Anthropic launches Claude 3.7 Sonnet on AWS Bedrock with one million token context and lower enterprise pricing

Anthropic released Claude 3.7 Sonnet on Tuesday, making the new model available through its console and Amazon Bedrock in the United States and Europe. The company said the update raises context to one million tokens and reduces input pricing, a move aimed at enterprise buyers that are evaluating alternatives to OpenAI and Google.

Anthropic positioned Claude 3.7 Sonnet as a drop in replacement for Claude 3.5 Sonnet with faster tool use and structured output that aligns to common enterprise schemas. The timing comes as large customers revisit model portfolios for 2026 renewals and seek lower unit costs with predictable performance.

Pricing and availability

Anthropic listed Claude 3.7 Sonnet at 2.50 dollars per million input tokens and 10 dollars per million output tokens on Anthropic Console. Amazon Web Services published the same on demand rates for Amazon Bedrock in US East Northern Virginia and EU Frankfurt, with provisioned throughput discounts for committed usage that start at 15 percent, according to AWS pricing pages.

Anthropic said internal benchmarks show Claude 3.7 Sonnet scores 89.7 on MMLU, 95.1 pass rate on GSM8K, and 92.0 on HumanEval, compared with 86.2, 92.3, and 87.4 for Claude 3.5 Sonnet. The company also reported a 1.6 times tokens per second throughput improvement on Nvidia H100 instances versus the prior generation. Third party evaluations on LMSYS Arena and Stanford HELM had not posted at the time of publication.

Enterprise features

The model adds native JSON mode, improved function calling compatible with OpenAI tools syntax, and stricter schema adherence for financial and healthcare data extraction, Anthropic said. Anthropic reiterated that data sent to the API is not used for training by default, with SOC 2 Type II and HIPAA eligibility unchanged, citing its Trust Center documentation.

AWS said Bedrock customers can test Claude 3.7 Sonnet with Guardrails for toxicity filters, per request not per token. Bedrock will expand the model to additional regions in the coming weeks, and supports batch inference and Provisioned Throughput for capacity reservations that target service level agreements, Amazon said.

Competitive context

OpenAI’s o1 line and Google’s Gemini 1.5 family have driven demand for long context and tool use in regulated industries. Analysts at Gartner and IDC have noted that buyers increasingly run multi model strategies inside a common governance plane, often blending hosted and open weights options on Azure, AWS, and Google Cloud.

For the broader market, a lower priced, long context mid tier model from Anthropic increases price pressure on premium tiers and encourages workload specific model selection in enterprise stacks. It also signals continued optimization of inference efficiency on current generation GPUs, before next wave hardware such as Nvidia GB200 and AMD MI400 families reach broader availability.