AI pricing didn’t just “trend” in Q3 2025 - it hit a whole new regime.
In the first half of 2025, the story was: pricey frontier models + simple per-token pricing + early premium subscriptions.
By the end of Q3 (July-September), we had:
- U.S. government-wide access to top models for $1 per agency - and even $0.42 in one case
- Matured $100-$250/month “power user” tiers across multiple vendors
- An intensified API price war (with Chinese competitors openly undercutting U.S. models)
- AI bundled into SaaS seats, telco plans, and productivity suites rather than sold only as tokens
Here’s a clean breakdown of what changed in Q3 vs the earlier part of 2025 - using only U.S. and global mainstream sources.
1. Baseline: What AI pricing looked like before Q3 2025
Token & model pricing
Through Q1-Q2 2025, pricing looked something like this:
- OpenAI’s o3 reasoning models were launched and then sharply repriced: in June, OpenAI publicly announced an 80% price cut for o3 in the API, with forum posts describing new rates as low as about $0.40 per million input tokens and $1.60 per million output tokens, matching GPT-4.1 mini. OpenAI Developer Community+2OpenAI Developer Community+2
- Frontier models from other vendors (Claude, Gemini, etc.) were still generally priced as premium, with higher rates for long context and advanced reasoning.
So by the end of Q2, the “luxury model” price ceiling had already moved down - but the downstream market response was still forming.
Early premium subscriptions
Parallel to API pricing, consumer/pro seats were already getting more expensive and more stratified:
- Anthropic’s Claude plans evolved to a three-tier ladder: Free, Pro (~$18/month), and Max starting at $100/month, with a higher option at $200/month for 5×-20× Pro usage limits and priority access. Anthropic+2TechCrunch+2
- Google’s AI subscriptions were converging around a premium stack:
- Google One AI Premium / Gemini Advanced at $19.99/month, bundling Gemini with 2TB storage and other One benefits. blog.google+2blog.google+2
- By I/O 2025, Google announced AI Pro at $19.99/month and AI Ultra at $249.99/month, positioning them as full suites including Gemini, Flow, NotebookLM and more. The Verge+3blog.google+3blog.google+3
AI as an add-on (with early bundling signals)
Coming into 2025, AI in office suites often required separate add-ons:
- Google charged around $20/user/month for Gemini Business on top of Workspace. Early 2025 changes removed that add-on fee and instead raised Workspace prices by about $2/user/month, effectively bundling AI inside the core subscription. TechCrunch+1
By mid-year, the pattern was:
API prices trending down; subscription ARPU trending up; and early shifts from “AI add-on” to “AI included.”
Q3 is where this all crystallized.
2. Q3’s biggest shock: Government AI went almost free
The single most dramatic Q3 shift came from U.S. federal government deals.
OpenAI: ChatGPT Enterprise for $1 per agency
In August, the U.S. General Services Administration (GSA) announced a OneGov deal with OpenAI that gives federal executive branch agencies access to ChatGPT Enterprise for a nominal fee of $1 per agency for one year, plus a 60-day period of unlimited access to advanced models. U.S. General Services Administration+2OpenAI+2
Anthropic: Matching and expanding the $1 offer
Days later, Anthropic struck its own OneGov deal and followed up with a blog post: Claude for Enterprise and Claude for Government available to all three branches of the U.S. government (executive, legislative, judicial) for $1 per agency for up to a year. TechCrunch+3Anthropic+3U.S. General Services Administration+3
Google & Gemini for Government
The GSA’s public “Buy AI” catalog now lists Gemini for Government at $0.47 per agency under a OneGov deal, alongside OpenAI and Anthropic’s $1 offers. U.S. General Services Administration+1
xAI: Grok undercuts everyone at $0.42
In late September, xAI reached a deal with the U.S. government’s purchasing arm to sell Grok models to federal agencies for $0.42 per organization over an 18-month period, explicitly undercutting both OpenAI and Anthropic. Bloomberg+2TechCrunch+2
Taken together, you now have:
- ChatGPT Enterprise - $1/agency
- Claude for Government/Enterprise -$1/agency
- Gemini for Government - $0.47/agency
- Grok for Government - $0.42/agency
All targeted at agencies inside a U.S. federal IT budget that regularly exceeds $100B/year. The Verge+1
Compared to early 2025, when government AI pricing resembled classic enterprise software contracts, Q3 marks a shift to:
“Logo land grab” pricing - distribution, lock-in, and influence first; revenue later.
3. The other side of the barbell: Ultra-premium power plans go mainstream
While government agencies get symbolic pricing, Q3 cemented the other end of the barbell: $100-$250/month power-user plans.
Anthropic’s Max Plan
Anthropic’s April announcement framed Claude Max as a plan “for people who want Claude to essentially be a full-time copilot,” starting at $100/month, with a $200/month tier for 20× the Pro usage limit and earlier access to new capabilities. Anthropic+2TechCrunch+2
By Q3, this was no longer an experiment - it was clearly positioned as a standard tier between Pro and enterprise.
Google AI Pro & AI Ultra
Google’s I/O 2025 updates turned the Google One ecosystem into a stacked ladder: The Verge+3blog.google+3blog.google+3
- AI Pro - $19.99/month
- A suite of AI tools (Gemini app, Flow, NotebookLM, etc.) replacing and expanding on Gemini Advanced. blog.google+2blog.google+2
- AI Ultra - $249.99/month
- Highest usage limits, access to Gemini 2.5 Pro “Deep Think” mode, Veo video generation, advanced editing tools, 30TB of storage, and YouTube Premium. The Verge+3blog.google+3blog.google+3
Bundles & telco partnerships
On top of that, U.S. telcos began reselling these AI bundles:
- Verizon launched a Google One AI Premium add-on for $10/month in its myPlan and myHome offerings - about half of Google’s direct $19.99/month price - showing how AI subscriptions are being distributed via carriers at aggressive promotional rates. The Verge+2blog.google+2
By late Q3, the pattern is clear across U.S. players:
Free → ~$20 Pro → $100-$250 Ultra/Max is now the standard ladder for serious AI users.
In contrast, in early 2025, $200/month chat plans felt experimental; now they’re normalized.
4. API price war: OpenAI’s cut becomes the benchmark, and challengers undercut from below
4.1 OpenAI’s o3 cut sets expectations
OpenAI’s June announcement - cutting o3 API pricing by 80% - was the inflection that Q3 fully built upon. OpenAI Developer Community+2OpenAI Developer Community+2
Community posts and pricing trackers show that o3’s new pricing brought it in line with OpenAI’s lighter models, with per-million token costs falling into the low-single-digit (or sub-dollar) range depending on configuration.
Ecosystem of price comparison tools
By Q3, you don’t have to guess: there’s a small but growing ecosystem of tools that track and compare LLM costs:
- AgentOps tokencost on GitHub maintains up-to-date token pricing for “major LLM providers” and exposes helpers to estimate costs per prompt. GitHub
- Public web tools like LLM-price.com let teams compare per-token and per-request costs for ChatGPT, Claude, Gemini, Llama, Mistral, and more. LLM Price+2GitHub+2
Compared to the first half of 2025 - when most teams just “used OpenAI and paid the bill” - Q3 looks much more price-aware and multi-vendor.
Zhipu GLM-4.5: competitive pressure from China
In September, Chinese startup Zhipu rolled out a migration plan targeting Claude API users, offering its GLM-4.5 model with: Reuters+1
- A simple “change the API URL” migration
- 20 million free tokens
- A developer package it claims costs one-seventh of Claude’s price with three times the usage capacity
Even though this is aimed at Chinese-run entities (and triggered by Anthropic’s access restrictions), it matters globally: it signals the floor on frontier-class pricing is still falling, with non-U.S. vendors willing to undercut aggressively.
5. From selling “tokens” to selling “AI inside” other products
Another major Q3 shift: AI is increasingly part of someone else’s subscription, not a separate meter.
Workspace & productivity suites
Google pushed this hard:
- Google One AI Premium / Gemini Advanced gives AI across Gmail, Docs, Slides, Sheets, Meet and more for $19.99/month with storage. The Verge+3blog.google+3blog.google+3
- For businesses, Google removed the separate Gemini for Workspace add-on and made AI free inside Workspace, raising base pricing from about $12 to $14 per user per month instead. TechCrunch+1
- Later, Gemini Enterprise tiers for workplace AI were formalized around $21-$30 per seat per month, depending on edition and term. TechCrunch
This is classic bundling: the perceived marginal price of AI inside the suite falls to zero, even though ARPU rises.
Richer bundles = more value stacking
The high-end subscriptions now stack media and storage on top:
- AI Ultra’s $249.99/month includes Project Mariner (agentic workflows), Gemini’s Deep Think mode, Veo video generation, 30TB storage, and YouTube Premium - turning it into a kind of “AI + cloud + media” super-bundle. The Verge+3blog.google+3The Verge+3
So compared to early 2025, where AI features were:
“Pay us extra for AI”
Q3 is more:
“Pay us a bit more for this suite; AI is just what makes it work.”
That bundling dynamic is especially important for SaaS founders who are trying to charge separately for AI features.
6. Q3 vs earlier 2025: side-by-side
Here’s the shift in one table:
| Dimension | Prior to Q3 2025 (Q1–Q2) | Q3 2025 |
|---|---|---|
| Gov & public sector pricing | Traditional enterprise-style contracts; no symbolic ultra-low offers | OneGov deals put ChatGPT Enterprise and Claude at $1/agency, Gemini for Gov at $0.47, and Grok at $0.42 per agency — effectively near-free for a year. (U.S. General Services Administration) |
| Power-user subscriptions | Early experiments with $100–$200 tiers (e.g., Claude Max, high-end Gemini / Gemini Advanced) | Clear, normalized Free → ~$20 Pro → $100–$250 Ultra/Max ladders across Anthropic and Google; telcos resell AI at discounted rates. (Anthropic) |
| API token pricing | Frontier models priced as premium; OpenAI still charging higher o3 rates | OpenAI’s 80% o3 price cut sets new expectations; price-comparison tools and non-U.S. challengers (Zhipu) push costs even lower. (OpenAI Developer Community) |
| Packaging model | AI often sold as a separate add-on (e.g., Gemini Business add-on) | AI increasingly bundled into core subscriptions (Workspace price uplift, AI included) and sold as part of broader “productivity + storage + media + AI” bundles. (TechCrunch) |
| Buyer behavior | Many teams default to a single provider (often OpenAI) and accept costs | Growing FinOps mindset: multi-model architectures, cost calculators, and switching incentives encourage mixing providers by workload. (GitHub) |
7. What to do with this if you’re building or buying AI
If you’re a product or pricing leader at an AI/SaaS company
- Design for falling unit costs, not today’s prices.
Assume per-token costs for high-end models will keep dropping and regional competitors will undercut. Peg your pricing to value and outcomes, not just a markup on current API rates. - Build a barbell in your own pricing.
Mirror what the platforms are doing:- A free or low-priced entry point
- A solid $20-$50/month “Pro” tier
- A true power tier (maybe $100+) that unlocks high usage and advanced workflows
- Expect customers to ask: “Why is this expensive when government gets it for $1?”
Be ready to explain that those government deals are promotional, volume-based, and distribution-driven, not reflective of sustainable unit economics. - Decide whether AI is “inside your seat price” or a separate meter -on purpose.
If you’re competing with Workspace, Microsoft 365, or telco bundles, you may need to treat AI as table stakes in your base seat and monetize on workflow depth, integrations, and specialized outcomes.
If you’re an AI buyer (enterprise, gov, or startup)
- Use the Q3 pricing bar as leverage.
When negotiating, reference the OneGov deals and public price cuts. You may not get $0.42/agency, but those anchors exist in the market. - Adopt a multi-model, cost-aware stack.
Use cheap models for routine tasks and reserve premium reasoning models or Ultra/Max-like tiers for critical workflows - and use tools like LLM-price or tokencost to keep suppliers honest. GitHub+1 - Don’t lock yourself into one vendor’s bundle blindly.
AI inside Workspace or a telco plan might be “free,” but that doesn’t mean it’s your best choice for core production workloads. Treat bundles as a floor of capability, not the ceiling.