Insight / Tool

Flat-rate coding has ended.

Coding assistance now behaves like cloud infrastructure: model choice, context size, agentic sessions, and developer habit turn seat subscriptions into metered intelligence.

Interactive cost modelAI tooling economicsLast reviewed May 2, 2026

Exposure model

Estimate the next-quarter bill before the invoice arrives.

Start with the old seat bill, then model how quickly usage intensity and frontier model selection can turn that line item into variable compute exposure.

Tool context

Seat plan

Usage intensity

Token bands model coding agents as multi-step loops. A single substantial feature request can consume millions of input/cache tokens before producing a comparatively smaller output, especially when the agent repeatedly reads files, tool results, and prior context.

Model mix

Projected exposure

82 active developers modelled from 120 licensed seats at 68% usage.

Previous quarter

$7,200

Seat plan

$20 / seat

Next-quarter bill

$182,479

+$175,279 vs last quarter (+2,434%).

Per active developer

$742 / month

Base case for GitHub Copilot, standard usage, frontier-led model mix.

Sensitivity range

$135,034 - $244,522

The strongest pressure in this scenario is usage intensity.

Token model

31.4B tokens over 65 working days.

$34/ active dev / day

Input$87,737
Output$70,190
Cached$4,679

Gross

$182,119

Included

$6,840

Net add

$175,279

Assumption basis

Copilot moves to usage-based billingAll Copilot plans transition on June 1, 2026, with usage calculated from input, output, and cached tokens.Model pricing and AI CreditsAdditional Copilot usage is billed in AI Credits at per-token model rates, where 1 AI Credit equals $0.01.

Full evidence ledger and source links are below the model.

The estimate intentionally uses transparent scenario multipliers. Exact vendor invoices depend on plan terms, model routing, cached-token behavior, code review usage, and any negotiated enterprise discounts.

What changed

The subsidy is becoming visible.

01Copilot is becoming metered infrastructure.From June 1, 2026, GitHub Copilot moves from premium request counting to token-based AI Credits across input, output, and cached tokens.GitHub Blog: Copilot moves to usage-based billing02Frontier models change the unit economics.In GitHub pricing, GPT-5.5 is listed at $5 per 1M input tokens and $30 per 1M output tokens, compared with GPT-5.4 at $2.50 and $15.GitHub Docs: Model pricing and AI Credits03Claude Enterprise exposes usage separately.Anthropic describes Enterprise seats as access, with every token consumed across Claude, Claude Code, and Cowork billed separately.Anthropic Help: Enterprise usage billing

The implication

Finance cannot govern what the organisation cannot see.

The old seat-price mental model made AI coding tools feel like productivity software. The new pricing reality makes them look much more like cloud infrastructure: model selection, context size, agent loops, cached tokens, code review, and developer habit all affect the bill.

That does not mean the answer is to use weaker tools. Frontier models can be worth the premium when the work is complex, high-stakes, or time-sensitive. The operational problem is using those models by default, without any shared view of who is using them well, who is burning context inefficiently, and which workflows deserve the additional spend.

This is where the cost question becomes a capability question. AI fluency starts with judgment. A capable user chooses a suitable model and frames the problem clearly. They spot weak output and stop before useful exploration becomes waste.

See how Dojo frames AI fluency

Sources

Evidence base

The pricing shift is already visible in vendor documentation.

Last reviewed: May 2, 2026

GitHub Blog2026Copilot moves to usage-based billingAll Copilot plans transition on June 1, 2026, with usage calculated from input, output, and cached tokens.GitHub Docs2026Model pricing and AI CreditsAdditional Copilot usage is billed in AI Credits at per-token model rates, where 1 AI Credit equals $0.01.GitHub Docs2026Premium requests and multipliersPremium request usage varies by feature and model, with frontier models consuming materially more allowance.Anthropic Help2026Enterprise usage billingEnterprise seats provide access, while Claude, Claude Code, and Cowork usage is billed separately at standard API rates.Claude Code Docs2026Claude Code cost managementClaude Code team spend varies by usage pattern and can be governed through workspace tracking and limits.Claude Code Docs2026Agent usage trackingClaude Agent SDK exposes per-step input, output, cache read, cache creation, per-model, and cumulative cost data for agent loops.Anthropic Docs2026Prompt caching and token accountingCache read and cache creation tokens are reported separately and should be included when calculating total coding-agent input usage.Cursor Docs2026Cursor usage and Max ModeCursor usage depends on model selection, and Max Mode can use larger context windows up to 1M tokens for complex codebase work.OpenAI2026GPT-5.5 launch and pricingGPT-5.5 is priced above GPT-5.4 while being positioned as more capable and more token-efficient for many Codex users.

Next step

Cost exposure is a measurement problem.

Access is only the starting point. The harder question is who can use the tools with discipline. A capable user chooses a suitable model, frames the task clearly, and stops costly loops when the work stops improving.

Dojo measures that judgment instead of counting AI use.

Explore DojoTalk about AI fluency