LLM API Pricing Comparison: OpenAI, Anthropic, and Google by Tier

The three tiers, side by side

All prices below are standard, short-context, per million tokens, copied from each vendor's own price page on 2026-10-13. Cached input is the price of reading a repeated prompt prefix. A tier here means a rough capability class, so the pairing is a judgment call: check quality on your own task before you treat two models as interchangeable.

The top tier is Claude Fable 5.1 at $10 input, $0.25 cached input, and $50 output; gpt-6-astra at $10, $1.00, and $50; and Gemini 4 Argon at $2, $0.10, and $10 through November 30, 2026, rising to $4, $0.20, and $20 on December 1. The middle tier is Claude Sonnet 5.5 at $2, $0.10, and $10; gpt-6.1-sol at $2, $0.10, and $10; and Gemini 3.8 Flash at $0.75, $0.075, and $3.75 through December 31, 2026, rising to $1.50, $0.15, and $7.50 on January 1, 2027.

The small tier is Claude Haiku 5.5 at $0.10, $0.01, and $0.50 for prompts up to 100,000 tokens; gpt-6-luna at $0.10, $0.01, and $0.50; and Gemini 3.5 Flash-Lite at $0.30 input and $2.50 output for text.

One workload priced on every model

Take a workload of 10 million input tokens and 2 million output tokens per month, with no caching and no batching. Input cost is 10 times the input rate and output cost is 2 times the output rate.

• Top tier: Claude Fable 5.1 costs $200 and gpt-6-astra costs $200. Gemini 4 Argon costs $40 through November 30 and $80 from December 1.
• Middle tier: Claude Sonnet 5.5 costs $40 and gpt-6.1-sol costs $40. Gemini 3.8 Flash costs $15 through December 31 and $30 from January 1.
• Small tier: Claude Haiku 5.5 costs $2 and gpt-6-luna costs $2. Gemini 3.5 Flash-Lite costs $8.

What moves the number more than the vendor

Tier choice is the biggest lever. On this workload gpt-6-astra costs $200 and gpt-6-luna costs $2, a 100x difference, while the gap between OpenAI and Anthropic at any single tier is zero. If a small model passes your quality check on a request type, routing that request there saves far more than switching vendors.

Caching is next. If 8 million of the 10 million input tokens are cache reads, Claude Sonnet 5.5 costs 2 million at $2 plus 8 million at $0.10 plus the $20 output charge, which is $24.80 instead of $40. gpt-6.1-sol gives the same $24.80. That excludes cache writes, which are priced above uncached input: $2.50 per million on both models, and Anthropic lists a separate $4 rate for one-hour writes.

Batch is the third. Anthropic lists Claude Sonnet 5.5 at $1 input and $5 output in batch, and OpenAI lists gpt-6.1-sol at the same $1 and $5, so the middle-tier workload above costs $20 in batch. Claude Haiku 5.5 and gpt-6-luna are both $0.05 input and $0.25 output in batch, which makes the small-tier workload $1.

Pricing rules that break simple comparisons

Claude Haiku 5.5 is priced by prompt length. Requests with prompts over 100,000 tokens pay $0.50 input and $2.50 output, five times the base rate, and Anthropic says the prompt length counts cache reads and cache writes. On the workload above, that would turn a $2 Haiku bill into $10 if every request crossed the threshold.

OpenAI lists a second, higher column for long context on the gpt-6 models. For gpt-6.1-sol it is $4 input and $15 output, so long prompts double the input rate. Google's two scheduled increases matter for planning: a team that budgets Gemini 4 Argon at today's $40 will see $80 on the same traffic after December 1. Rates also differ by service mode, so confirm whether you are looking at standard, batch, flex, or priority rows before you copy a number.