One developer on GitHub's forums reported burning through 8 percent of their monthly GitHub Copilot Pro+ allocation in two hours under the new billing system. At that rate, a 7,000-credit monthly budget would last less than two days. That was June 1, the day GitHub switched Copilot from flat premium requests to token-metered AI Credits. Two weeks later, Anthropic dropped Claude Opus 5 at half the price of Claude Fable 5, with near-equivalent benchmark scores on agentic tasks. If your engineering team set its AI tooling budget before July, every number in that budget is wrong.
This is not a story about one model launch. It is about two structural changes - one to how frontier models are tiered, one to how coding tools are billed - landing in the same month. Together, they reshape the cost math for any team running agentic coding workflows.
What actually changed in Anthropic's model lineup
Fable 5 landed on June 9 and reset the frontier. Then Opus 5 launched on July 24 at near-Fable-5 quality - while Fable 5 itself went API-only on July 9. That two-step move matters more than either release alone.
Claude Opus 5 launched on July 24, 2026, at $5 per million input tokens, while Claude Fable 5 released on June 9, 2026, at exactly double that.
Anthropic says Opus 5 is the new state-of-the-art on coding and knowledge-work evaluations like Frontier-Bench and GDPval-AA, and on some benchmarks it actually beats Fable 5 outright - which flips the usual assumption that the pricier model always wins.
The practical consequence: the lineup now runs Opus 5 as the agentic-coding default, Fable 5 as the capability ceiling, Sonnet 5 for speed and cost, and Mythos 5 as the invitation-only security tier.
One API change buried in the Opus 5 release notes is worth flagging for teams building agents. Changing the tool list mid-conversation used to invalidate the entire prompt cache. Opus 5 lets you add or remove tools while keeping the cache, via a beta header. That matters for the cost of long-running agents.
There is also a gotcha for teams that disable thinking. With thinking disabled, Opus 5 sometimes writes tool calls out as body text - meaning the tool never runs while the turn still looks successful - and can leak internal XML tags into the output. If you are building agents, keeping thinking on and dialing effort down is the safer path.
How GitHub Copilot's billing shift changes the tool-stack math
GitHub Copilot's move to metered billing is a separate change, but it lands on top of the model pricing shift in a way that forces teams to think about both at once.
On June 1, 2026, all GitHub Copilot plans transitioned to usage-based billing. Instead of counting premium requests, every Copilot plan now includes a monthly allotment of GitHub AI Credits, with usage calculated based on token consumption - input, output, and cached tokens - at the listed API rate for each model.
One credit equals $0.01, and the output-token spread is roughly 40x across the model menu.
The asymmetry that catches teams off guard: code completions and next edit suggestions are unlimited on all paid plans and never consume credits. Only chat, CLI, agents, Spaces, Spark, and code review draw from the credit allowance. A developer who uses Copilot mainly for inline suggestions will notice almost nothing. A developer running multi-step agent sessions against a frontier model will hit the wall fast.
"This is a staggering shift from a 'predictable subscription' to a 'stressful meter-based' service," wrote one developer on GitHub's user forum, who said they burned through about 8 percent of their monthly AI Credits allocation in two hours on the new $39-per-month Copilot Pro+ plan.
The old model throttled you with premium request counts - hit the limit, and Copilot fell back to cheaper, lower-quality models. Annoying but survivable. The new model is fundamentally different: every feature except inline completions now draws from a monthly pool of AI Credits priced by token consumption. When the pool runs out, there's no fallback.
There is also a visibility gap.
GitHub added an ai_credits_used field to the Copilot usage metrics REST API on June 19, so you can pull per-user credit consumption programmatically. The catch: it is a per-user daily total only - not broken down by model, feature, or surface. If you need to know which model or which project burned the credits, the API will not tell you.
The model-vs-harness question that pricing obscures
Here is the insight that gets lost when teams focus on per-token costs: for long, messy coding tasks, the harness around the model often matters more than the model itself.
Model intelligence dominates short, clean tasks. Harness architecture dominates long, messy ones. A frontier model inside a thin terminal loop can be the wrong choice for a legacy migration. A slightly smaller model inside a mature harness with worktrees, checkpoints, and reliable resume can finish more real work - and consume fewer tokens along the way, since it does not re-read the full context after every failed step.
In the MemoryArena study, swapping an active memory agent for a long-context-only baseline dropped task completion from over 80% to roughly 45% on interdependent multi-session tasks. The gap between "has memory" and "does not have memory" is often larger than the gap between different LLM backbones. Investing in memory architecture can yield returns that rival - or exceed - model scaling.
The practical implication for agentic coding: before you decide whether to pay for Fable 5 over Opus 5, check whether your tool's harness supports parallel worktrees, session resume, and test-driven verification. The winning workflow is not autonomous magic. The practical workflow is a managed loop: specify the change, let the agent work, force evidence, run tests, review diffs, and preserve a rollback path. A better harness running a cheaper model will often beat an expensive model in a naive loop - and the token bill will show it.
Beagle in action#eng-tools, 10:22amWhat teams should actually do this week
The pricing landscape stabilized enough this month to make a decision - Opus 5 at $5/$25 is Anthropic's own recommended starting point for agentic coding, Sonnet 5 at $2/$10 (promotional through August 31) covers high-volume lighter work, and Fable 5 at $10/$50 is now a deliberate upgrade for specific tasks, not a default.
A quick audit checklist for any team running agentic coding tools:
- Copilot users: pull the
ai_credits_usedfield from the usage API and sort by user. Outliers burning 4-5x the team average are almost certainly running multi-step agent sessions against a frontier model. Decide whether that is intentional. - Claude Code users on Max plans: Claude Code now uses Opus 5 as its top-tier model on Max plans ($100/$200 per month), with Sonnet 5 as the default model. Check whether your team's work warrants the Max plan or whether API-key access with explicit model routing gives better cost control.
- Anyone who set Fable 5 as a default before July 9: it is no longer available on subscription plans. Fable 5 is now available only through the Anthropic API at $10/$50 per million tokens or through GitHub Copilot with Fable 5 enabled. Update your model IDs before something silently downgrades or errors.
- Teams evaluating harness vs. model: run the same real task through your current stack and a cheaper model in a better-structured loop. Measure patch correctness, test pass rate, and total token cost including retries. The result is often surprising.
On paper, GitHub's billing change is a pricing model change. In practice, it signals something bigger: AI coding tools are moving from subscription software economics to cloud compute economics. That shift is not going to reverse. Teams that treat model selection as a one-time configuration choice - set it once, forget it - will find the bill teaching them otherwise.
Agentic coding tool pricing: common questions
What is the difference between Claude Opus 5 and Claude Fable 5 for coding?
If you're deciding which Anthropic model to point your agent at, you now have two flagship-class options that sit close together in capability but far apart in price. Claude Opus 5 launched July 24, 2026, at $5 per million input tokens, while Fable 5 released June 9, 2026, at exactly double that. The obvious question is whether Fable 5's higher tier buys you anything Opus 5 doesn't already deliver. For most agentic coding tasks, Anthropic's own answer is no.
How does GitHub Copilot's AI Credits billing work?
Every Copilot plan now includes a monthly allotment of GitHub AI Credits, with the option for paid plans to purchase additional usage. Usage is calculated based on token consumption - input, output, and cached tokens - at the listed API rate for each model. Inline completions are unlimited. Agents, chat, and code review are metered.
Why did some developers burn through Copilot credits so fast?
Agentic sessions send large context windows repeatedly. GitHub's own explanation: Copilot "now powers far more complex, agentic workflows that consume far more compute" than the old request-count model assumed. A single agent task that reads a large codebase, iterates on a fix, and runs tests can consume more tokens in one session than a week of inline completions.
Is it worth paying for Fable 5 over Opus 5?
Opus 5 at $5/$25 suits complex agentic coding, long-horizon tasks, and enterprise work - Anthropic's own recommended starting point. Fable 5 at $10/$50 is justified when you need the capability ceiling and the benchmark gap is worth double the token cost. For most teams, the default answer is Opus 5 unless a specific task shows a measurable gap.
Does harness choice affect agentic coding costs?
Yes, often more than model choice does. A harness with session resume, parallel worktrees, and test-driven verification prevents re-reading full context after every failed step - which is where token costs spike on long tasks. Evaluate the tool's architecture, not just its model menu, before assuming that the cheapest model in the best harness costs more than an expensive model in a naive loop.