteLLM Routing | In-house router |
As described in Claude Platform on AWS — Enterprise AI Custom Development, when using Claude via AWS Bedrock, combining it with AWS-side cost guards is also an effective pattern.
Which projects it fits best
| Suited projects | Benefit |
|---|---|
| Projects running Claude Code company-wide | Prevent credit exhaustion |
| Concurrent operations across multiple client projects | Project allocations are visualized |
| IPO preparation / audit compliance | Retain evidence for API consumption |
| Growth stage of an AI startup | Monthly burn rate management |
| Enterprise custom development | Ensure safety via environment-specific quotas |
Six clauses to include in client contracts
| Clause | Details | What the client should verify |
|---|---|---|
| Monthly credit limit | Cap amounts by project × environment | Demarcation of responsibility during overruns |
| Behavior upon overrun | Automatic shutdown / continuation upon approval | Tolerance level for production impact |
| Model modification authority | Permission for dynamic routing by custom development team | Balance with output quality |
| Consumption reports | Monthly / weekly / daily | Alignment with audit requirements |
| Emergency quota expansion procedure | Quota expansion workflow within 24 hours | Communication channels and decision-makers |
| Credit refunds upon cancellation | Handling of remaining balances | Settlement upon contract termination |
Four common pitfalls
Pitfall 1: "Running client projects on individual plans"
Running custom development projects on individual Pro / Max plan API quotas constitutes a triple violation of privacy, contract terms, and cost transparency. We explicitly state in contract clauses from the outset that Org / Team plans + dedicated programmatic allocations will be used.
Pitfall 2: Not "annualizing single-month consumption"
Simply converting single-month credit consumption by assuming "multiplying by 12 gives the annual budget" causes teams to overlook the peaks and valleys of busy seasons. We determine seasonal coefficients based on the peaks and valleys of the past six months.
Pitfall 3: Leaving quality degradation unaddressed after model downgrades
There are cases where downgrading to Haiku results in reduced output quality. We establish rules to continuously measure a "quality gate (test pass rate / review rejection rate)" and revert to Sonnet if it falls below the threshold.
Pitfall 4: Inability to explain invoice allocations to clients
If an invoice merely states "Project A: 270,000 yen," clients cannot approve it. Design tracking with the granularity to break costs down by agent, such as "Refactoring Agent: 120,000 yen / Test Agent: 150,000 yen."
Summary — Toward a custom development standard that "never exhausts credits"
The introduction of Anthropic monthly programmatic credits marks a turning point from an era of "saving Claude API usage through sheer willpower" to "governing it through architecture." Building in three-axis quotas, tiered alerts, and dynamic model routing from the outset becomes the baseline for protecting monthly costs in custom development projects.
How far to build depends on each individual project. For organizations that can break down their past three months of consumption by project, environment, and agent, simply inserting three-axis quotas into a proxy might be enough to get over the hump. Conversely, if consumption breakdowns are not tracked in the first place or if production cannot be stopped under any circumstances, inventorying and agreeing on the demarcation of responsibility during overruns comes first; model routers and allocation reporting platforms come later. This determination cannot be made purely on paper without examining actual consumption data and contract terms. We welcome inquiries via our contact form starting from stages such as "Claude Code costs are unpredictable" or "We want to establish project allocation reporting."
Sources
- Anthropic to Introduce Dedicated Monthly Programmatic Credits for Paid Claude Plans (gihyo.jp)
- GitHub Copilot Pay-as-You-Go Token Governance Client Services (GH Media)
- Claude Code Operational Cost Optimization 2026 (GH Media)
- Claude Code Auto Mode Approval Gate Client Engagements (GH Media)
- Claude Platform on AWS — Enterprise AI Custom Development (GH Media)
