When you launch a consumer‑facing chat app in 2026, the first line item that climbs into the budget is the token cost of the model you choose. A glance at Anthropic’s cheapest model, claude‑3‑haiku, shows a $0.25 input rate and a $1.25 output rate per 1 M tokens, while OpenAI’s gpt‑oss‑20b offers $0.03 input and $0.13 output for the same token bucket. The difference is stark: the output price of Claude is almost ten times higher than OpenAI’s lightest offering, and the input price is more than eight times higher.
## Token rates that matter
The table below pulls the exact figures from the official pricing pages. All rates are per 1 M tokens and are quoted in the vendor’s native currency.
| Vendor | Model | Input | Cached input | Output |
|---|---|---|---|---|
| Anthropic | claude‑3‑haiku | $0.25 per 1 M tokens | $0.03 per 1 M tokens | $1.25 per 1 M tokens |
| OpenAI | gpt‑oss‑20b | $0.03 per 1 M tokens | $0.03 per 1 M tokens | $0.13 per 1 M tokens |
| OpenAI | gpt‑5‑nano | $0.05 per 1 M tokens | $0.005 per 1 M tokens | $0.4 per 1 M tokens |
| Anthropic | claude‑haiku‑4.5 | $1 per 1 M tokens | $0.1 per 1 M tokens | $5 per 1 M tokens |
The raw numbers suggest that for high‑volume, low‑complexity chat, OpenAI’s OSS and nano models are the clear cost leaders. The cached‑input discount on gpt‑5‑nano is especially attractive when a system prompt is reused across many turns; the $0.005 rate is a fraction of the $0.03 charged by Claude for cached input. Even when you look at higher‑tier models, the gap widens. Anthropic’s claude‑sonnet‑5 costs $2 for input and $10 for output, whereas OpenAI’s comparable gpt‑4.1‑nano costs $0.1 for input and $0.4 for output, a factor of 20‑plus in both dimensions.
## Choosing the right model
Beyond the numbers, the ecosystem matters. Anthropic’s API guarantees lower latency and includes built‑in safety filters, which can reduce engineering overhead for compliance‑heavy applications. OpenAI, meanwhile, offers a broader language coverage and a more mature set of tooling, including fine‑tuning and prompt‑engineering utilities. If your product relies on the absolute lowest output cost and you can tolerate a newer provider, the nano and OSS tiers are unbeatable.
For startups that need to keep burn low and scale quickly, the recommendation is to start with gpt‑oss‑20b for the core chat loop and reserve claude‑haiku‑4.5 or higher for premium features that justify the premium. For enterprises that require the highest assurance of safety and latency, or that need to host the model in a private cloud, Anthropic’s claude‑sonnet‑5 or claude‑opus‑5 may still be worth the price if the use case demands it.
Pricing verified on 2026-09-12.