July 2026 · Guide

Claude Code cost in 2026: plans, limits, and usage credits

On July 20, 2026, Anthropic changed how Fable 5 access works across paid plans. It remains included within limits for Max and premium Team or Enterprise seats, while Pro and standard seats use pay-as-you-go usage credits for that model. Claude Code itself remains included with Pro and Max plan limits, with optional API-credit usage after those limits. Check Anthropic's current plan guidance before making a purchase decision.

GitHub moved Copilot usage to AI Credits on June 1, 2026. The useful takeaway is narrower than “every plan became pay-as-you-go”: AI coding cost now mixes subscription allowances, shared limits, model-specific credits, and optional metered overages.

Why the total is hard to explain

Ask a heavy Claude Code user what their AI setup costs per month and you'll usually get the subscription price. That may be the cash charge, but it does not explain capacity consumed or whether separate API/provider charges exist. Those facts live in different places: a subscription meter here, a provider report there, a second tool with its own credit pool, and local agent-session evidence.

The raw material for a useful comparison already exists — Claude Code and Codex write detailed session logs to your machine as you work. Pricing supported token usage at published API-equivalent rates provides one input to a plan decision. It does not prove incremental spend, remaining entitlement, or what a subscription covers; combine it with detected or declared plan context and provider-reported limits when those limits are available.

Check your local evidence

npx aibill reads the session logs already on your machine—locally, with no signup, provider connection, or upload—and shows:

• Observed API-equivalent value — supported local usage priced at published rates, broken down by project and model where the transcript exposes them.
• Plan context and comparison — API-rate value beside a detected or user-declared plan label. It is comparison math, not proof of plan coverage or the cheapest option.
• Evidence-ranked action candidates — local transcript aggregates are treated as observed exposure, not invented monthly savings. Apply asks the coding agent to inspect the exact source, propose one reversible change, wait for approval, and compare matched future sessions. A dollar result requires a source-supported counterfactual and, for a cash claim, provider cost.
• Context inventory and invocation evidence — items that are discoverable, invoked, MCP-configured, explicitly requested as always-loaded, hook-injected, unmeasured, or invocation-unobservable. Configuration alone does not prove a schema payload was loaded; aibill reports no matching invocation only where the transcript supports it.

Local estimate vs provider report

Numbers derived from local logs are estimates at published API rates — useful for comparisons and cost investigations, but not a bill. If you want official provider-reported cost, connect a user-referenced org admin/owner credential with the required read permissions. aibill keeps that report beside—not merged into—the local estimate. A final invoice can still include credits, discounts, tax, or adjustments. Anything the source cannot establish remains labeled accordingly.

Local-first, because it's your evidence

Default transcript analysis runs on your machine with no signup: transcripts, prompts, file names, and dollar amounts are never uploaded. The CLI's one disclosed signal is anonymous command counts — no code, no prompts, no file names — announced by a printed notice before the first byte and ended by aibill telemetry off or DO_NOT_TRACK. A deliberate provider connection sends the referenced credential only to that provider's official API; an explicit MCP result goes only to the AI client you invoked. The code is MIT-licensed and open source. aibill never sits in the inference path and never stores, prints, or proxies provider credentials. The meters are multiplying; know which number you are looking at before acting on it.

Correction (August 24, 2026): an earlier version of this post said the default run involves no telemetry. Since v0.9.2 the CLI sends disclosed, anonymous command counts as described above.