Claude Sonnet 5 in GitHub Copilot: what the usage-based pricing shift actually costs

Claude Sonnet 5 reached general availability in GitHub Copilot on June 30, 2026 (across Pro, Pro+, Max, Business, and Enterprise), landing in a billing system that changed underneath it at the same time. Since June 1, 2026, Copilot meters actual token consumption as usage-based AI Credits (1 credit = $0.01) instead of the old fixed premium-request multiplier, where a one-line chat and a multi-hour agent session could cost the same number of requests. GitHub's CPO Mario Rodriguez put it plainly: "Today, a quick chat question in a multi-hour autonomous coding session can cost the user the same amount. GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable." Code completions and next-edit suggestions stay unlimited and unbilled. Everything else is now metered, with no free fallback to a cheaper model once credits run out.

Pricing mechanics

Rates are provider list prices, billed per token:

Sonnet 5's introductory rate is a discount, not the real price. It lands exactly on Sonnet 4.6's rate card once the promo ends. The legacy request-multiplier system (Sonnet 4.6 = 9x, Opus 4.7/4.8 = 27x, code review = 13x) still applies only to Pro/Pro+ annual subscribers who stayed on request-based billing. Everyone else is now paying for the tokens they actually burn.

What the reviews actually found

CodeRabbit's hands-on review is split. On the agentic side, they left a task running over lunch and came back to a model still iterating: "it kept cleaning up its own solution, pass after pass, because it was chasing the best answer instead of the first one that happened to work." On raw code review, though, it regressed. Their fixed known-bug benchmark puts Sonnet 5 at roughly 50–51% bug-catch recall, against Sonnet 4.6's ~63% and their own production baseline's ~57%. Precision improved (29% → ~38–40%), but it also posts three to four times more nitpicks than the baseline. Their verdict: fine for careful agentic builds, not yet a drop-in review upgrade.

Artificial Analysis found the same story from the cost side: Sonnet 5 scores 53 on their Intelligence Index (a 6-point gain over Sonnet 4.6's 47) but costs $2.29 per task, about 2x Sonnet 4.6 and 15% more than Opus 4.8, which AA says is "driven entirely by increased token usage." At max effort it burns roughly 40% more output tokens per task and up to 3x the agentic turns. Latency follows: 52 tokens/second at medium effort, which AA flags as "notably slow."

The Hacker News thread on the release (item 48736605) converges on a routing rule: don't max out Sonnet 5's effort dial on hard problems, escalate to Opus instead. As one commenter put it, "for certain tasks, Opus 4.8 is cheaper than Sonnet 5, and does better than Sonnet 5; if you're doing something hard, just use a bigger model." At extra-high effort, Sonnet 5 costs almost the same as Opus 4.8 for worse output. Not a good trade.

The September 1 cost cliff

Three separate effects land on or around September 1, 2026:

• Price reversion. The $2/$10 promo rate expires and reverts to standard $3/$15. That's a 50% increase on output tokens.
• Credit allowance step-down. Copilot's promotional included allowances shrink back to their real level: Business drops from 3,000 to 1,900 credits/user (≈ −37%), Enterprise from 7,000 to 3,900 credits/user (≈ −44%).
• Tokenizer inflation. Sonnet 5 ships a new tokenizer that maps the same input to more tokens: Anthropic states 1.0–1.35x depending on content; Simon Willison's independent measurement found ~1.4x for English, ~1.33x for Spanish, ~1.28x for Python, and no change for Simplified Mandarin.

No published source has computed a single compounded three-factor figure for Copilot specifically. The closest published estimates (a two-factor price-plus-tokenizer calculation from Finout, "effective cost lands 20–35% above where you started," and a dollar-worked English-workload estimate from CodingFleet, "a real cost increase of 27–42%") are bounds, not a fact, and neither includes Copilot's credit-allowance cut. Treat any single "X% more expensive" number for this stack as illustrative, and model the three effects separately when budgeting.

What to actually do about it

• Set a cost-efficient default model in org-wide model policy rather than letting Opus/xhigh become the default. Model and effort choice is now the single largest cost lever.
• Turn on the budget hard-stop. "Stop usage when budget limit is reached" is off by default; without it, spend keeps going past whatever cap you set.
• Route by effort, not by habit. Low/medium effort for routine work, and escalate genuinely hard tasks to a frontier model instead of cranking Sonnet 5 to xhigh. The Pareto frontier favors that split.
• Re-baseline token counts. Don't reuse Sonnet 4.6 consumption numbers to forecast Sonnet 5 spend; the new tokenizer changes the conversion on its own.
• Keep human review in the loop on Sonnet 5-authored code, given the bug-catch regression versus Sonnet 4.6.

Budget for the post-September numbers now, not the summer promo headline. The benchmark has the standard $3/$15 rate already baked in for exactly that reason.