Analysis
A year ago, building anything serious on top of a large language model meant watching the meter. By June 2026, the meter barely moves. The price of running a million tokens through a capable model has dropped so far that the question for most teams has flipped from "can we afford this?" to "which of these near-identical cheap options do we even pick?"
That shift is the real story behind the 2026 pricing war. Dozens of models now sit on a price ladder that runs from roughly $0.15 a million tokens at the bottom to $50 at the very top. Reports put the cheapest capable models near the $0.15/$0.60 mark, while a frontier model can still cost more than thirty times that. The gap is enormous, and it tells you more about strategy than about cost.
For an Australian business team, the upshot is simple. The cheap end is now cheap enough that price is rarely the thing standing between you and a working tool. What matters is matching the model to the job, and knowing which providers are charging you for real capability versus charging you for a service contract.
Here is how the market has sorted itself out, and why the prices look the way they do.
The Pricing Tiers
The market has settled into roughly five tiers, each making a different promise.
The budget tier ($0.15-$0.35/$0.60-$1.20) covers the cheapest capable models. DeepSeek's low-cost line is reportedly around $0.15/$0.60, MiniMax M3 sits at $0.30/$1.20 (OpenRouter - MiniMax M3, a 50% promo off a $0.60/$2.40 list price), and a Gemini Flash model is reportedly near $0.35/$0.70. These give you solid capability at prices that make high-volume work practical. They are the default for jobs where cost beats marginal quality: content moderation, document classification, data extraction, customer service triage.
The mid-tier ($0.50-$3/$2-$15) includes Kimi K2.7-Code (reportedly $0.50/$2.00), Qwen 3 (around $1/$3, an estimate rather than a published rate), a GPT-5.5 Instant tier reportedly near $0.50/$1.50, and Claude Sonnet 4.6 at $3/$15. These handle harder tasks while staying cheap enough for most production use. They are the workhorses of enterprise AI.
The premium tier ($5/$25-$30) is Claude Opus 4.8 at $5/$25 and GPT-5.5 at $5/$30. These earn their price on the hardest tasks, with better safety profiles and stronger enterprise features. You reach for them when failure is expensive: financial analysis, legal review, medical decision support.
The ultra-premium tier ($8/$40+) is occupied by a GPT-5.5 Pro tier that the article puts at $8/$40, though published rates for Pro are reportedly far higher. This tier targets enterprises that need the highest rate limits, priority support, and contractual guarantees. The model capability is usually the same as the premium tier; you are paying for the service-level agreement, not extra performance.
The suspended tier belongs to Claude Fable 5 ($10/$50), pulled offline after a US government export-control directive on 12 June 2026 (BetaNews - US order forces Anthropic to disable two Claude models). Its removal left a vacuum at the very top. Fable 5's launch materials listed an 80.3% score on SWE-Bench Pro, a figure some independent evaluators dispute, and no model currently on sale matches it.
The Cost Drivers
So what actually sets the price? Three things do most of the work: inference cost, competitive positioning, and strategic subsidy.
Inference cost is the compute needed to push a million tokens through the model, and it swings a lot with architecture and efficiency. A Mixture-of-Experts design like DeepSeek's keeps inference cheap, which is part of how the low price points are possible. A larger model such as GLM-5.2 (reportedly an MoE design around 753B total, roughly 40B active per token, not a dense model) needs more compute, and its pricing reportedly lands higher, around $0.80/$2.40 in this article versus higher published figures elsewhere. Providers with leaner inference setups, whether through custom silicon, better software, or sheer scale, can charge less at the same margin.
Competitive positioning shapes the rest. Google's low Gemini Flash pricing is widely read as a market-share play; the interpretation is that Google will accept thinner margins to pull customers off OpenAI and Anthropic (DevTk.AI - Gemini API Pricing Guide 2026). Chinese labs like DeepSeek and MiniMax use price as a weapon, undercutting Western providers to build a presence.
Strategic subsidy is when a provider prices below cost to win something else. Meta gives its Llama models away as open weights, which is the ultimate subsidy. Google's Flash pricing may be propped up in part by the wider Google Cloud business, though that reading is interpretation rather than confirmed fact. And DeepSeek's parent, the quantitative trading firm High-Flyer, can in principle run the AI side at a loss for a long time on the back of its trading profits.
How Low Can Prices Go?
The theoretical floor is the marginal cost of inference: the energy, compute, and operating cost of one more token. For efficient models at scale, the author estimates this at $0.02-0.05 per million input tokens. Current budget pricing sits roughly 3-15x above that, which hints at room to fall further as efficiency improves and competition bites. Treat those bands as estimates, not measured figures.
Most providers, though, aren't running at theoretical efficiency. A more realistic floor for a typical provider is estimated at $0.08-0.15 per million input tokens for budget models. If that holds, today's cheapest prices are already brushing up against sustainable limits, and the next big drop will need a genuine efficiency breakthrough rather than another round of subsidy.