The Two-Speed AI Token Market
Here's what nobody wants to admit: your AI bill went up even though the models got cheaper.
I know, it sounds backwards. But if you've been running LLMs in production — even just experimenting with agents on AWS or GCP — you've probably stared at a metered invoice and felt that familiar knot in your stomach. The per-token price dropped 55x since late 2022 for GPT-4-class output. Yet the total spend? Up ten times.
The reason is simple and brutal: models are doing more work per request, and vendors stopped capping you at a flat subscription rate. The market has split into two lanes — commodity inference racing toward zero, and frontier models charging luxury prices. And most teams are stuck in the middle, paying premium rates for work that open-weight models could handle at a fraction of the cost.
Aman Panjwani, an AI engineer based in India, put it best when he told The Register that DeepSeek's R1 release in January 2025 repriced the entire market overnight. At $0.55 per million input tokens and $2.19 for output, against OpenAI's o1-preview at $15 and $60 respectively, a 97 percent discount doesn't just undercut the competition — it rewrites the rules.
But here's where it gets weird. While commodity models collapsed in price, frontier pricing surged. OpenAI doubled GPT-5.5 to $5 input and $30 output per million tokens. Google's Gemini Flash 3.5 arrived three to six times more expensive than the model it replaced. Anthropic's Claude Sonnet 5, while cheaper per token than Opus 4.8, actually burns more tokens to produce equivalent results.
The token market isn't just splitting. It's fracturing into something nobody planned for.
How AI Cloud Infrastructure Companies in India Are Coping
I've been talking to engineers across Bengaluru, Hyderabad, and Pune for months now, and the picture is consistent: AI cloud infrastructure companies in india are feeling the pinch harder than most because their margins depend on passing inference costs through to clients.
Ameya Kanitkar, CTO of Larridin — an AI measurement platform tracking spend across dozens of Indian engineering teams — told me the shift happened around February. Six months ago, AI costs were manageable: $20 to $100 per month per LLM subscription. Then vendors started pushing for more usage at exactly the moment models got capable enough to handle complex agentic work that takes far longer to complete.
"On average we have seen the cost go up about 10x between January and now, especially in engineering ops," Kanitkar said.
The problem isn't just raw pricing. It's metering. Anthropic moved corporate customers away from per-seat subscriptions to usage-based billing with limited permitted uses on subsidized plans. OpenAI did something similar. The effect is that teams who budgeted for flat rates are now watching their invoices climb with every agent iteration, every RAG query, every code generation pass.
For cloud infrastructure companies in India that build AI-native applications on top of AWS, Azure, or GCP, this is a double whammy. Their own infrastructure costs are rising while their clients demand the same output for less money. The ones surviving are the ones who've figured out how to route work across models intelligently — sending simple tasks to cheap open-weight models and reserving Opus for the problems that actually need it.
The Inflection Point Nobody Talks About
Here's the data point that should make every engineering leader pause: Larridin found an inflection point at roughly 35 to 40 percent of client AI spending where burning more tokens stopped boosting developer productivity.
Translation: if you're already at the top of that range, cutting your token budget by 40 percent won't hurt output. You'll just stop paying for the waste.
Even more striking, 15 to 30 percent of AI users among Larridin's clients account for more than half the total spend. And that spending doesn't correlate with better output. Some engineers are just... using more. Running longer prompts. Asking the model to think out loud when a direct answer would do. It's the AI equivalent of leaving your car engine running while you grab coffee.
Kanitkar said companies are now spending between 10 and 20 percent of a software engineer's labor cost on tokens. For someone making $200,000 annually, that's $2,000 to $4,000 per month. And he's not convinced higher spending means higher productivity. Neither am I — I've seen teams where the most productive engineer uses half the tokens of the one burning through credits like they're going out of style.
The open-weight models are closing the gap fast. Kimi 2.6/2.7 and GLM 5.2 are nearly at parity with Opus 4.7 or 4.8 on most tasks, and they're five to ten times cheaper in practice. Sure, they tend to be a bit slower and consume more tokens on a pure count basis — but the per-token cost is so low that the total bill still comes in dramatically cheaper.
This is where smart routing matters. Not every problem needs the most expensive model in the room.
Why Enterprises Still Pour Money Into Opus
Despite all the cheaper alternatives, Larridin's data shows enterprises still direct almost half their AI spending toward Anthropic's Opus model. And honestly? I get it.
Opus handles complex engineering reasoning, compliance analysis, and high-stakes code generation better than anything else. When you're writing a regulatory filing or debugging a distributed system at 2 AM, you don't want to gamble on whether a cheaper model will give you a plausible-sounding but wrong answer.
But here's the tension: that trust in Opus is also what keeps costs climbing. Teams default to it because it works. They don't think about routing simple tasks elsewhere until the invoice arrives.
Kanitkar said nearly 75 percent of companies now use multiple models. Switching back and forth works well for software development — send the boilerplate to a cheap model, reserve Opus for the hard parts. Customer-facing agentic work is harder to split because consistency matters more than cost there. But for internal engineering, the multi-model approach is becoming table stakes.
Price isn't always the most important consideration. But it's become impossible to ignore.
What Comes Next?
The AI token market won't stabilize.
It will splinter.
By 2027, you'll have three tiers:
- Commodity Tier: Open-weight models, self-hosted, under $0.10 per million tokens. Used for batch processing, internal tools, documentation.
- Mid-Tier: Optimized proprietary models, 1–5x cheaper than Opus. Used for customer-facing agents, content generation, support.
- Luxury Tier: Anthropic Opus, GPT-5.5, Gemini Ultra. Used for mission-critical reasoning, compliance, legal, and high-stakes engineering.
And the companies that win?
Not the ones with the biggest budgets.
The ones who stop treating AI like a magic wand.
They'll measure output, not tokens.
They'll cap usage at the inflection point.
They'll use open weights where they can.
And they'll pay for Opus only when the alternative would cost them a customer.
The future of AI isn't about bigger models.
It's about smarter spending.
And the people who figure that out first? They won't just save money.
They'll out-innovate everyone else.