Claude
Claude is a proprietary frontier-model developed by AI safety research lab Anthropic. The largest variant, Opus, matches or surpasses GPT-4 on many benchmarks. Anthropic focuses on AI safety with the development of their models, and leverages a novel technique called Constitutional AI.
Airtrain supports Claude models by calling the Anthropic API on your behalf. You do not need your own API key to call Anthropic via Airtrain.
Supported Variants
Claude 3 generation
Opus
The largest and most capable (though also most expensive) variant of Claude 3, this beats GPT-4 on most standard benchmarks, with particularly notable performance on math and coding benchmarks (ex: on grade school math benchmark GSM8K, Claude 3 Opus scores 95.0% vs GPT-4s 92.0% and on the code benchmark HumanEval Opus scores 84.9% against GPT-4's 67.0%).
Sonnet
The intermediate variant of Claude 3, this is still competitive with GPT-4 on many benchmarks but being priced at about a quarter of the price.
Haiku (coming soon)
The smallest and fastest variant of Claude 3 and is more capable than models in the class of GPT-3.5 and Mixtral by most standard benchmarks.
Legacy
2.1 Full
The most capable Claude variant before the creation of the Claude 3 series. In general this model may be slightly less capable than "Haiku" while at a price and latency point more similar to Opus.
1.2 Instant
An older variant of Claude optimized for inference speed while having capabilities below those of Haiku and pricing similar to Sonnet.
Updated 9 months ago