Anthropic releases Claude 3.5 Haiku at $1 per million input tokens, up 4x from Claude 3.0 Haiku's ¢25 per million tokens, and without image analysis features
Anthropic released Claude 3.5 Haiku today, a few days later than expected … X: @anthropicai : During final testing, Haiku surpassed Claude 3 Opus, our previous flagship model, on many benchmarks—at a fraction of the cost. As a result, we've increased pricing for Claude 3.5 Haiku to reflect its increase in intelligence: https://anthropic.com/.... @anthropicai : Claude 3.5 Haiku is now available on our API, Amazon Bedrock, and Google Cloud's Vertex AI. Haiku is fast and particularly strong at coding. It outperforms state-of-the-art models—including GPT-4o—on SWE-bench Verified, which measures how models solve real software issues. [image] Paul Gauthier / @paulgauthier : New Claude 3.5 Haiku comes in 4th on the aider code editing leaderboard with 75%. This is just behind the old 3.5 Sonnet 06/20. 84% Sonnet 10/22 80% o1-preview 77% Sonnet 06/20 75% Claude 3.5 Haiku 47% Claude 3 Haiku https://aider.chat/... [image] Simon Willison / @simonw : Claude 3.5 Haiku is out - two surprises: 1. It's priced differently from Claude 3 Haiku. 3.5 Sonnet had the same price as 3 Sonnet, but 3.5 Haiku costs ~4x more than 3 Haiku did 2. No image input support yet 3.5 Haiku beats 3 Opus though, and Opus cost 15x the new Haiku price! James Grugett / @jahooma : We just shipped the new Claude 3.5 Haiku to production! It's faster and one third the cost of Sonnet, while still being really good at picking relevant files for your request. Impact on Manicode usage: - 2x faster file picking step - 33% cheaper overall Bindu Reddy / @bindureddy : Haiku 3.5 is 4x more expensive than Haiku 3.0 Strange move by Anthropic. They claim it's because Haiku is a good model. It doesn't make much sense... With all fast-moving tech, we should see both costs go down by 10x while performance increases by 10x The correct answer was @yawnxyz : Haiku 3.5 is now Anthropic's cheapest model and it's a disappointment. At $1/M in and $5/M out, it's almost as much as gpt-4o ($2.5M / $10M). While some of us though Haiku would slay 4o-mini ($0.15M/$0.6M), it doesn't. It just awkwardly sits somewhere in between. [image] Mike Krieger / @mikeyk : 3.5 Haiku is out! Impressively, our smallest model now outperforms Claude 3 Opus, our largest model from earlier this year, on many benchmarks, at a fraction of the cost. Swami Sivasubramanian / @swamisivasubram : Excited to share that Anthropic's Claude 3.5 Haiku is available for customers to use in Amazon Bedrock: https://aws.amazon.com/... This model shows improvements on speed and intelligence and can speed up response times for customer service, e-commerce solutions, and education
Context & Ripple Effects
Anthropic positioned the original Haiku as a fast, low-cost option for high-volume, latency-sensitive use cases in its initial Haiku rollout. The Claude 3 family had already established a tiered lineup spanning Haiku, Sonnet, and Opus.
This release changes the economics of the small-model tier: Anthropic is charging more than it did for Claude 3 Haiku while claiming performance that reaches beyond its former flagship on some benchmarks.
First-order effects
- API customers on Anthropic, Amazon Bedrock, and Google Cloud Vertex AI get a faster coding-oriented Haiku option, but face a fourfold increase in input-token pricing versus Claude 3 Haiku.
- Workloads requiring image analysis cannot consolidate on Claude 3.5 Haiku, despite its claimed text and coding gains.
Second-order effects
- Buyers will need to evaluate cost per completed task rather than token price alone: the reported production result of faster file selection and lower overall cost suggests higher-priced inference can still reduce application spend.
- Anthropic’s cloud-distribution partners gain another differentiated model tier, while competing model providers face pressure to match the combination of low latency, coding performance, and usable cost.
Third-order effects
- The small-model category is shifting from “cheapest tokens” toward segmented pricing based on capability and workload value, an instance of Haiku’s original high-volume positioning evolving into a stronger performance tier.
- If this pattern persists, model selection will become more workload-specific—text, coding, and multimodal tasks may use separate models—rather than a single low-cost model serving every request.
The trend: Frontier-model vendors are raising the value ceiling of smaller models while pricing inference around useful application outcomes rather than raw token volume.