Two of the AI coding tools most engineering teams run daily quietly got more expensive today. Not through a headline price hike — through introductory and promotional pricing simply expiring on schedule. If you set a budget in June or July based on launch pricing, that budget is wrong as of this morning.
What Actually Changed Today
Claude Sonnet 5 shipped on June 30, 2026 with introductory API pricing of $2 per million input tokens and $10 per million output tokens, good through August 31. Standard pricing — $3 input / $15 output per million tokens — took effect September 1. That’s a 50% increase on both sides of the meter, on the model most teams already migrated their agent workloads to because it’s the new default for Claude Code and ships with a 1M-token context window and adaptive thinking on by default.
GitHub Copilot Studio moved on the same day. Business seats drop from $30 to $19 of included monthly credits; Enterprise drops from $70 to $39. The subscription price is unchanged — what you get for it shrank. On top of that, any Copilot harness-built agents and workflows created before August 3, 2026 — including ones sitting in Dev or Trial environments you forgot about — start consuming Copilot Credits for the first time today, for both authoring and runtime execution.
Neither change is a surprise if you were reading the fine print in June. Both are a surprise if you weren’t, because promotional pricing rarely gets a second announcement when it ends — it just ends.
Why This Keeps Happening
This isn’t coincidence, and it isn’t the last one. The pattern is now predictable enough to plan around:
- A new model or feature ships with introductory pricing to drive adoption and undercut the incumbent (Sonnet 5 undercutting Opus-tier pricing at launch, Copilot Studio’s promotional credit pool undercutting steady-state consumption).
- Teams migrate fast because the discounted rate makes the switch a no-brainer — nobody delays adopting a cheaper, better model.
- The promotion has a fixed end date from day one, usually 60-90 days out, buried in the announcement post rather than the headline.
- The increase lands as “back to normal” rather than “price hike,” which is technically true and still means your run-rate jumps.
If you’re budgeting AI tooling spend on a quarterly cycle, the actionable takeaway is: treat every “introductory pricing through [date]” line as a calendar reminder, not a footnote.
Doing the Actual Math
Here’s what the Sonnet 5 change means for a team running agentic workflows at moderate volume — say 50M input tokens and 15M output tokens a month, a realistic number for a team of 6-8 engineers running Claude Code with subagents for code review, test generation, and refactors:
Before Sept 1 (introductory):
Input: 50M tokens × $2/M = $100
Output: 15M tokens × $10/M = $150
Total: $250/month
After Sept 1 (standard):
Input: 50M tokens × $3/M = $150
Output: 15M tokens × $15/M = $225
Total: $375/month
Delta: +$125/month (+50%)Fifty percent sounds alarming until you compare it to what you were paying for Opus-tier reasoning before Sonnet 5 existed — this is still likely cheaper than your pre-June baseline. The risk isn’t that the new price is bad value. It’s that a team that sized its budget off the March number and never revisited it is now 50% over plan with no corresponding line item explaining why.
For Copilot Studio, the credit math is more binary: if your org has agents built before August 3 that you assumed were “free” because they hadn’t started consuming credits yet, check your Power Platform admin center consumption report this week, not next quarter. A dormant Dev-environment agent that suddenly counts against a shrunk credit pool is exactly the kind of thing that produces a surprised Slack message from finance in October.
What to Actually Do About It
Audit token consumption by workflow, not by tool. A blanket “we spend $X on Claude” number hides which specific agent workflows are output-token-heavy (code generation, long-form documentation) versus input-token-heavy (large-context code review, RAG-style retrieval). Output tokens took the bigger relative hit in the Sonnet 5 change ($10→$15, a 50% jump on the more expensive side of the ledger), so workflows that generate a lot of text — not just read a lot of context — are where the delta concentrates.
Set a recurring calendar check for promotional pricing end dates, not just for the tools you’re using today but for the next model you adopt. When you migrate to a new model because of launch pricing, write the promotion’s expiration date into your cost-tracking doc the same day you flip the switch.
For Copilot Studio specifically, inventory dormant agents now. Anything built before August 3 in a Dev or Trial environment that nobody’s touched since is either worth promoting to a real, monitored workflow or worth deleting — leaving it in limbo just means it silently starts drawing from a smaller credit pool with no owner watching the meter.
Model your effort parameter usage. Sonnet 5’s adaptive thinking (controlled by the effort parameter: low/medium/high/max/xhigh) means output token volume is now a tuning knob, not just a function of task complexity. A workflow defaulting to high effort for tasks that don’t need it is paying the new, higher output rate for reasoning tokens it didn’t need to generate. Auditing effort settings against actual task complexity is now a legitimate cost-optimization lever, the same way choosing the right model tier used to be.
The Broader Lesson
AI tooling pricing in 2026 behaves less like traditional SaaS (stable list price, occasional plan changes) and more like cloud compute spot pricing crossed with a subscription — cheap to start, adjusting toward a steady state as the vendor’s own unit economics settle. That’s a reasonable business model for the vendors. It’s a planning problem for you if your budget process assumes prices are static between quarterly reviews.
The fix isn’t complicated: treat AI tooling costs the way a mature team already treats cloud infra costs — a monitored, alertable line item with known step-function risk, not a fixed subscription you set once and forget. Today’s cliff was visible in June if you were looking. The next one will be too.
Thuận Lương is a Technical Lead with 15+ years in .NET, cloud architecture, and AI systems. He writes about real-world lessons from building production systems.