Evidence correction, 14 October 2026: earlier versions of this post said the three billing changes landed within 30 days, overstated what the Claude credit covers, attributed figures to sources that do not contain them (a 15x to 30x Claude subsidy, per-session token counts, Copilot credits gone in hours, a $29 to $750 bill credited to The Register), credited arXiv 2604.22750 to Stanford and Microsoft Research, and misread Datadog's caching figure. They also got the routing example's arithmetic wrong, included unsourced savings and quality percentages, and did not note that Anthropic paused the Claude credit change on June 15, 2026. Those claims are removed or corrected below.
Three billing changes in three months. The flat-rate era is officially over.
On June 1, 2026, GitHub flipped every Copilot plan to usage-based billing. Instead of counting "premium requests," each plan now includes a monthly allotment of AI Credits pegged to token consumption. Copilot Pro+ gets $39 in credits. Business gets $19 per seat. Enterprise gets $39 per seat. When the credits run out, you pay per token or you stop.
Source: GitHub Blog, "GitHub Copilot is moving to usage-based billing," April 27, 2026
Source: GitHub Changelog, "Updates to GitHub Copilot billing and plans," June 1, 2026
On June 15, 2026, Anthropic moves Agent SDK and headless (claude -p) usage off the shared subscription limits and onto a separate monthly credit: $20 for Pro, $100 for Max 5x, $200 for Max 20x. Interactive use stays on the plan's normal limits. The credits meter at full API rates. No rollover. Update: Anthropic paused this change on June 15, 2026, the day it was due to take effect, so for now Agent SDK and claude -p usage still draw from subscription limits. The pause is described as temporary.
Source: InfoWorld, "Anthropic puts Claude agents on a meter across its subscriptions," May 14, 2026
And in April, OpenAI updated Codex pricing to align with API token usage instead of per-message billing, extending the change to all Enterprise plans by April 23.
Source: OpenAI Help Center, "Codex rate card"
Three platforms. Three billing overhauls. The same direction: every token now has a visible price tag.
The subsidy math never worked at scale.
The reason all three moved within a few months is straightforward. Agentic coding workflows consume orders of magnitude more tokens than chat, and the flat-rate subscriptions were never priced for that volume.
One analysis cited by TechTimes estimated that OpenClaw users got about $236 a month of API-equivalent compute from a $20 Pro plan, roughly 12 to 1. At chat-era usage levels, the subsidy was invisible. At agentic usage levels, it broke the model.
Source: TechTimes, "Anthropic Ends Subscription Subsidy for Agents June 15," June 2, 2026
GitHub saw the same dynamic. One developer told The Register they used about 8% of a monthly allowance in two hours. One Redditor quoted by TechCrunch projected their bill rising from about $29 to nearly $750 a month.
Flat-rate pricing worked when AI coding meant autocomplete. One study of agentic coding found that agentic tasks consume about 1,000x more tokens than code reasoning and code chat tasks, with input tokens driving most of the cost. The subscription model could not absorb that volume at those prices.
$20 in Claude credits runs out faster than you think.
At Sonnet 4.6 pricing ($3 per million input tokens, $15 per million output tokens), $20 covers roughly 6.6 million input tokens or 1.3 million output tokens. That sounds like a lot until you measure what a coding agent session actually consumes.
CloudZero estimates that a serious agentic session, with a big context and many turns, lands around $5 to $15 in API terms. At that rate, a $20 credit covers only one to four such sessions ($20 / $15 to $20 / $5).
For teams running CI pipelines, automated PR reviews, or scheduled agent tasks through the Agent SDK, the credit can run out well before the month ends. The overflow option exists: enable "usage credits" and pay at standard API rates for anything above the monthly allowance. But that is exactly the scenario where per-token cost optimization matters most.
Source: FindSkill.ai, "Claude Code Pricing After June 15: The Decision Table"
GitHub's credit math is similar. 1 AI Credit equals $0.01 USD. GitHub's docs now list 7,000 credits a month for Copilot Pro+, 3,900 base and 3,100 flex. They also warn that a complex agentic session uses far more credits than a quick chat question.
Source: GitHub Docs, "Usage-based billing for individuals"
Most coding agent calls do not need a frontier model.
The critical insight is not that metered billing is expensive. It is that most of the metered tokens are going to the wrong model.
Augment Code published a routing guide in 2026 showing that three-tier Claude routing saves 51% compared to uniform Opus deployment. Their breakdown maps the coding agent workflow to model tiers:
- Opus 4.6 for coordination decisions that cascade through every downstream agent
- Sonnet 4.6 for high-volume implementation work (79.6% SWE-bench score at $3/MTok)
- Haiku 4.5 for the hundreds of file operations that need speed over reasoning depth
- GPT-5.2 for async code review that benefits from exhaustive investigation
Source: Augment Code, "Best AI Model for Coding Agents in 2026: A Routing Guide"
On SWE-bench Verified, Sonnet 4.6 scores 79.6%, within 1.2 points of Opus 4.6, per Augment Code. Augment still assigns Opus to coordination, where it reports the reasoning gap is largest.
This matches the broader production data. Datadog measured that 69% of all LLM input tokens are system prompts, tool schemas, and policy definitions that repeat on every call. In a May 2026 press release covering 2.4 billion API calls on its platform, API aggregator AI.cc reported a median $2.31 per million tokens for customers using multi-model routing, versus $18.40 for frontier-only deployments.
Source: Datadog, "State of AI Engineering 2026"
When every token was free (or felt free under a flat-rate plan), nobody measured this. Now that every token has a price, the waste is visible.
The new math: what routing does to your metered bill.
Let us run the numbers on a developer using the full $200 Agent SDK credit on a Max 20x plan, assuming every call has the same token mix.
Without routing (current default): All agent calls go to Opus 4.8 at $5/$25 per million tokens. Total: $200/month.
With three-tier routing:
- 60% of calls to Haiku at $1/$5: $24/month ($120 of Opus-priced usage at one fifth the rate)
- 30% of calls to Sonnet at $3/$15: $36/month ($60 of Opus-priced usage at three fifths the rate)
- 10% of calls stay on Opus at $5/$25: $20/month
- Total: $80/month. Savings: $120/month per developer.
That $200 credit now covers 2.5 times the usage ($200 / $80). Or the same usage costs $80 instead of $200. For a 50-developer team, the monthly savings are $6,000 (50 x $120).
The same math applies to Copilot. A Pro+ credit allowance goes further when simple file reads and boilerplate generation route to a cheaper model instead of burning credits at frontier rates. GitHub's credit system charges based on the model's API price, so routing to a cheaper model directly reduces credit consumption per request.
For teams on overflow billing (paying API rates above the monthly credit), the savings are even larger. Every request above the credit threshold is pure API spend. Routing those overflow requests captures the full 5x spread between Haiku and Opus on every call.
Cursor already figured this out.
Cursor moved to usage-based billing in June 2025, a full year before GitHub and Anthropic followed. Their response was architectural: they built Composer 2.5, an in-house coding model on Moonshot's open-weight Kimi K2.5 checkpoint. It lists at $0.50/$2.50 per million tokens against $5/$25 for a frontier model, a 10x difference. Cursor's pricing docs now give its own models, including Composer 2.5, a separate pool with significantly more included usage than third-party frontier models.
Source: Cursor Blog, "June 2025 pricing," June 2025
Source: Cursor Blog, "Introducing Composer 2.5," May 18, 2026
Source: Cursor Docs, "Models & Pricing"
The logic is clear. When every token is metered, the vendor that routes to the cheapest capable model gives the user the most value per dollar. Cursor built the routing into the product. The question for Claude Code and Copilot users is whether they wait for Anthropic and GitHub to do the same, or route at the API layer now.
Cached tokens are the other half of the equation.
Every coding agent provider now offers cached token pricing, and the discounts are substantial. Anthropic charges roughly 10% of the standard input rate for cached tokens. OpenAI offers similar discounts on Codex cached inputs.
Datadog's finding that 69% of input tokens are repeated system prompts means caching can do a lot of work. If those repeated tokens are all cache hits at a 90% discount, caching cuts total input cost by about 62% (0.69 x 0.90).
Source: Datadog, "State of AI Engineering 2026"
Yet Datadog found that only 28% of LLM calls showed any cached input tokens. Now that every cached token saves real money against a visible credit balance, expect adoption to accelerate. But caching only reduces the cost of repeated content. Routing reduces the cost of new content. You need both.
Five things to do before June 15.
1. Measure your per-session token consumption. Pull your API logs or Claude Code usage data. How many tokens does a typical coding session consume? How much of that is system prompts, tool schemas, and file reads versus actual reasoning? The answer determines how much metered billing will cost you.
2. Classify your agent calls by complexity. Tag a sample of 100 coding agent interactions. File reads, linting, formatting, simple refactors, test generation from templates: these are Haiku-tier tasks. Architecture decisions, complex debugging, multi-file refactors with dependency analysis: these need Opus. Count how many land in the cheap tier.
3. Enable prompt caching. If your coding agent sends the same system prompt, tool schemas, or project context on every call, you may be paying full price for most of your input tokens. OpenAI caches prompt prefixes automatically, and Anthropic caches them once you mark the cacheable prefix with cache_control. Enable it before metered billing starts.
4. Route per call, not per application. Stop hardcoding model selection. A classifier that evaluates each agent call independently and routes to the cheapest capable model captures the 5x pricing gap between Haiku and Opus on every task the cheap model can handle. This is not a workflow change. It is a base URL change.
5. Set overflow alerts. If you enable usage credits (overflow billing) on Claude Code, set a monthly cap. Without a cap, overflow usage is billed at API rates with no ceiling.
The market is moving toward routing as default.
IDC predicts 70% of top AI enterprises will use dynamic model routing by 2028. The metered billing shift accelerates this timeline. When every token was subsidized, routing was an optimization. Now that every token is metered, routing is cost control.
Source: IDC Blog, "The future of AI is model routing," November 17, 2025
The enterprise AI cost crisis is well documented. Fortune reported that Uber used up its 2026 AI coding tools budget in four months. Microsoft began cancelling most internal Claude Code licenses in its Experiences and Devices division, according to The Verge. An AI consultant told Axios that one unnamed client spent about $500 million on Claude in a single month. The FinOps Foundation reports 98% of teams now manage AI spend, up from 31% two years ago.
Source: The Verge, Microsoft is discontinuing most internal Claude Code licenses, May 14, 2026
Source: Axios, "AI sticker shock hits corporate America," May 28, 2026
Source: FinOps Foundation, "State of FinOps 2026 Report"
Metered billing does not create the cost problem. It makes the cost problem visible. And visibility is the prerequisite for optimization.
Where Nadir fits.
Nadir can sit beside the coding agent's existing gateway and return the model, reasoning effort, cache, context, and policy for each call. Start in shadow mode and report outcomes before enforcement. The older 60% / 98% RouterBench result is a reference-assisted research ceiling, not deployed performance.
The integration is two lines: change the base URL, set model="auto". Complete non-streaming responses report the served model and actual cost; a configured, priced benchmark adds the avoided-spend comparison.
Anthropic's announced Agent SDK credit is $20 to $200 a month, if it takes effect. Your Copilot plan has a fixed monthly credit allowance. Your Codex budget is whatever your team approved. Routing is the lever that makes those credits cover more work. Not by using AI less. By using the right model for each call.
Sources: [GitHub Blog, "GitHub Copilot is moving to usage-based billing"](https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/) (April 27, 2026). [GitHub Changelog, "Updates to GitHub Copilot billing and plans"](https://github.blog/changelog/2026-06-01-updates-to-github-copilot-billing-and-plans/) (June 1, 2026). [The Register, "Angry devs vow to flee GitHub Copilot as metered billing takes hold"](https://www.theregister.com/ai-and-ml/2026/06/02/github-copilot-users-threaten-exit-as-metered-billing-kicks-in/5249826) (June 2, 2026). [TechCrunch, "'What a joke': GitHub Copilot's new token-based billing spurs consternation among devs"](https://techcrunch.com/2026/05/30/what-a-joke-github-copilots-new-token-based-billing-spurs-consternation-among-devs/) (May 30, 2026). [TechTimes, "Anthropic Ends Subscription Subsidy for Agents June 15"](https://www.techtimes.com/articles/317625/20260602/anthropic-ends-subscription-subsidy-agents-june-15-credit-pool-replaces-flat-rate-access.htm) (June 2, 2026). [InfoWorld, "Anthropic puts Claude agents on a meter"](https://www.infoworld.com/article/4171274/anthropic-puts-claude-agents-on-a-meter-across-its-subscriptions.html) (May 14, 2026). [DevOps.com, "Anthropic Hits Pause on Claude Agent SDK Billing Change, For Now"](https://devops.com/anthropic-hits-pause-on-claude-agent-sdk-billing-change-for-now/) (June 18, 2026). [OpenAI Help Center, "Codex rate card"](https://help.openai.com/en/articles/20001106-codex-rate-card). [Bai et al., "How Do AI Agents Spend Your Money?", arXiv:2604.22750](https://arxiv.org/abs/2604.22750) (April 2026). [Augment Code, "Best AI Model for Coding Agents in 2026: A Routing Guide"](https://www.augmentcode.com/guides/ai-model-routing-guide). [Datadog, "State of AI Engineering 2026"](https://www.datadoghq.com/state-of-ai-engineering/). [AI.cc press release, "Enterprise Token Costs Drop 67% Year-Over-Year"](https://www.einpresswire.com/article/911544568/aicc-report-enterprise-token-costs-drop-67-year-over-year-as-multi-model-ai-adoption-hits-record-high) (May 2026). [Cursor Blog, "June 2025 pricing"](https://cursor.com/blog/june-2025-pricing) (June 2025). [Cursor Blog, "Introducing Composer 2.5"](https://cursor.com/blog/composer-2-5) (May 18, 2026). [Cursor Docs, "Models & Pricing"](https://cursor.com/docs/account/pricing). [FindSkill.ai, "Claude Code Pricing After June 15"](https://findskill.ai/blog/claude-code-pricing-after-june-15-decision-table/). [CloudZero, "Claude Code Pricing In 2026"](https://www.cloudzero.com/blog/claude-code-pricing/). [GitHub Docs, "Usage-based billing for individuals"](https://docs.github.com/en/copilot/concepts/billing/usage-based-billing-for-individuals). [IDC Blog, "The future of AI is model routing"](https://www.idc.com/resource-center/blog/the-future-of-ai-is-model-routing/) (November 17, 2025). [Fortune, "Uber burned through its entire 2026 AI budget in four months"](https://fortune.com/2026/05/26/uber-coo-ai-spending-tokens-claude-code/) (May 26, 2026). [The Verge, Microsoft discontinuing internal Claude Code licenses](https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad) (May 14, 2026). [Axios, "AI sticker shock hits corporate America"](https://www.axios.com/2026/05/28/ai-spending-roi-enterprise-costs) (May 28, 2026). [FinOps Foundation, "State of FinOps 2026 Report"](https://data.finops.org/) (2026). Anthropic, OpenAI, GitHub model pricing as of June 2026.