The AI price war has started
AI vendors are racing to tier prices and product teams now have to treat model choice as a margin and rollout decision, not just a quality one.
This morning's window is a little shorter than usual, but the pricing signal is obvious: AI model vendors are cutting and tiering prices faster than product teams can treat them as static infrastructure. The market now looks less like a single leaderboard and more like a ladder of tiers, each priced for a different job. For PMs, that changes vendor choice, margin math, fallback planning, and launch timing.
| Vendor | Current pricing | Packaging signal |
|---|---|---|
| OpenAI | GPT-5.6 Sol is $5 input / $30 output per 1M tokens, Terra is $2.50 / $15, and Luna is $1 / $6. GPT-5.6 is available across ChatGPT, Codex, and the OpenAI API. 1 | OpenAI now sells three durable tiers, not one default model. It also added explicit cache breakpoints and a 30-minute minimum cache life. 1 |
| Meta | Muse Spark 1.1 is priced at $1.25 input / $4.25 output per 1M tokens. Meta also gives every new API account $20 in free credits and is exposing the model through a public-preview developer portal. 2 | Meta is using aggressive pricing to buy adoption, and the API is gated to its own properties for now. 23 |
| xAI | Grok 4.5 is priced at $2 input / $6 output per 1M tokens. 4 | xAI is pitching Grok 4.5 as a lower-cost frontier model for coding and agentic work. 4 |
| Anthropic | Haiku 4.5 is $1 / $5, Sonnet 5 is $2 / $10 on introductory pricing through August 31, 2026, Opus 4.8 is $5 / $25, and Fable 5 is $10 / $50. 5 | Anthropic still owns the premium end of the range, but it now sits beside much cheaper tiers inside the same stack. 5 |
The important part is not that one vendor got cheap. It is that the whole market now spans a much wider cost band, and the same product can land in very different economics depending on the tier you pick. CNBC framed the shift as a move from bigger models toward routing, cost, control, and compute, which is the right diagnosis for product teams even if the term sounds technical. 6
What PMs should do
First, stop asking which model is best in the abstract. Ask which task deserves premium inference and which task can live on the cheap tier. A classification job, a short summary, or bulk enrichment can usually tolerate a lower-cost model. A planning step, a complex code task, or a user-facing answer with high downside still needs a stronger model.
Second, treat model cost as part of product margin, not procurement trivia. Meta is pricing like it wants adoption share, not just revenue per token. Anthropic still charges a wide premium for its top tier. 5 That means a feature that looks fine in demo traffic can become expensive fast once usage scales.
Third, spec fallback before launch. Every AI feature should declare a primary model, a fallback model, and the degradation mode if the preferred model is unavailable or too expensive. Meta’s API is still in public preview and limited to Meta’s own properties, which is a reminder that access and packaging can change under you. 2
Fourth, use caching wherever the prompt prefix stays stable. OpenAI now exposes explicit cache breakpoints and a 30-minute minimum cache life, which matters if your product reuses the same instructions, policy text, or context across many requests. 1
A simple rollout spec is enough to keep the team honest:
- Define the task class for each AI feature.
- Assign a primary model and a fallback model.
- Set a max cost per task and an acceptable degradation path.
- Mark which prompt prefixes should be cached.
- Add a launch buffer if the feature depends on preview access or a newly released frontier model.
That is the practical shift. The team is no longer choosing a model once and moving on. It is managing a price ladder, a fallback plan, and a cost envelope at the same time.
References
- 1
- 2
- 3
- 4
- 5Anthropic Pricing
anthropic.com
- 6
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.