Cloud AI Prices Are Only Going One Direction

The GPT-5.5 Price Shock

OpenAI launched GPT-5.5 on April 17, 2026, with a pricing structure that broke the downward trajectory the industry had been riding since 2023. The standard model costs $5 per 1M input tokens and $30 per 1M output tokens — double what GPT-5.4 charged at launch just months earlier. The Pro tier goes further: $30 per 1M input and $180 per 1M output, a 12x multiplier over GPT-5.4's pricing.

For teams already running GPT-5.4 at scale, this isn't incremental. A development team processing 50M input tokens and 10M output tokens per month on GPT-5.4 pays $275,000 per year. On GPT-5.5, that same workload jumps to $510,000 — an $235,000 annual increase for a model upgrade. The Pro tier? $1.56M per year for the same workload.

Three Years of Launch Pricing: The Trend Is Clear

The story you've been told is that AI is getting cheaper. GPT-4's $30/$60 pricing in 2023 fell to GPT-4o's $2.50/$10 in 2024. That's real. But the trend reversed in 2025, and the new generation of frontier models is priced higher than the generation it replaces.

The chart below tracks output token pricing at launch for every major OpenAI model since 2023. The blue line shows standard flagship models. The purple line shows premium tiers. Both lines are now pointing upward.

The pattern is clear. From 2023 to mid-2024, competition and efficiency gains drove prices down. GPT-4 at $60 per 1M output tokens gave way to GPT-4 Turbo at $30, then GPT-4o at $10. But then OpenAI introduced the premium tier with o1 at $60 in September 2024 — the same price as the original GPT-4. The "cheaper AI" narrative only applied to the commodity tier. The frontier tier, which is where the best capabilities live, has never been cheap.

Now even the standard tier is reversing. GPT-5.4 at $15 was already 50% more expensive than GPT-4o. GPT-5.5 at $30 tripled GPT-4o's price. GPT-6 Astra, launched in September 2026 at $10/$50, took it to five times. At the top of the range, the race to the bottom is over.

Below the flagship, the story is different — and we should say so plainly. OpenAI cut GPT-6.1 Sol to $2/$10 and GPT-6 Luna to $0.10/$0.50, and Anthropic priced Claude Opus 5.5 20% below Opus 5. Cheaper tiers got cheaper. The best model kept costing more. And in both directions, the vendor sets the rate and changes it without asking you.

Subscriptions: The Same Story, Monthly

API pricing is only half the picture. Consumer and team subscriptions have followed an identical trajectory:

Three price tiers in two years. The $20/month plan still exists, but it doesn't include the models that matter most. GPT-5.5 Pro and its equivalents live behind the $200/month paywall — and each new capability OpenAI ships will be gated the same way. As of October 2026 the ladder hasn't moved: ChatGPT Pro is $200/month, and Claude’s Max plans are $100 and $200/month. The pattern is consistent: launch at a low price, build dependency, then move the best features to the most expensive tier.

What This Costs a Real Team

A 10-person development team using AI for code generation, code review, and documentation provides a concrete picture of the cost spiral:

That's a 3x cost increase in two years for the same team doing the same work. Not because they're using more AI, but because the price of the model they rely on kept rising. And there's no ceiling in sight at the top: each new flagship has launched higher than the last, and GPT-6 Astra and Claude Fable 5.1 now list at $50 per 1M output tokens. Cheaper tiers exist, and some got cheaper in 2026 — but the vendor sets that price too, and promotional rates end. At publication, OpenAI's reported annualized revenue run rate was above $5 billion, and its compute costs were growing faster than revenue. The pricing spiral at the frontier isn't a bug. It's the business model.

The Open Model Alternative

In April, open-weight models had reached near-parity with proprietary models on the benchmarks that mattered — and their pricing told a very different story. Launch-week prices, April 2026:

At April prices, Kimi K2.6 cost 12x less than GPT-5.5 on output tokens, and scored within 1.5 percentage points on SWE-bench Pro. GLM-5.1 cost 10x less. These aren't toy models — they're production-grade models that compete with the proprietary frontier on the hardest benchmarks.

Six months later the picture is more mixed. List prices per 1M tokens as of October 2026, with Artificial Analysis Intelligence Index v4.3 scores where published:

Some honest reading. OpenAI’s GPT-6.1 Sol now undercuts Kimi K3’s API on both input and output, and Google’s Gemini 4 Argon matches Sol — until its launch promotion ends. The cheapest open APIs, DeepSeek V4.1 Flash and MiniMax M3, still cost a fraction of any flagship. And on the rescaled v4.3 index, the gap to the very top widened: the best open model, MiMo-V2.6-Pro, scores 46 against Claude Opus 5.5’s 58, though Artificial Analysis puts its cost at about $0.13 per index task against $1.99–$7.63 for the 53-point closed tier.

But even cloud API pricing for open models is a variable cost that your cloud provider controls. The real escape is on-premises.

On-Premises: Fixed Cost, Zero Surprises

On-premises AI infrastructure replaces the variable-cost spiral with a single, predictable cost. A fixed monthly plan covers the hardware. You download the models. You run inference as much as you need. No per-token charges. No subscription tiers. No price hikes when a new model launches.

A Faraday Machines Growth plan at $599 US a month includes hardware and support for the full 36-month term. The whole term, $21,564, costs less than two months of GPT-5.5 API spending for a 10-person team. Beyond the plan, the only ongoing cost is electricity. Every additional token is free.

References