On this page · 9 sections
Summary. The old question was simple: pay a flat seat for GitHub Copilot or Cursor, or bring your own API key (BYOK) and pay the model provider per token. In 2026 that framing broke. On June 1, 2026, GitHub moved every Copilot plan to usage-based billing, so a Business seat is still $19 per user per month but now includes exactly $19 of GitHub AI Credits, with overage billed at each model's published API rate. Cursor did the same in spirit, splitting its $40 Teams seat into separate first-party and third-party usage pools. So both a $19 Copilot Business seat and a raw Anthropic key now bill your Claude tokens at the same underlying rate, roughly $2 per million input and $10 per million output for Claude Sonnet 5 during its introductory period through August 31, 2026. The seat price is no longer the number that decides your bill. Usage intensity and model choice are. This guide gives engineering leads the actual math: a worked cost model across light, medium and heavy developers, a decision table, and the governance costs that never show up on the pricing page.
This is written for the person signing off on a dev-tools budget, not for a pricing-page skim. The numbers below use published seat prices and API rates as of July 2026, plus a transparent token-consumption model you can re-run with your own telemetry.
What actually changed in 2026
For two years the mental model was "subscription equals predictable, BYOK equals cheap but fiddly." Both halves are now wrong.
GitHub retired premium request units and replaced them with GitHub AI Credits on June 1, 2026. Every plan includes a monthly credit allotment equal to its price, and usage is metered by token consumption, input, output and cached, at the listed API rate for whichever model you pick. GitHub's own framing is blunt about why. As GitHub Chief Product Officer Mario Rodriguez wrote when announcing the change, "Today, a quick chat question and a multi-hour autonomous coding session can cost the user the same amount," and that was no longer sustainable once agentic sessions became the norm. Base seat prices did not move: Copilot Pro stays $10 per month, Pro+ $39, Business $19 per user, and Enterprise $39 per user, each bundling an equal dollar value of credits. Code completions and Next Edit suggestions stay included and do not draw down credits.
Cursor moved the same direction. Its plans in July 2026 are Free, Pro at $20 per month, Business or Teams at $40 per user, and an Ultra tier at $200 per month for developers running agents all day. The Teams seat now separates usage into a first-party pool for Cursor's own Composer models and a third-party pool for Claude, GPT and Gemini calls, each metered.
The consequence is the headline of this article. Once your usage crosses the credit that comes bundled with the seat, a subscription and BYOK bill the same tokens at the same provider rate. What you are really buying with a seat above that line is tooling, not cheaper inference.
The three ways to pay, defined
There are three real options in 2026, and most teams end up on a mix.
The first is a bundled subscription: a seat in GitHub Copilot or Cursor that includes a token budget and bills overage at API rates. You get the integrated editor experience, completions, code review and agent modes, plus admin budget controls, and you accept the vendor's model catalogue.
The second is pure BYOK: an open tool such as Cline, Aider, Kilo Code, OpenCode or Goose (avoid Roo Code, which was archived in May 2026), pointed at your own provider key. You pay Anthropic, OpenAI, Google, AWS Bedrock or Azure OpenAI directly, with no tool markup, and you can route to any model including cheap or self-hosted endpoints via Ollama. GitHub also supports a BYOK path: admins connect keys for Anthropic, OpenAI, xAI, Microsoft Foundry, Bedrock or Google AI Studio, and that usage is billed by the provider rather than against Copilot credits, which lets teams spend down an existing committed-use contract.
The third is hybrid: a paid seat for the polished day-to-day experience, plus a BYOK agent for heavy or specialised work billed to a provider contract. This is where most funded teams land, and it is why "subscription or BYOK" is a false binary.
| Decision vector | Bundled subscription | Pure BYOK | Hybrid |
|---|---|---|---|
| Unit of billing | Seat + metered credits | Provider tokens only | Seat for some devs, tokens for others |
| Predictable monthly floor | Yes, the seat price | No, usage-driven | Partly |
| Marginal cost per token | Provider API rate after credits | Provider API rate, no markup | Provider API rate |
| Tooling and UX | Best, fully integrated | Varies, often CLI-first | Best where it matters |
| Data control and residency | Vendor's terms | Your provider contract and region | Mixed, needs policy |
| Admin governance and budget caps | Built in | You build it | Partial, two systems |
| Best for | Light users, teams wanting one bill | Heavy users with provider contracts | Most mid-to-large teams |
The token math: a worked model
Seat price stops mattering the moment token cost dominates, so the only honest way to compare is to model token consumption. Agentic coding is token-hungry because every step re-sends context. The figures below are a model, not a measurement: plug in your own numbers from a week of billing telemetry. They assume Claude Sonnet 5 at its introductory $2 per million input and $10 per million output, and Google's Gemini 3 Flash at roughly $0.50 per million input and $3 per million output, both as listed in July 2026.
| Developer profile | Tokens per month (input / output) | Cost on Gemini 3 Flash | Cost on Claude Sonnet 5 | Covered by a $19 seat credit? |
|---|---|---|---|---|
| Light (completions, occasional chat) | 5M / 0.5M | about $4 | about $15 | Yes, on either model |
| Medium (daily agent use) | 30M / 3M | about $24 | about $90 | Only on Flash |
| Heavy (agent sessions most of the day) | 120M / 12M | about $96 | about $360 | No |
| Very heavy (parallel agents) | 300M / 30M | about $240 | about $900 | No |
Three things fall out of this table, and they are the whole decision.
First, for a light user the bundled seat is the cheapest option, not the most expensive. A $19 Copilot Business seat includes $19 of credits, which covers roughly $15 of Sonnet usage, and completions are free on top. Going BYOK to save money here is a false economy once you count the tooling you give up.
Second, for a medium user on a premium model, the seat's included credit is irrelevant. A developer burning $90 of Sonnet tokens a month pays the $19 seat plus about $71 of overage at the same API rate BYOK would charge. The subscription did not make the tokens cheaper. It bought the editor and the admin controls. That can be worth $19. It is not worth pretending the token cost went away.
Third, for heavy users the lever is model and caching, not seat versus BYOK. Moving that heavy developer from Sonnet to Gemini 3 Flash cuts the monthly token bill from about $360 to about $96, a far bigger saving than any seat decision. Prompt caching, which bills cached input at a fraction of the standard rate, moves the number again. This is where a documented routing policy pays for itself, and it is the same discipline we describe in our LLM hybrid routing and API-spend decision framework.
When BYOK wins, and when it does not
BYOK wins on cost in three specific situations. If you already hold a committed-use or enterprise contract with Anthropic, OpenAI, Google or a cloud provider, BYOK lets your coding tools draw down that commitment at your negotiated rate instead of paying a second vendor. If your team is heavy enough that tokens dominate the bill, removing even a small tool markup and routing to a cheaper model matters. And if you need data residency or contractual data-handling that a seat vendor will not give you, your own provider contract in your own region is the only way to get it.
Subscriptions win everywhere else. For light and many medium users, the bundled credit plus free completions is genuinely cheaper than assembling a BYOK stack, and the integrated experience is better. For teams that do not want to run billing operations, reconcile provider invoices, or build their own budget guardrails, a seat with admin caps is less total cost of ownership even at a slightly higher unit price. The right question is not "which is cheaper per token," because past the bundled credit they cost the same. It is "who should own the token bill, the tool vendor or us." For a deeper view of squeezing model spend itself, our guide on Claude Opus 5 effort levels to cut token spend and the piece on GPT-5.6 inference cost for enterprise AI go under the hood.
The costs that never show on the pricing page
The per-token rate is the easy part. The costs that decide real total cost of ownership are the ones a pricing page hides.
Governance and budget control is the first. A seat vendor gives you admin budget caps, pooled credits across the organisation and per-user limits out of the box. GitHub, for example, lets admins set budgets at the enterprise, cost-center and user level and pool unused included usage so it is not stranded. With pure BYOK you build that yourself, or you discover the gap when one engineer's runaway agent posts a five-figure provider bill.
Billing operations is the second. BYOK means provider invoices to reconcile, keys to rotate, and usage to attribute back to teams. That is real finance and platform work, and it is the reason many organisations accept a tool markup they could technically avoid.
Data control is the third, and for regulated teams it is decisive. When code and context leave your environment for a third-party model, you inherit that provider's data-handling and retention terms. Under India's Digital Personal Data Protection Act, 2023 (DPDP Act), a team handling personal data in its codebase or test fixtures has to be able to say where that data went and under what terms. A BYOK contract in a specific region can answer that; a default seat may not. Rate limits and model access round out the list: a raw provider key can hit throttles a managed seat smooths over, while BYOK can reach models a seat vendor has not yet added.
India-specific considerations
For an Indian engineering team, the seat-versus-BYOK decision also has a currency and a compliance dimension. A ten-developer team on Copilot Business seats runs about $190 a month, roughly ₹16,000, before any token overage. The same team going heavy on Claude Sonnet 5 could add several hundred dollars of tokens on top, which is why the model-routing lever matters more here than the seat choice. Converting the heavy-user row above, a single agent-all-day developer on Sonnet is about ₹30,000 a month in tokens, versus about ₹8,000 on Gemini 3 Flash.
The compliance dimension is DPDP. If your repositories or prompts carry personal data, sending them to a model API is a cross-border processing decision, and the fiduciary, your company, keeps the liability. BYOK with a provider region and contract you control makes that defensible in a way a consumer-grade seat may not. Teams weighing the broader build decision often read our Claude Opus 5 coding agents cost versus Fable analysis alongside this one, and both sit under our wider comparison of Gemini 3.5 Pro versus GPT-5.6 versus Claude Fable 5.
FAQ
How eCorpIT can help
eCorpIT helps engineering teams model and control what their AI coding stack actually costs. We instrument a week of real token telemetry, build the seat-versus-BYOK break-even for your usage, set up model-routing and prompt-caching policies, and put budget guardrails on both subscription and provider-key spend. As a CMMI Level 5 and ISO 27001:2022 certified organisation, we design these workflows aligned with DPDP Act data-handling requirements. If you want the numbers for your own team rather than a generic model, talk to our engineering team.
References
_Last updated: July 28, 2026._
