Should Coding Agents Run on Your Subscription or API Credits?

Published October 12, 2026 · by the AQ team

Whether a coding agent run costs you subscription allowance or API credits is decided by how the tool authenticates, not by which tool or model you picked. As of October 2026 both Anthropic and OpenAI run two meters side by side: an agent CLI signed in with a plan (Claude Code on a Claude plan, Codex on a ChatGPT plan) draws from that plan's included usage windows at a fixed monthly price, while the same software running on an API key is billed per token against a prepaid credit balance. October 2026 redrew the Anthropic side of this line: Claude Max and Team plans now include monthly API credits, claimed through the Claude Console, that cover the Claude API and the Agent SDK but explicitly not Claude Code.

The two meters, defined

A subscription allowance is the usage included in a flat monthly plan, metered in opaque units that reset on a schedule: Claude plans combine a session limit that resets every five hours with a weekly limit, shared across the Claude apps and Claude Code; ChatGPT plans give Codex and chat one shared pool with per-five-hour estimates and possible weekly limits. You cannot overspend it or roll it over. When it runs out, you wait, buy extra usage, or switch accounts.

API credits are a prepaid per-token balance. Every request is priced by exact input and output tokens at the published rate (as of October 2026, Claude Opus 5.5 is $4 per million input tokens and $20 per million output; Claude Sonnet 5.5 is $2 and $10; Claude Haiku 5.5 is $0.10 and $0.50 for prompts up to 100,000 tokens). There are no five-hour windows and no weekly caps, only rate limits and whatever spend limits you configure.

What Anthropic changed on October 7, 2026

Anthropic's October 7, 2026 update gave Claude Max and Team subscribers a monthly allowance of real API credits: $100 per month on Max 5x, $200 on Max 20x, and on Team $20 per Standard seat and $100 per Premium seat, pooled per team and capped at $500 per month. You claim them once by linking a Claude Console organization from your plan's billing settings (new subscribers wait seven days), and the credits then arrive each billing cycle, no payment method required.

The boundaries matter more than the headline. As of October 2026, per Anthropic's own documentation:

• The credits cover the Claude API, the Claude Agent SDK, Claude Managed Agents, and the Console Playground. They do not cover Claude Code or extra usage in the Claude apps, and they do not apply on Amazon Bedrock, Google Cloud, or Microsoft Foundry.
• Unused credits expire at the end of each billing cycle, with no rollover.
• Claiming them changes nothing about your plan's usage limits in Claude or Claude Code.
• Each plan links to exactly one Console organization; changing the link requires support.

This lands against a confusing backstory. Anthropic had announced that Agent SDK and claude -p usage would move off subscription plans onto a separate monthly credit, then paused that change on June 15, 2026, the day it was due to take effect. As of October 2026 the pause still holds: the Agent SDK, claude -p, and third-party apps signed in with your plan still draw subscription limits. The new monthly API credits are the other path: run the same SDK code against an API key from your linked Console organization and it burns the included credits instead.

Separately, Anthropic sells usage credits (extra usage) on Pro and Max: once an included limit is used up, requests can continue at standard API rates billed to a prepaid balance, and this does apply to Claude Code. You can cap it with a monthly spend limit, and in most regions these credits do not expire. Confusingly, the new monthly API credits cannot be spent as usage credits; they live in the Console, not in your plan.

How the same line runs through Codex

OpenAI draws the identical boundary with different names. As of October 2026, OpenAI's pricing documentation says Codex is included in ChatGPT plans (Free, Go, Plus, Pro, Business, Edu, and Enterprise), sharing one usage pool with chat. For Plus and standard Business seats it publishes per-five-hour estimates by model (for example 15 to 160 local messages on GPT-6.1 Sol, depending on task size), notes that weekly limits may also apply, and says Pro currently has no five-hour limit. When the pool runs dry, Plus and Pro users can buy additional credits.

Sign Codex in with an API key instead and every request is billed at standard API pricing with no plan windows at all. The tradeoff is not just money: as of October 2026 OpenAI's docs state that cloud features such as GitHub code review and the Slack integration are unavailable when Codex runs on an API key. The plan is not only an allowance, it is also the product surface.

The rule that decides the meter

Across both vendors one rule holds: the credential decides the meter, never the binary. The same Claude Code install meters against your plan when signed in with claude.ai OAuth and against per-token credits when it finds an API key. The same Agent SDK script is plan usage under a subscription login and API spend under a key. A harness that drives the real CLI under each person's own login consumes that person's plan; a product that calls the API with its own key is buying tokens and will price them into its bill somehow.

That rule produces the surprises worth knowing in advance:

• The new Max API credits cannot rescue a Claude Code session that hit its weekly limit. Relief comes from extra usage credits, the reset, or another connected account.
• An unattended run on a subscription login can stall mid-task when a window closes, which an API-key run never does. Decide which failure mode you prefer for long work: a paused run or an unbounded bill. Overnight coding agent runs covers the safety half of that decision.
• CI and scheduled automation should not run on a person's plan login: it competes with their interactive work for the same windows and couples the pipeline to one human's account.

Where a run lands (October 2026)

Vendor mechanics verified on Anthropic and OpenAI documentation, October 2026.

Which should your team use?

Most teams end up running both meters on purpose, split by workload:

• Interactive daily coding: the subscription. For an engineer driving an agent all day, a flat plan is almost always cheaper than the same tokens at API rates, and the cost is predictable. The plan windows are the constraint, so learn their resets.
• Automation, CI, and bulk work: API credits. Per-token billing with explicit spend limits is the right shape for unattended work, and the new Max and Team monthly credits make a real sandbox free: $100 to $500 a month covers a lot of Haiku-class or carefully routed token spend, and Anthropic's Batch API takes 50 percent off both input and output for work that can wait.
• Overflow: decide before you hit it. On Claude, turning on extra usage with a monthly cap converts a hard stop into a bounded bill. On Codex, purchased credits do the same. Teams that instead run several connected accounts per person should treat that as the deliberate choice it is.

Whatever the split, keep it visible: the failure mode is discovering at invoice time that a harness quietly routed a month of usage through a marked-up key, or that the team burned its weekly windows on work a $5 batch job could have done.

Where AQ fits

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. AQ sits deliberately on the subscription side of this line: agents run as the real CLIs (Claude Code, Codex, Cursor Agent, Kimi, Grok, or plain shells) in persistent tmux sessions on your team's VM, each signed in with the person's own Claude or OpenAI account. A run in AQ meters exactly like the same run on your laptop (no inference markup, no shared vendor key).

The plan-window math is also why the team layer matters. Sessions in AQ survive a closed laptop and resume from any device, teammates can open the same workspace and watch the same live session, and each workspace gets its own isolated git worktree, so running several agents in parallel is normal. Parallel sessions reach a five-hour window faster than one does, and a session that outlives your laptop can hit a weekly cap while you sleep, so the dashboard that shows what every agent is doing is also where you notice usage before the meter does.

Pricing is two plans: Free is a personal sandbox for one person (AQ creates a private machine, nothing to install, no time limit), and Team is $50 per user per month early access ($200 standard, billed monthly), covering VMs from your own cloud or a dedicated always-on AQ-managed VM, rate locked for your first 12 months.