A practical, numbers-first comparison of API token prices for OpenAI, Anthropic and Google models — with worked cost examples and context-window limits.
All figures are public list input prices in USD, as of November 2026. Output tokens are billed higher (see below).
| Model | Provider | Input $/1M | Context |
|---|---|---|---|
| GPT-6.2 Astra | OpenAI | $8.00 | 1.05M |
| GPT-5.5 | OpenAI | $5.00 | 1.05M |
| Claude Opus 5.5 | Anthropic | $4.00 | 1M |
| GPT-5.4 | OpenAI | $2.50 | 1.05M |
| GPT-6.1 Sol | OpenAI | $2.00 | 1.05M |
| Claude Sonnet 5.5 | Anthropic | $2.00 | 1M |
| Gemini 3.1 Pro | $2.00 | 1M | |
| GPT-5.4 mini | OpenAI | $0.75 | 1.05M |
| Gemini 3.7 Flash | $0.75 | 1M | |
| GPT-5 mini | OpenAI | $0.25 | — |
| Claude Haiku 5.5 | Anthropic | $0.10 | 1M |
| GPT-5 nano | OpenAI | $0.05 | — |
Context values are maximum context windows. Some vendors bill long-context requests at a higher tier. Always confirm on the official pricing page before budgeting.
Cost is simply (tokens ÷ 1,000,000) × price. Here is what common workloads cost at input price:
| Workload | GPT-6.2 Astra ($8) | Claude Sonnet 5.5 ($2) | GPT-5 nano ($0.05) |
|---|---|---|---|
| 10K tokens | $0.080 | $0.020 | $0.0005 |
| 100K tokens | $0.80 | $0.20 | $0.005 |
| 500K tokens | $4.00 | $1.00 | $0.025 |
| 1M tokens | $8.00 | $2.00 | $0.05 |
The gap compounds fast. A workload processing 20M input tokens per month costs about $160 on GPT-6.2 Astra versus $40 on Claude Sonnet 5.5 versus $1 on GPT-5 nano — a 160× spread for the same token volume.
The prices above are input only. For every family, generated (output) tokens are billed at a higher rate — commonly 2–4× the input price. Chat and agent workloads often produce far fewer output tokens than they consume, so input price is a good first-order estimate. But for generation-heavy tasks (long-form writing, code synthesis), budget using the output rate instead.
It depends on the tier. Like-for-like mid-tier models (GPT-6.1 Sol and Claude Sonnet 5.5) are both $2 / 1M input tokens. The flagships differ: GPT-6.2 Astra is $8 versus Claude Opus 5.5 at $4, so Claude's top model is half the input price of OpenAI's.
Yes — all three providers offer prompt caching and batch discounts that can reduce effective input cost substantially. Those discounts are not reflected in list prices, so heavy users should model them separately.
Use the free token counter on this site. It runs the real o200k_base / cl100k_base encoders locally in your browser and shows the estimated cost for your exact text across all the models above. For a deeper explanation of tokens, read the What are tokens? guide.