GitHub Copilot, one week in: real developer bills since AI Credits went live

What developers are actually posting

The June 2 math piece worked from the published rate card. One week of real bills tells you what those rates produced when developers ran their normal workflows. The reports below are the ones with attribution we could verify in primary or first-tier coverage. Paraphrased numbers floating in third-party recaps without source links are excluded.

“This is a staggering shift from a ‘predictable subscription’ to a ‘stressful meter-based’ service.”

Posted by an anonymous developer on the GitHub community forum and quoted in The Register's June 2 coverage. The same user reported burning roughly 8% of their monthly allocation in two hours of regular development work. Project that pace and a 7,000-credit Pro+ pool depletes inside two days of sustained activity.

“Not after a day of usage… After ONE request.”

A second anonymous user on the same thread, reporting over $6 in credits charged for a single project change request. At Claude Opus 4.5 through 4.8 rates of $5 per million input and $25 per million output, $6 corresponds to roughly a 50,000-token input and 220,000-token output interaction. That is one large agentic refactor producing one substantial generated file. Realistic for an Opus-class one-shot.

“1,180 credits used. 16% of my monthly Pro+ allowance. Gone. For basically nothing.”

Posted on Reddit and cited by The Register, from a user testing Claude 4.8 on website editing work. At $11.80 against Opus 4.8 rates ($5 in / $25 out per million tokens), the math lands in the 80,000-input plus 450,000-output range. That is a multi-file generated artifact: think a website edit that touches several pages, each producing 20,000 to 40,000 tokens of revised code. The session is the kind of thing many developers ran weekly under the premium-request model without ever seeing the underlying token cost. The credit system makes it visible for the first time.

“I've opted to stick to Pro+, burn through my allocated credit in a week, and then pivot to using OpenRouter.”

Quoted in The Register's June 2 piece from the same GitHub community thread. The structural observation worth flagging: this user is not leaving Copilot because the Pro+ subscription got more expensive. The base price did not change. They are leaving because they want to move the meter to a place where they can see and control it directly. OpenRouter charges the provider's per-token rate with no credit-pool abstraction.

On the same thread, a separate Chinese-language post reported four code agents consuming roughly 53% of a Pro+ pool (3,707 of 7,000 credits) on day one. We are not rendering that one as a direct English quote, but the number is consistent with what the rate card predicts for sustained Opus or GPT-5.5 use across multiple parallel agent sessions.

The single-request problem

The pattern across the week-one reports is not gradual depletion. It is single-call burns that consume 10% to 16% of the monthly pool in one shot. The math piece explained why this is structurally possible. One week of bills makes it concrete.

Three model-rate facts drive most of the variance. First, GPT-4.1 and GPT-5 mini are included on paid plans and consume zero credits regardless of how much you use them. Second, Claude Sonnet 4.6 lands at $3 input and $15 output per million tokens, meaning a 10,000-input plus 5,000-output chat costs roughly 10 credits or 10 cents. Third, Claude Opus 4.5 through 4.8 and GPT-5.5 sit at the top of the rate sheet with output rates of $25 and $30 per million tokens respectively (both at $5 per million input), meaning a single agentic refactor producing 200,000 to 500,000 output tokens can consume 600 to 1,200 credits in one call.

Before June 1, every model call counted as one premium request regardless of input or output size. A 50,000-token Opus refactor and a one-line Sonnet completion looked identical on the meter. The credit system makes the difference legible. Developers who built habits around model-agnostic billing are now seeing the per-call cost for the first time. Many of the “sticker shock” reports are not bills that got higher than before. They are bills that became visible for the first time.

That said, three categories of user are genuinely paying more. Heavy agentic users who relied on Opus or GPT-5.5 every day are paying more, because the included pool covers fewer of those calls than the old premium-request system did. Users who picked premium models for short chats out of habit are paying more, because they were not paying for the model selection before. And users running parallel agents (multiple Copilot agents working on the same repo or different repos in the same plan) are paying multiples of single-developer use without admin-level budget controls.

The migration paths users are citing this week

The thread surfaces five common alternatives. The honest framing matters: most of them just move the meter to a different vendor.

Four of the six paths still have a meter. Only the last two genuinely remove it. The distinction matters because moving to a different platform meter does not solve the underlying problem (you cannot predict the bill without modeling your token consumption). It just changes which dashboard you check.

What one week of bills tells you that the rate card could not

Reading the rate card on June 1, you could calculate the cost of any individual call if you knew the input and output token counts. Most developers did not. One week of bills tells you three things only visible after the meter ran.

One. Heavy agentic users deplete in days, not weeks, even on Pro+. The 7,000-credit pool covers about 10 to 12 sustained Opus sessions of the kind the “1,180 credits” report describes. After that, you are paying overage or rationing model choice. Max at $200 in credits stretches the runway to roughly two weeks of comparable use.

Two. The lock-in is not the rate card. It is the workflow you built under the old pricing. Agentic refactors became normal because the meter did not exist. Now the same workflow costs what it costs. The friction in the migration threads is not about $6 per call. It is about retrofitting habits that were free for two years.

Three. For users who live in inline completions and lightweight chat, the change is invisible. Code completions and Next Edit suggestions stay unlimited on every paid plan and do not consume credits. GPT-4.1 and GPT-5 mini chats cost zero. Pro at $10 per month covers the same workflow it did on May 31.

Before extrapolating wildly from week one, check your July 1 bill. The reports above are real, but they describe a specific user profile. If your day is mostly inline completions and short chats, the system did not touch you. If you ran an Opus refactor every morning, it did.

Three off-ramps that actually remove the meter

For the user profile that the credit system genuinely hurts (heavy agentic, premium-model, multi-call workflows), three off-ramps remove the meter instead of moving it.

1. A BYOK extension. Cline and Continue.dev run as VS Code extensions in bring-your-own-key mode. You bring your own API key, pay your model provider directly, and the meter lives at the provider not the platform. No credit pool, no flex allotment, no transition window. Both are free under Apache 2.0. The tradeoff is that the IDE host stays the same; you are still working inside VS Code.

2. A local model with zero inference cost. Run a coding-class open-weight model via Ollama, LM Studio, or llama.cpp and the meter goes away entirely. Qwen3-Coder, DeepSeek V4, GLM-4.7 Thinking, and Kimi K2.6 Thinking all run locally on consumer hardware in 2026 and post competitive scores on coding benchmarks. The tradeoff is hardware: a credible local-coding setup wants 16 to 48 GB of VRAM, and the highest-performing open-weight models want more. The VRAM calculator sizes this for your hardware.

3. A free or one-time-purchase IDE. Bodega One Code is currently free for everyone in open beta, including commercial use. At full release, Personal stays free for personal use on one machine. Pro will be $39 one-time for commercial use on two machines. There is no monthly renewal at any tier, no credit pool, no flex allotment, and no premium-model paywall. Standalone desktop app, Monaco editor, autonomous coding agent, 10+ provider presets via BYOLLM, and an air-gap mode for regulated environments where the code cannot leave the machine. Point it at your own API keys for Anthropic, OpenAI, Groq, OpenRouter, or any OpenAI-compatible endpoint, or point it at a local Ollama or LM Studio instance for zero-cost inference. Download free.

Between the three, the choice is really about how far you want to move from the GitHub Copilot starting point. A BYOK extension keeps your editor and your habits and just changes the billing. A local model removes the inference bill entirely but asks you to manage the model. A new IDE replaces the front door but removes the meter at every layer.

For the full side-by-side breakdown of features, training-data policy, and architecture, read our GitHub Copilot comparison page. For the published rate card the credit math derives from, read the June 2 math piece. For the local-model picks that actually run on your hardware, start at /local-llms.

Sources

• GitHub Copilot models and pricing (rate authority, 1 AI credit = $0.01 USD, full model list and per-token rates): docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
• Updates to GitHub Copilot billing and plans (June 1 changelog, sign-up-pause verbatim, Actions-minutes for Copilot code review): github.blog/changelog/2026-06-01-updates-to-github-copilot-billing-and-plans
• GitHub community discussion #192948 (958 down vs 24 up, 534 comments, 845 replies, thread locked June 1): github.com/orgs/community/discussions/192948
• The Register, June 2, 2026: Angry devs vow to flee GitHub Copilot as metered billing takes hold (source of four verified developer quotes including the 1,180-credit and $6-per-request reports): theregister.com
• gHacks, June 2, 2026: GitHub Copilot Usage-Based Billing Takes Effect, Drawing Developer Backlash Over Rapid Credit Depletion: ghacks.net
• MLQ News, June 1, 2026: GitHub Copilot Switches to Token-Based Billing June 1, Drawing Developer Backlash: mlq.ai
• dev.to, June 2, 2026 (Anand Karanjkar): GitHub Copilot Token Billing 2026: Full Cost Guide and Alternatives (source of the by-user-tier monthly cost estimates): dev.to/akaranjkar08
• Bodega One Code, June 2, 2026: GitHub Copilot AI Credits, decoded (the rate-card math piece this post builds on): bodegaone.ai/blog/github-copilot-ai-credits-math-explained