Anthropic launches haiku 4.5, a lightweight model that achieves 73% in SWE-Bench and 41% in Terminal-Bench at a lower cost

By Padfoot

Anthropic launched Claude Haiku 4.5, a small model designed to offer similar performance to larger models at a lower cost and with reduced latency, and made it immediately available on all of its free plans.

***

Haiku 4.5 achieves 73% in verified SWE-Bench and 41% in Terminal-Bench, according to tests cited by TechCrunch.
Anthropic claims that the model offers performance comparable to Sonnet 4 and GPT-5 and Gemini 2.5 in various benchmarks, at a third of the cost and more than twice the speed.
The light version comes for use cases with high sensitivity to latency, such as parallel agent development and deployment tools.

🚀 Amazing launch! Anthropic presents Haiku 4.5.

Lightweight model achieves 73% in SWE-Bench and 41% in Terminal-Bench.

It promises similar performance to Sonnet 4 and GPT-5, at a third of the cost and more than twice the speed.

Now available in all free plans, ideal… pic.twitter.com/igUTAzcRFv

— Diario฿itcoin (@DiarioBitcoin) October 16, 2025

On October 15, 2025, Anthropic presented a new version of its reduced model, Claude Haiku 4.5. The company reported that this iteration seeks to deliver a balance between capacity, speed and cost. TechCrunch, in an article written by Russell Brandom published at 10:00 PDT, covered the launch and provided the company’s data and statements.

What anthropic and benchmarks said

Anthropic claimed that Haiku 4.5 can deliver performance similar to Sonnet 4 at one-third the cost and more than twice the speed. That assertion comes from a company blog post cited by TechCrunch.

In the internal tests cited, Haiku scored 73% in verified SWE-Bench and 41% in Terminal-Bench, which is focused on command-line tasks. TechCrunch pointed out that these results are below Sonnet 4.5, but tie with the reports of Sonnet 4, GPT-5 and Gemini 2.5 in the same benchmarks.

In addition to SWE-Bench and Terminal-Bench, Anthropic published comparisons in benchmarks aimed at tool use, computer use, and visual reasoning. In all these test sets, the company stated that Haiku 4.5 exhibits similar results to larger models, according to the report cited by TechCrunch.

The company presented these figures as evidence that a lightweight model can sustain significant capabilities without imposing the same computational load as larger models. TechCrunch reproduces the metrics and puts them in the context of the company’s recent launches.

Availability, costs and implementation proposals

Anthropic indicated that Haiku 4.5 will be available immediately under all of the firm’s free plans. The company believes that this option will be especially attractive for free versions of AI products, where reducing the load on servers and minimizing latency are priorities.

The company’s strategy, explained in statements cited by TechCrunch, is to allow new production configurations. Mike Krieger, product director at Anthropic, stated that Haiku will allow models of different complexity to be deployed within the same architecture. Krieger said: “It opens up entirely new categories of what is possible with AI in production environments, with Sonnet handling complex planning while Haiku-powered subagents execute at speed.”

The lightweight nature of the model makes it easy to run multiple Haiku agents in parallel or combine them with a more sophisticated model. In practice, this can translate into architectures where Sonnet acts as a planning brain and Haiku as a fast executor of specific tasks.

Anthropic emphasizes that cost savings and latency improvements respond to an intentional design for scale and high concurrency scenarios. TechCrunch adds that the company has promoted Haiku as a tool to reduce operational costs in massive deployments.

Use cases and market reactions

The most immediate applications of Haiku 4.5 will likely be in software development tools, where Claude Code is already widely used and latency is often critical. Anthropic and TechCrunch highlight that the combination of speed and cost is attractive for response time-sensitive workflows.

In statements provided by Anthropic, Andrew Filev, CEO of Zencoder, described the new version of Haiku as “unlocking a whole new set of use cases.” That opinion suggests commercial interest in adopting Haiku for functions that require fast, parallel execution.

Outside of software development, architectures with Haiku subagents could power hybrid conversational assistants, automation pipelines, and task orchestration in the cloud. The alternative of using smaller models for one-off tasks can reduce overall latency and cost per query.

Analysts and operators will closely watch whether Haiku 4.5’s performance holds up in independent testing and production environments. TechCrunch collects the data published by Anthropic, but warns that final comparisons will depend on external evaluations and real-world testing.

Launch context and next steps

Haiku 4.5 arrives a few weeks after the release of Sonnet 4.5, announced by Anthropic two weeks earlier, and approximately two months after Opus 4.1, another of its recent updates. The previous version of Haiku was released in October 2024, according to the timeline cited by TechCrunch.

The pace of releases shows the company’s strategy: frequent iterations and specialized models for different workloads. Anthropic seeks to position Haiku as the lightweight piece in an agent toolbox, complementary to its larger models.

For technical teams, immediate availability on free plans can accelerate proofs of concept and pilots. If Haiku 4.5 maintains the published metrics, organizations with cost or latency constraints could integrate lightweight subagents into their pipelines.

TechCrunch, in its coverage written by Russell Brandom, reports Anthropic’s statements and data and contextualizes the launch within the company’s series of recent moves. It remains to be seen how competitors will respond and what independent validations will confirm the initial numbers.

Original image from DiarioBitcoin, created with artificial intelligence, free to use, licensed under Public Domain.
This article was written by an AI content writer and reviewed by a human editor to ensure quality and accuracy.

WARNING: DiarioBitcoin offers informative and educational content on various topics, including cryptocurrencies, AI, technology and regulations. We do not provide financial advice. Investments in crypto assets are high risk and may not be suitable for everyone. Do your research, consult an expert and check applicable laws before investing. You could lose all your capital.