AI Model Prices Fall Up to 50% as Microsoft Meters Copilot Agents: What It Means for Businesses

Executive Summary

This week, Anthropic and OpenAI cut the price of their leading AI models on the same day. Both companies say typical work now costs about 40–50% less. Three days later, Microsoft said Copilot’s AI agents and most powerful models will be billed by usage instead of a flat licence. Meanwhile, some AI data-centre projects ran into power and financing problems, and governments tightened control over which AI suppliers can be used. Anyone who budgets for AI should take one lesson from this: AI work is getting cheaper, but the bill is getting harder to predict.

Estimated reading time: 15 minutes

Key Takeaways

• Anthropic and OpenAI say their new models cut the cost of typical AI work by about 40–50%. Business cases priced earlier this year are probably out of date.
• Judge AI on the cost of a finished task, not the list price. A model that looks cheaper on paper can cost more if it needs more attempts.
• Microsoft now bills Copilot agents by usage. Put spending limits in place before anyone switches them on.
• Power shortages and complex debt are slowing some AI data-centre projects. That risk sits underneath the falling prices.
• A U.S. court ruling and pressure from the White House show that choosing an AI supplier now carries political and regulatory risk.
• Using more than one AI supplier is becoming a sensible safeguard, not just a way to save money.

What Happened?

Four developments shaped the week. Two of them changed what AI costs and how it is billed. The other two changed the conditions around it.

Date (2026)	What happened
22 September	Anthropic released Claude Opus 5.5. OpenAI released GPT-6 Sol and GPT-6 Luna about 90 minutes later.
23 September	Brookings published a paper estimating $10.3 trillion of U.S. AI infrastructure investment from 2025 to 2032.
24 September	Oracle sent a “force majeure” notice on a major data-centre project. Akamai announced an $11.6 billion deal with Anthropic. Politico reported that the White House asked AI labs to hold new models back from UK testers.
25 September	Microsoft unveiled a rebuilt Copilot with usage-based billing. A U.S. appeals court upheld the Pentagon’s “supply chain risk” label on Anthropic.

Anthropic and OpenAI cut model prices on the same day

AI companies charge by the “token”. A token is a small chunk of text, often part of a word. Prices are quoted per million tokens, with separate rates for the text you send in (input) and the text the model writes back (output).

Model	Input price	Output price	Change
Claude Opus 5 (previous)	$5	$25	—
Claude Opus 5.5	$4	$20	20% lower per token
GPT-5.6 Sol (previous)	$4	$20	—
GPT-6 Sol	$2	$10	50% lower
GPT-5.6 Luna (previous)	$0.20	$1.20	—
GPT-6 Luna	$0.10	$0.50	50% lower

Prices in U.S. dollars per million tokens. Sources: Anthropic, OpenAI.

Anthropic says Opus 5.5 costs about 40% less to run than Opus 5 on typical work. The per-token price fell 20%, and the model also uses fewer tokens to finish a job. Anthropic says it writes its output more than 30% faster and performs close to its more expensive Fable 5.1 model on most work. Opus 5.5 is available in Claude and through Amazon Web Services, Google Cloud and Microsoft Azure.

OpenAI says Sol and Luna were trained using methods similar to those behind its flagship model, GPT-6 Astra. Both are available in ChatGPT Work and Codex for paid plans, and to developers through OpenAI’s API. OpenAI also improved “caching”, which gives a 90% discount when the same text is sent to the model again.

Microsoft rebuilt Copilot and changed how it charges

Microsoft reorganised Microsoft 365 Copilot around three areas (Microsoft):

• Home brings chat and Cowork together, with Word, Excel and PowerPoint built into Copilot.
• Code lets staff without programming skills build apps, trackers, dashboards and automations by describing them in plain language.
• Autopilot is an AI agent that keeps working on tasks in the background. An AI agent is software that carries out multi-step work on its own rather than answering one question at a time.

The bigger change is billing. The standard per-user licence still covers chat and the Office features. Cowork, Code, Autopilot and top models such as OpenAI’s Astra and Anthropic’s Fable now use metered “Copilot Credits”. Microsoft also announced new tools to track and control AI spending.

Home and Code will reach Microsoft’s early-access programme in the coming weeks. Autopilot enters a limited preview at the end of September. Microsoft has not announced dates for general release.

The infrastructure behind AI showed strain

A Brookings paper by Columbia professor Stijn Van Nieuwerburgh estimates that the U.S. will invest about $10.3 trillion in AI infrastructure from 2025 to 2032. That is about 3.6% of GDP each year. The paper finds that more of this spending is being funded through private credit, leases and other arrangements that do not appear on big tech companies’ balance sheets (Brookings).

Oracle sent a force majeure notice to the developer of Project Jupiter, a planned data-centre campus in New Mexico. The campus is part of the Stargate AI infrastructure programme and is meant to open in 2028. A force majeure clause excuses delays caused by events outside a company’s control. According to people familiar with the matter, Oracle is using it to protect itself if power delays push the project back. Oracle says the project “remains on our planned schedule” (CNBC).

Kansas City Federal Reserve President Jeff Schmid said the Fed needs to understand whether AI is becoming a “too-big-to-fail AI ecosystem”. Demand is still strong, though: Akamai signed an $11.6 billion, seven-year agreement to run computing work for Anthropic (Akamai).

Governments tightened control over AI suppliers

A U.S. appeals court voted 2–1 to uphold the Pentagon’s decision to label Anthropic a national-security “supply chain risk”. The court pointed to Claude’s built-in refusals for certain uses, including some involving autonomous weapons and mass domestic surveillance. A separate court in California struck down a related designation in August. Anthropic says the label has cost it billions of dollars in lost business and is considering further appeals (CNBC).

Separately, Politico reported that the White House asked OpenAI and Anthropic not to share new models with the UK’s AI Security Institute until U.S. officials had tested them. The report relied on unnamed sources. The Information also reported that Google, OpenAI and Anthropic are planning a joint industry standards body. None of the three companies has formally announced it.

Brief Signals

• AI coding tools reach scale. Cognition, maker of the Devin coding agent, says it passed $1 billion in annualised revenue less than two years after Devin launched (Cognition).
• AI can move costs rather than cut them. The Blue Cross Blue Shield Association estimates that hospitals’ AI billing tools added about $942 million to its member plans’ costs over 2024–25, without matching changes in treatment. Hospital groups dispute the finding (Fierce Healthcare).
• Shopping by AI agent moves closer. Stripe made its checkout pages easier for AI agents to use, and Meta announced shopping and travel connections for its Muse assistant. Six major banks, including Bank of America, ING and NatWest, warned about new fraud and dispute risks.
• Regulators ask for AI inventories. New York’s utility regulator ordered utilities to list every way they use AI within 60 days.
• More launches ahead. OpenAI holds its DevDay developer event on 29 September.

Why It Matters

For the past two years, many AI projects stalled because they cost too much to run at scale. This week removed a large part of that barrier.

When the two leading AI labs cut prices on the same day, it shows real competition rather than a one-off promotion. It also shows that what you should measure has changed. Both companies stressed that their new models finish work in fewer steps. What you actually pay depends on how efficiently a model completes a task, not just its list price.

The market now has clear price tiers. Premium models such as GPT-6 Astra handle the hardest work. Mid-tier models such as Opus 5.5 and GPT-6 Sol cover most professional work. Very cheap models such as GPT-6 Luna make high-volume tasks affordable.

Microsoft’s change matters just as much, because so many companies run on Microsoft 365. Its new model is simple: the licence pays for an assistant, and agents that do real work are metered. Many other software vendors are likely to follow the same pattern.

Important: Most performance and cost figures from this week’s launches come from the vendors themselves. No independent test comparing Opus 5.5 and GPT-6 Sol under the same conditions was available at launch. Treat the 40–50% savings as a starting point for your own testing, not a guarantee.

The infrastructure and government stories add a warning. Cheap models run on data centres that are expensive, sometimes delayed and increasingly funded with debt that is hard to see. Governments are also showing that they can limit which AI suppliers a company can use.

Business Impact

Enterprise

Large organisations have the most to gain and the most to manage. The price cuts could reduce the cost of AI used in software development, document analysis, financial modelling and customer support. Several early testers quoted by Anthropic reported large savings: trading firm Optiver said its coding workload cost 40–50% less, and Box said Opus 5.5 used a third of the tokens Opus 5 did.

At the same time, Microsoft’s metered billing turns part of AI spending into a variable cost, more like cloud computing than a software licence. Finance and IT teams will need forecasts, spending caps and a way to charge costs back to the departments that use them.

Small and Mid-Sized Businesses

Cheaper models put more tasks within reach of smaller budgets. GPT-6 Luna, at $0.10 per million input tokens, makes it affordable to sort emails, summarise documents or tag customer requests in large volumes.

Hypothetical example: a 40-person agency spending $2,000 a month on AI through an API could, if the vendor savings hold for its work, pay closer to $1,000–$1,200 for the same output. The only way to know is to run the same tasks through the new models and compare.

Finance and Procurement

Existing Copilot agreements may not cover the new agent features at all. Budget owners should also avoid long AI contracts at fixed prices, because some prices have halved in about two months.

Software Development

Coding is where the new models are strongest and where spending is highest. Both vendors focused their claims on coding. Cognition’s $1 billion revenue milestone shows how much companies are already spending on AI coding tools.

Organisations with Government or Cross-Border Work

Defence contractors may need to avoid Claude in work covered by the upheld Pentagon designation. Businesses outside the U.S., including in Europe, may see new U.S. models later if U.S.-first testing becomes standard practice. That is not confirmed.

Key Changes

• Lower prices: GPT-6 Sol and Luna cost 50% less per token than their predecessors. Opus 5.5 costs 20% less per token and, according to Anthropic, about 40% less per task.
• Efficiency as a selling point: Vendors now compete on how few steps and tokens a model needs to finish a job.
• Metered agents in Microsoft 365: Copilot’s agent features and top models are now billed by usage.
• Multiple models inside Copilot: Microsoft offers both OpenAI and Anthropic models, and admins can choose which model families different groups of staff can use.
• Automatic model choice: Copilot can pick a model for each request based on accuracy, speed and cost.
• Supplier eligibility set by government: A court ruling confirmed that a leading AI lab can be excluded from U.S. defence supply chains because of its usage policies.

A simple way to use the new price tiers is to route each task to the cheapest model that can handle it well:

text

New AI task
   │
   ├── Routine work (sorting, tagging, short summaries)
   │        └──► Low-cost model, e.g. GPT-6 Luna
   │
   ├── Everyday professional work (drafting, analysis, coding)
   │        └──► Mid-tier model, e.g. GPT-6 Sol or Claude Opus 5.5
   │
   └── Hard, high-stakes work (complex research, critical code)
            └──► Premium model, e.g. GPT-6 Astra or Claude Fable 5.1

The tiers above are Future Signal’s grouping, not an official vendor classification.

Industry Reaction

Early outside testing gave a mixed picture:

• Artificial Analysis ranked Opus 5.5 first on its Intelligence Index.
• Browser Use found that GPT-6 Sol outscored Opus 5.5 on web-browsing tasks and cost about 3.5 times less on its platform.
• Independent tester Nate Herk preferred Opus 5.5 on seven of ten real jobs, but it cost about $213 against about $74 for GPT-6 Sol.

Together, these results suggest Opus 5.5 may lead on quality in some work, while Sol leads on price.

Enterprise testers quoted by Anthropic were positive. Deloitte said Opus 5.5 at its lowest setting caught 72% of known bugs in code reviews, compared with 56% for Opus 5 at a high setting. These are selected customer quotes, not independent studies.

On Copilot, Microsoft CEO Satya Nadella said AI agents for business could become a market far larger than cloud computing. Wharton professor Ethan Mollick warned that automatic model choice matters a great deal, because a system could quietly send a hard task to a weaker model. Analysts at FourWeekMBA and WindowsForum argued that the billing change, not the new features, is the real news.

Financial markets reacted quickly. Oracle’s shares fell about 3–4% after its force majeure notice. Akamai’s shares reportedly rose more than 20% in after-hours trading on its Anthropic deal. Anthropic said it disagrees with the court ruling and remains confident in its position.

Opportunities

• Cheaper existing work: Tasks you already run with AI may cost substantially less on the new models.
• Projects that were too expensive before: Processing large volumes of documents or running assistants all day may now pay off.
• Smarter spending: Sending each task to the right price tier can cut costs without lowering quality.
• Faster internal tools: Microsoft’s Code feature could let teams build simple trackers and dashboards without waiting for developers.
• Stronger bargaining position: With two leading suppliers cutting prices at once, buyers have more leverage in negotiations.

Risks & Limitations

• Unverified claims: Most savings and performance figures come from vendors. Independent comparisons are still limited.
• Unpredictable bills: Microsoft has not published what Copilot Credits cost. Agents that run for long periods can use credits quickly.
• Hidden quality trade-offs: Automatic model selection may choose a cheaper model without users noticing.
• Early-stage products: Microsoft’s new Copilot features are in preview, with no general release date.
• Supply risk: Power shortages and financing pressure could delay new capacity and slow further price cuts.
• Regulatory and vendor risk: Government decisions can limit which suppliers you can use. Relying on a single vendor increases that exposure.
• Access limits in sensitive fields: Anthropic sends some cybersecurity and biology requests to older models unless the user’s organisation is verified. That could affect legitimate work in those fields.

Is your organisation relying on a single AI supplier for anything important? This week showed why that deserves a second look.

Future Outlook

Type	What to expect
Confirmed	Anthropic says Claude Sonnet 5.5 and Haiku 5.5 will follow “in the coming weeks”. Microsoft’s Autopilot preview starts at the end of September, and Home and Code reach early access in the coming weeks. OpenAI’s DevDay is on 29 September.
Likely	AI prices keep falling, with pressure from low-cost Chinese and open models such as DeepSeek. Other business software vendors move to a similar “licence plus usage” model. Managing AI spending becomes a routine finance and IT task.
Editorial interpretation	Competition shifts away from benchmark scores towards reliability, writing quality, integration and cost control. Power and financing constraints may matter more for AI prices than chip supply. Supplier choice increasingly depends on regulation as well as capability.

The Future Signal

AI is starting to behave like a utility.

Until now, many businesses bought AI mainly as a subscription: a fixed fee per person, with much of the real cost absorbed by the vendor. This week showed that model ending. The price of each unit of AI work is falling fast, but the work is now measured, metered and billed like electricity or cloud computing.

That changes who needs to pay attention. AI decisions used to sit mainly with IT and innovation teams. Now they belong with finance, procurement and risk as well. The companies that do well will be the ones that treat AI spending like any other operating cost, with budgets, controls and regular reviews.

The less visible signal is fragility. Cheaper models depend on a huge, debt-funded build-out that is already hitting power limits, and on suppliers whose access governments can restrict. Falling prices are real, but they rest on foundations that are still being tested.

The long-term winners will not be the companies that use the cheapest model. They will be the ones that can switch models, measure results and control costs as the market keeps moving.

What Businesses Should Do Next

These steps are ordered from most to least urgent for most organisations:

• Re-test your main AI tasks on the new models. Run the same real work through Opus 5.5, GPT-6 Sol and GPT-6 Luna, and compare cost per finished task and quality.
• Update your AI budget and business cases. Projects rejected as too expensive this year may now pass.
• Set Copilot spending policies before enabling agents. Decide who gets access to Cowork, Code and Autopilot, and set limits.
• Ask Microsoft how Copilot Credits are priced. Get figures before signing new agreements.
• Avoid long fixed-price AI contracts. Prices are falling too quickly to lock in today’s rates.
• Map your AI suppliers. List which vendors you depend on, and check whether any government, defence or cross-border work limits your options.
• Plan a backup supplier for critical work. Make sure important processes could switch to another model if needed.

Future Signal Tip: Pick three tasks your team runs every week. Give each one to two different models and record the cost and the quality of the result. One afternoon of testing will tell you more than any vendor benchmark.

Frequently Asked Questions

Will my AI costs actually fall by 40–50%?

Possibly, but not automatically. The figures come from vendors and apply to “typical” work. Your savings depend on your tasks, the model you switch to and how it is set up. Test before you budget.

Should I switch from one AI provider to another?

Not straight away. Test both on your own work first. Early results suggest Opus 5.5 may lead on quality in some tasks while GPT-6 Sol leads on price, so the right choice depends on the job.

Does Microsoft’s change increase my current Copilot bill?

Not directly. The standard licence still covers chat and the Office features. Extra costs apply only if your organisation uses the metered features, such as Cowork, Code, Autopilot or the premium models.

What are Copilot Credits?

They are the units Microsoft uses to bill metered Copilot features. Microsoft has not yet published credit prices, so ask your Microsoft account team for details before enabling these features.

Is the AI data-centre boom a bubble?

The evidence does not support that conclusion yet. The author of the Brookings paper says it is too early to call it a systemic risk. What it does show is that the build-out relies more and more on complex debt, and that power supply is becoming a real constraint.

Can my business still use Claude?

Yes, for most businesses. The upheld Pentagon designation affects U.S. defence supply chains. If your organisation does defence-related work in the U.S., check with legal counsel.

Why does the White House request matter to businesses outside the U.S.?

If new U.S. models are tested in the U.S. before other governments see them, customers and regulators elsewhere may get access later. This is based on reporting from unnamed sources and has not been confirmed as formal policy.

What is the simplest thing I can do this week?

Re-test your most common AI task on one of the new models and compare the cost with what you pay now.

Conclusion

The biggest story of the week is that high-end AI got much cheaper, while Microsoft showed how the agents doing the work will be billed from now on. Together, these changes give businesses more power for less money, but also more responsibility for managing costs.

The practical response is to re-test your key AI tasks, update your budgets and set spending controls before switching on metered agents. At the same time, watch the less visible risks: strain in AI infrastructure and governments’ growing influence over which suppliers you can use.

Prices are likely to keep falling. The businesses that benefit most will be the ones ready to measure, compare and switch as the market moves.