When you need the absolute top‑tier model for a high‑stakes product launch, the first thing you’ll notice is the spread between the three flagship APIs. Claude Opus 5 charges $5 per 1M input tokens, $0.5 per 1M cached input tokens and $25 per 1M output tokens, while OpenAI’s GPT‑5‑nano sits at $0.05 per 1M input, $0.005 per 1M cached input and $0.4 per 1M output. Google’s nearest competitor, Gemini‑3.5‑Flash‑Lite, asks $0.3 per 1M input, $0.03 per 1M cached input and $2.5 per 1M output. The gap is stark: Claude’s output price is ten times higher than GPT‑5‑nano and tenfold the Gemini flash tier, even though its input cost is only a few cents more.
### Where the money goes
The three tables below lay out the raw token rates for the flagship offering of each vendor. Claude Opus 5’s pricing is uniform across its recent sub‑versions, so the numbers you see are the same for Opus 5, 4.8, 4.7 and 4.6. OpenAI splits its tier into a nano model that still carries a safeguard tier (gpt‑oss‑safeguard‑20b) but the flagship GPT‑5‑nano is the only one that lines up with the "flagship" label in the prompt. Google’s Gemini line offers a flash‑lite variant that is the most comparable in latency and cost to the other two flagships, even though the brand name differs.
| Vendor | Model | Input | Cached input | Output |
|---|---|---|---|---|
| Anthropic | claude‑opus‑5 | $5 per 1M tokens | $0.5 per 1M tokens | $25 per 1M tokens |
| OpenAI | gpt‑5‑nano | $0.05 per 1M tokens | $0.005 per 1M tokens | $0.4 per 1M tokens |
| Google Gemini | gemini‑3.5‑flash‑lite | $0.3 per 1M tokens | $0.03 per 1M tokens | $2.5 per 1M tokens |
### Who each one is for
If you are building a mission‑critical, compliance‑heavy system where latency and hallucination risk are non‑negotiable, Claude Opus 5’s premium output price reflects its enterprise‑grade guardrails and the cache‑friendly pricing can still make sense at massive scale. For startups that need to prototype fast, iterate on prompts, and keep the bill under control, GPT‑5‑nano’s sub‑dollar output cost is hard to beat, especially when the cached input discount is already baked in. Gemini‑Flash‑Lite lands in the sweet spot for developers who want a middle ground: the input cost is modest, the output price is higher than GPT‑5‑nano but far below Claude, and the model shines on multimodal tasks like image‑augmented prompts.
In practice the choice narrows to budget versus risk tolerance. When every cent counts and you can tolerate a slightly higher hallucination profile, GPT‑5‑nano wins hands‑down. If you can justify the spend for tighter safety and the ability to cache large prompt fragments, Claude Opus 5 becomes the logical pick. For teams that need a balanced, cost‑effective flash model with decent safety, Gemini‑Flash‑Lite offers the most pragmatic compromise. Pricing verified on 2026-09-10.