Nobody Knows What AI Will Cost Next Quarter. Budget for It Anyway.

You already decide how much intelligence a task deserves every time you hand work to a person. For the last three years, AI subscriptions were the one place that skill didn’t apply, because once the seat was paid for, every model in it cost the same. That ended this year. Every major vendor put a meter on its top models, so the best available intelligence now has a price per use, and that price has changed about once a month since April, in both directions.

I run the AI practice at an IT company, so I hit this months before most businesses will. I plan projects with Fable 5 and hand implementation to GPT-5.6 Terra through Codex. Until July, both were covered by subscriptions we were already paying for. Then Fable moved onto usage credits, came partway back into the plan two weeks later, and the Codex work started billing by use. The mix still works, and for the first time I can see what each piece costs.

What Changed in 2026

The announcements ran April through September, and the second half of the list is prices coming down or increases being cancelled.

On April 23, OpenAI released GPT-5.5 and doubled its flagship line’s per-token prices to $5 per million input tokens and $30 per million output.⁹ On June 1, GitHub moved every Copilot plan to usage-based credits: each plan carries a monthly allowance, and going past it means setting a paid budget that bills at the end of the month.^15,16 Anthropic scheduled the same move for its agent tooling on June 15, then paused it the day it was due to take effect, and it is still paused.⁷ And on July 7, Fable 5, the most capable model Anthropic sells, left subscription plans for usage credits at $10 per million tokens of input and $50 per million tokens of output, twice the price of its next model down.^1,2

A token is the unit these prices are quoted in, and one token is about three quarters of a word.¹ A fifty-page contract runs around 30,000 tokens, so having Fable read it costs about thirty cents.

Then prices started moving the other way. On July 20, Anthropic put Fable 5 back into Max plans and premium Team seats at up to half of weekly usage limits; Pro and standard Team seats stay on usage credits, softened by a one-time credit.³ On July 24 it released Opus 5 at $5 / $25, the same price as Opus 4.8, at what it calls near-Fable intelligence for half the cost per task.⁴ On July 30, OpenAI cut GPT-5.6 Luna by 80% and Terra by 20%, and on August 21 cut Sol to $4 / $20, a promotional rate that runs through November 21.^10,11 Anthropic cancelled the Sonnet 5 price increase it had scheduled for September 1, so the $2 / $10 rate is now the standard price.¹ Google shipped three Flash models over the summer at $0.75 / $3.75, introductory through December 31 and then doubling.^12,13 And on September 1, Fable 5.1 arrived at the same $10 / $50 with cache reads at a quarter of the old price, which Anthropic says cuts typical workloads by about 25%.⁵

This isn’t confined to developer tools. On June 16, Microsoft took Copilot Cowork, the agent side of Microsoft 365 Copilot, to general availability on usage-based Copilot Credits, billed on top of the Copilot seat itself, which runs up to $30 per user per month.^17,18 Estimates built on Microsoft’s own cost calculator put a light task at a dollar to three and a heavy one at seven and up.¹⁷ If your AI budget still says one flat number per seat, it’s out of date.

Developers on community forums reported monthly Copilot costs jumping from $29 to around $750, and in one case from $50 to $3,000.¹⁴ Those are the worst cases that made the rounds, well above the average.

The reason is agents. A chat question costs a vendor a fraction of a cent to answer. But agentic workflows, where the model plans, calls tools, checks its own work, and iterates, consume tokens at a different order of magnitude: the Harvard Business School AI Institute estimates that simple tool-calling agents burn 5,000 to 15,000 tokens per task, and complex multi-agent workflows routinely run 200,000 to over a million tokens for a single goal.⁸ Flat-rate pricing was set to win chat customers, and it didn’t cover agent compute. Once a $20 subscriber could burn hundreds of dollars of it in a month, every major vendor moved toward metering within about a year.

The subsidized era is over. It’s just not going to be as bad as it felt in July, and I don’t think it swings so far that hiring people back becomes the cheaper option. The disconcerting part is that the ground keeps moving: every promotional rate in that list has an end date, and the vendors have shown they’ll change the number on short notice. Budgeting on it feels like building on quicksand. The one thing that stays useful at any price level is knowing what a task is worth before you assign it.

Choosing a Model Is a Delegation Decision

Every organization already runs on metered intelligence. It’s called payroll. A law firm bills a partner at $900 an hour, an associate at $350, and a paralegal at $150, and nobody asks the partner to spend the afternoon photocopying exhibits. The partner could do it. Nobody would let them.

People have compared AI models to interns since the chatbots arrived, but when every model cost the same, nothing made anyone act on it.

So people developed a habit that would be fireable in any other context. They sent everything to the most capable option available, password resets and strategy questions alike, because it was free. With a price on every request, choosing a model becomes a delegation decision, and the managers who delegate well have a head start.

The analogy breaks in one place. A junior employee who is out of their depth slows down, asks questions, and escalates. A junior model doesn’t. It answers at the same speed and the same confidence as the frontier tier, and it never tells you it’s guessing. So the work you send down the ladder has to be the kind you’d hand a new hire with a checklist: structured, routine, repeatable, with a written procedure that says how to do it and what done and correct look like. Somebody then checks the output against that same checklist. That somebody can be the expensive model reading the cheap one’s work, which costs a fraction of having it do the job, but anything that goes to a client still gets a person’s eyes. Either way, the review belongs in the cost.

The Intelligence Ladder

Here’s what the current market looks like when you read it like a staffing chart instead of a spec sheet. The rows sort by price band, not by any benchmark. Prices are per million tokens, input / output, from the vendors’ pricing pages as of early September 2026.^1,6,12

Sol’s $4 / $20 and Gemini Flash’s $0.75 / $3.75 are promotional rates that end November 21 and December 31. Check the vendors’ pricing pages before you budget against any row.

If you buy seats rather than models, the ladder still applies; the seat decides which rungs are prepaid. On Claude, Max and premium Team seats include the frontier row up to half your weekly limit, and Pro and standard Team seats meter it on top.³ A Microsoft 365 Copilot seat covers the assistant, and Cowork’s agent work bills separately.¹⁸

Look at the spread before the absolute numbers: top to bottom of that ladder is more than a 40x difference in price. Run one real task at the top of the ladder and in the middle and you can see it in dollars. Say you hand a model forty vendor contracts, about 5,000 tokens each, and ask for a one-page summary of each: 200,000 tokens in, 40,000 out. That job costs about $4.00 on Fable 5.1, about $2.00 on Opus 5, and about $0.80 on Sonnet 5.¹ That’s five times the price between top and middle for one task, and for a lot of tasks the result is the same.

Some problems need the frontier tier, and sending them down the ladder gets you a confident wrong answer where it costs the most. The skill is telling those problems apart from the rest.

What a Hybrid Workflow Looks Like

In my own work, and in the vendors’ own cost guidance, one split keeps showing up: the expensive model plans and reviews, and a cheaper one does the work in between.¹

My coding workflow is the concrete version. A frontier model writes the plan: it reads the codebase, weighs the tradeoffs, and produces a plan with guardrails, edge cases, and a definition of done. Then GPT-5.6 Terra, at about a fifth of the price, does the implementation, which is where the hundreds of thousands of tokens get consumed. The frontier model comes back at the end to review what got built against the plan. The plan and the review together read a few hundred thousand tokens and write a few thousand, so at frontier rates they come to a few dollars. The split pays when the implementation burns several times what the plan and the review do. For a task the size of one email, use one model.

The plan is the part people skip, and it’s the part that makes the economics work. Give two tiers the same precise, well-scoped plan and, on routine execution work, the gap between them mostly closes. Most of the AI failures I get called into on that kind of work started with a thin plan. On long multi-step reasoning the gap is real and a better plan won’t close it. A good plan is the reason you can send work down the ladder at all, and the metered price is the reason it now pays to write one. The plan matters because the model can’t know what’s in your head until you write it down, which is the whole argument of our post on the Sally-Anne test.

Outside of code the pattern holds. Have the strong model design the analysis and the checklist, let the cheap model grind through the forty contracts, and have it flag anything that doesn’t fit the checklist. Then bring the strong model back for the flagged ones, and read a handful of the unflagged ones yourself.

What to Actually Do About It

If you run a business and this is the first you’re hearing about metered AI, four moves, none of them complicated:

• Set spending caps today. Every API console and the Microsoft 365 admin center support a hard limit. Subscription seats vary, so check whether an admin can cap usage credits on yours; if not, the cap is the seat count. Do it before the first surprise invoice.
• Find out what you’re running. Many companies discover mid-audit that AI spend is scattered across personal subscriptions, seat licenses, and API keys nobody remembers creating. Get it into one view before you cut anything.
• Write routing rules. They don’t need to be sophisticated. “Drafts and summaries go to the mid-tier model, anything touching a client commitment gets the senior tier plus human review” is a real policy.
• Expect the prices to keep changing. Four of the rates in the table have moved at least once since April and two expire before year end. Budget against measured usage rather than the sticker price, and recheck the pricing pages every quarter.

Paying frontier prices for routine work is the obvious failure, the partner doing data entry, and the invoice shows it. Forcing a junior model into judgment work to save eighty cents is the quieter one. You pay for it later in rework, bad decisions, or the wrong answer nobody caught, and none of that appears on the invoice.

If your team is trying to figure out where AI fits and what it should cost, that’s the work our thinkAI program does: thirteen weeks inside your real workflows, deciding what AI should do and what it shouldn’t, with a person leading it. If you’re earlier than that, start with our 90-day AI pilot playbook. See how thinkAI works or start a conversation.

Sources

1. Pricing. Anthropic Claude Platform docs, live page, checked September 2026. platform.claude.com

2. Redeploying Claude Fable 5. Anthropic, June 2026. anthropic.com

3. Claude Fable models on your plan. Anthropic Help Center, updated September 2026. support.claude.com

4. Introducing Claude Opus 5. Anthropic, July 24, 2026. anthropic.com

5. Introducing Claude Fable 5.1 and Claude Mythos 5.1. Anthropic, September 1, 2026. anthropic.com

6. Pricing. OpenAI Developer Platform, live page, checked September 2026. developers.openai.com

7. Use the Claude Agent SDK with your Claude plan. Anthropic Help Center, updated June 16, 2026. support.claude.com

8. Codex for (Almost) Everything, and the End of the AI Subsidy Era. Harvard Business School AI Institute, June 2026 memo. aiinstitute.hbs.edu

9. GPT-5.5 Pricing: Full Breakdown of API, Codex, and Subscription Costs. Apidog, June 2026. apidog.com

10. GPT-5.6 Pricing: Sol, Terra, and Luna Costs Explained. eesel AI, updated August 2026. eesel.ai

11. 20% price reduction for GPT-5.6 Sol: API, Codex credits and ChatGPT Work. OpenAI Developer Community, August 21, 2026. community.openai.com

12. Gemini Developer API Pricing. Google AI for Developers, live page, checked September 2026. ai.google.dev

13. Introducing Gemini 3.8 Flash and 3.8 Flash Cyber. Google, September 2, 2026. blog.google

14. The Tokenpocalypse: AI Coding’s Flat-Rate Era Ended in 2026. UsageBox, June 2026. usagebox.com

15. Updates to GitHub Copilot billing and plans. GitHub Changelog, June 1, 2026. github.blog

16. GitHub Copilot plans. GitHub, live page, checked September 2026. github.com

17. Copilot Cowork Pricing 2026: How Usage-Based Billing Works. Quisitive, June 2026. quisitive.com

18. Copilot Cowork is now generally available. Microsoft 365 Blog, June 16, 2026. microsoft.com

19. Understand usage-based billing and cost management for Copilot Credits. Microsoft Learn, updated August 2026. learn.microsoft.com

20. Set up Microsoft 365 Copilot pay-as-you-go services. Microsoft Learn, updated July 2026. learn.microsoft.com