LLM API prices / Comparison
gpt-6.1-sol vs claude-sonnet-5: API pricing
Last checked: 2026-10-07. These are each provider's direct standard API text-token rates. The comparison describes published charges only; model quality, latency, availability and answer length need a workload-specific test. A lower token rate does not guarantee a lower invoice if one model produces more tokens or uses an extra billed feature.
| Model | Input / 1M | Output / 1M | Cached input / 1M | Context | Official source |
|---|---|---|---|---|---|
| gpt-6.1-sol (OpenAI) | $2 | $10 | $0.1 | 1,050,000 | Official · 2026-10-07 |
| claude-sonnet-5 (Anthropic) | $2 | $10 | $0.2 | TBA | Official · 2026-10-07 |
- gpt-6.1-sol: Standard short-context text rates; long-context (>272K), regional, tools and processing tiers can change billed cost.
- claude-sonnet-5: Base API tokens; cache read is a separate rate. Full 1M context is billed at standard rates; US-only inference_geo adds 1.1x.
Calculate the same workload
Assume 2,000 input tokens, 500 output tokens and 1,000 requests per day for 30 days. That produces 60 million input tokens and 15 million output tokens. For gpt-6.1-sol, 60 × $2 plus 15 × $10 equals $270.00. For claude-sonnet-5, 60 × $2 plus 15 × $10 equals $270.00. This arithmetic assumes every input token is billed at the base rate and every output token at the listed output rate. It excludes cache hits, writes, batch discounts, taxes, tools and regional charges.
Where the invoice can diverge
gpt-6.1-sol is billed by OpenAI, while claude-sonnet-5 is billed by Anthropic. Their tokenizers can split the same text differently, so a fixed character count does not imply equal token counts. The context limit indicates the maximum request size in the cited material when verified; it does not determine the price of an ordinary short request. Long-context thresholds, where published, can raise both input and output charges once a prompt crosses the threshold. If the context field is TBA, check the official model specification before planning a large prompt.
Cached input rates matter only when a request actually qualifies for a cache hit. Some providers also charge to create or store a cache, and a cache may expire. For a fair pilot, log the provider's usage counters for uncached input, cached input and output separately. Use those totals with the relevant rate, then compare the invoices. If your application handles images or audio, consult each provider's modality rules; these text-rate examples are not an estimate for media-heavy traffic.
How to choose
Begin with the cheapest combination that meets your own accuracy and latency target on representative requests. Test both models on the same prompts, record their actual token usage and note any failures requiring retries. Retries and longer generated answers can outweigh a small rate difference. Review the linked official pages again before committing to spend because API prices and availability change. The calculator lets you replace this example with your measured average tokens and daily volume.