Value Index / Research / Agentic coding spend controls Spend-control matrix · checked 2026-07-18
Can agent work continue into paid usage after the included allowance runs out? Verdict: Copilot, Cursor, Antigravity, and Claude Max all expose different continuation mechanics. A published allowance does not mean paid continuation is blocked.
Safest default: turn off or cap paid continuation unless official documentation establishes both the continuation route and its control.
Compare the four controls ↓ One scan · no winner score
The included-usage wall and the switch that can open spend ▲ Relative or unpublished headroom is a negative predictability signal—not a numeric score.
Spend controls after included agentic-coding usage is exhausted Provider / evidence Allowance disclosure At exhaustion Paid continuation Blocking control Predictability Safest action Official receipt GitHub Copilot ● Published $ / credits Pro 1,500; Pro+ 7,000; Max 20,000 AI Credits per calendar month. 1 credit = $0.01. Credit-metered requests are blocked unless a budget permits additional paid usage. Yes—additional usage can continue against a configured budget. Do not set an additional-usage budget, or set the budget to $0. Concrete monthly credits and dollar value; model choice still changes burn. Keep additional paid usage at $0 until the team chooses a budget. GitHub Copilot plans ↗ GitHub individual usage billing ↗ Cursor ▲ Published $ / credits Pro includes $20, Pro+ $70, and Ultra $400 of API agent usage. The first-party Auto + Composer pool is only described as “generous.” API usage can move to on-demand when enabled. First-party pool exhaustion behavior: Not published—do not assume. Yes—on-demand usage is billed at model API rates when explicitly enabled. Keep on-demand usage off, or set its spend limit before agent work. Mixed: API-dollar pool is concrete; first-party pool headroom is unpublished. Leave on-demand off until a spend limit is set; do not budget from “generous.” Cursor usage and limits ↗ Cursor usage-based charges ↗ Google Antigravity ▲ Not published Five-hour and weekly quota behavior is documented, but an exact public allowance is Not published—do not assume. Quota availability waits for its five-hour or weekly reset; behavior beyond that is Not published—do not assume. AI Credit Overages provides Never or Always behavior. Set AI Credit Overages to Never. Negative: dynamic/relative quota without an exact public allowance. Choose Never before enabling agents; reject the plan if an exact budgetable allowance is required. Antigravity plans ↗ Changes to Antigravity plans ↗ Claude Max ▲ Relative headroom Max is published as 5× or 20× more usage than Pro, not as a stable token allowance. Included usage reaches a usage limit; paid continuation is a separate usage-credits route. Usage credits can continue at standard API rates. Enabling/default state: Not published—do not assume. Keep usage credits disabled, or set the documented monthly spend limit. Negative: relative capacity and changing usage conditions are not a fixed budgetable allowance. Do not enable usage credits without a monthly spend limit. Claude Code with Pro or Max ↗ Manage usage credits ↗
GitHub Copilot ● Published $ / credits
Allowance disclosure Pro 1,500; Pro+ 7,000; Max 20,000 AI Credits per calendar month. 1 credit = $0.01.
At exhaustion Credit-metered requests are blocked unless a budget permits additional paid usage.
Paid continuation Yes—additional usage can continue against a configured budget.
Blocking control Do not set an additional-usage budget, or set the budget to $0.
Predictability Concrete monthly credits and dollar value; model choice still changes burn.
Safest action Keep additional paid usage at $0 until the team chooses a budget. Cursor ▲ Published $ / credits
Allowance disclosure Pro includes $20, Pro+ $70, and Ultra $400 of API agent usage. The first-party Auto + Composer pool is only described as “generous.”
At exhaustion API usage can move to on-demand when enabled. First-party pool exhaustion behavior: Not published—do not assume.
Paid continuation Yes—on-demand usage is billed at model API rates when explicitly enabled.
Blocking control Keep on-demand usage off, or set its spend limit before agent work.
Predictability Mixed: API-dollar pool is concrete; first-party pool headroom is unpublished.
Safest action Leave on-demand off until a spend limit is set; do not budget from “generous.” Google Antigravity ▲ Not published
Allowance disclosure Five-hour and weekly quota behavior is documented, but an exact public allowance is Not published—do not assume.
At exhaustion Quota availability waits for its five-hour or weekly reset; behavior beyond that is Not published—do not assume.
Paid continuation AI Credit Overages provides Never or Always behavior.
Blocking control Set AI Credit Overages to Never.
Predictability Negative: dynamic/relative quota without an exact public allowance.
Safest action Choose Never before enabling agents; reject the plan if an exact budgetable allowance is required. Claude Max ▲ Relative headroom
Allowance disclosure Max is published as 5× or 20× more usage than Pro, not as a stable token allowance.
At exhaustion Included usage reaches a usage limit; paid continuation is a separate usage-credits route.
Paid continuation Usage credits can continue at standard API rates. Enabling/default state: Not published—do not assume.
Blocking control Keep usage credits disabled, or set the documented monthly spend limit.
Predictability Negative: relative capacity and changing usage conditions are not a fixed budgetable allowance.
Safest action Do not enable usage credits without a monthly spend limit. Provider evidence notes
What each official receipt establishes Every decisive row links directly to vendor documentation checked on 2026-07-18.
Methodology refusal
No comparable cross-provider token total exists. Dollar-denominated API pools, AI Credits, relative 5×/20× capacity, dynamic quota windows, and unpublished first-party headroom are not equivalent units. Converting them into one guaranteed token allowance would require model mix, routing, cache, context, and vendor behaviors the providers do not publish as one contract. This matrix therefore compares spend-control behavior only; it does not normalize or rank headroom.