OpenAI and Anthropic can land surprisingly close on standard token cost, but the similarity disappears when context bands, batch and fast service are mixed into one table.
This comparison uses exact first-party API identities from the Kingy.ai AI API Price Index, release kapi-price-index-2026-08-23T042959Z. Prices are US dollars per one million tokens. The representative workload is one million input plus 250,000 output tokens.
| Exact offering | Tier | Input | Cached input | Output | Workload |
|---|---|---|---|---|---|
GPT-5.4 (gpt-5.4) |
standard, ≤272K | $2.50 | $0.25 | $15.00 | $6.25 |
GPT-5.4 mini (gpt-5.4-mini) |
standard | $0.75 | $0.075 | $4.50 | $1.875 |
GPT-5.5 (gpt-5.5) |
standard, ≤272K | $5.00 | $0.50 | $30.00 | $12.50 |
Claude Haiku 4.5 (claude-haiku-4-5-20251001) |
standard | $1.00 | $0.10 | $5.00 | $2.25 |
Claude Sonnet 4.6 (claude-sonnet-4-6) |
standard | $3.00 | $0.30 | $15.00 | $6.75 |
Claude Opus 4.8 (claude-opus-4-8) |
standard | $5.00 | $0.50 | $25.00 | $11.25 |
The closest comparison is GPT-5.4 against Claude Sonnet 4.6
Both charge $15 per million output tokens in the accepted standard band. GPT-5.4 charges $0.50 less on input, so the representative workload is $6.25 versus $6.75. That eight-percent difference is real, but small enough that task success and latency can erase it quickly.
The cheaper tier has a different shape
GPT-5.4 mini costs $1.875 for the representative workload; Claude Haiku 4.5 costs $2.25. Haiku’s input is higher, while its output rate is only $0.50 above the mini. Output-heavy and input-heavy applications therefore see different relative gaps.
Premium pricing crosses
GPT-5.5 and Claude Opus 4.8 both charge $5 for standard input and $0.50 for cached input. Anthropic’s $25 output price puts the representative workload at $11.25, below GPT-5.5 at $12.50. That is a price result, not a quality verdict.
Batch helps Anthropic—but it is a separate service
Anthropic officially documents 50% batch prices for the accepted Claude models: Haiku at $0.50/$2.50, Sonnet at $1.50/$7.50 and Opus at $2.50/$12.50. The index does not copy that discount onto OpenAI rows without exact model-level evidence in the captured release.
Likewise, Claude Opus 4.8’s fast research preview is recorded at $10 input and $50 output. It should not appear as “the Opus price” without its tier label.
Long context changes OpenAI’s rate
Above 272,000 tokens, GPT-5.4 moves to $5 input and $22.50 output; GPT-5.5 moves to $10 and $45. Those rows are deliberately separate. The accepted Anthropic rows in this release do not invent an equivalent context threshold where the official evidence did not define one for this comparison.
Bottom line
For standard token cost alone, GPT-5.4 narrowly undercuts Claude Sonnet 4.6 in the representative workload, GPT-5.4 mini undercuts Claude Haiku 4.5, and Claude Opus 4.8 undercuts GPT-5.5. Batch can change Anthropic’s economics, while very long context can change OpenAI’s.
Use the interactive index to select the exact route and tier, then test quality on your own workload before making a buying decision.
Official evidence: OpenAI model documentation, Anthropic pricing, and the index’s versioned source-capture manifest.