Anthropic Claude Haiku 5.5 launched with deep price cuts as IPO details leak

Anthropic has launched its latest artificial intelligence model, Claude Haiku 5.5, slashing prices by an average of 75 per cent compared to its predecessor to match rival OpenAI in an intensifying price war.

The release of the lightweight model comes at a critical moment for the San Francisco-based startup. Details from Anthropic's confidential S-1 investor prospectus, obtained by Reuters, have leaked ahead of a planned initial public offering expected before Thanksgiving. The documents show the company's revenue skyrocketed twelvefold to $4.6 billion in 2025, but that growth came at an immense cost, with losses ballooning to $42 billion over the same period.

With the public listing looming, Anthropic is aggressively courting enterprise clients who are looking to rein in their ballooning AI budgets. Haiku 5.5 is specifically engineered for high-volume, repetitive tasks where speed and cost are more important than sheer computing muscle.

Slashed pricing to challenge OpenAI

Under the new pricing structure, developers will pay $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. That matches the pricing of OpenAI’s smallest model, GPT-6 Luna.

For larger jobs exceeding 100,000 tokens, which Anthropic says make up only about 10 per cent of previous Haiku workloads, the price rises to $0.50 for input and $2.50 for output.

The company is also reducing costs elsewhere in its lineup. It has halved the price of cache reads for its mid-tier Claude Sonnet 5.5 model to $0.10 per million tokens, a move it claims will make most agentic workflows about 20 per cent cheaper.

Faster speeds and desktop agents

Despite the lower price, Anthropic claims Haiku 5.5 easily outperforms the older Haiku 4.5 across coding, computer use, and general knowledge benchmarks. It has also introduced an "adjustable effort" setting, allowing developers to trade off processing speed against the intelligence of the output.

Software firm Asana, which tested the model early, reported a 30 per cent reduction in latency and up to 2.5 times faster response times when using Haiku 5.5 to triage bugs and search portfolios.

Haiku 5.5 is also being pitched as a lightweight companion to Anthropic's more powerful models, Sonnet and Opus. It can act as a "subagent" to handle quick, repetitive steps in complex workflows. The model is available immediately on major cloud platforms including Amazon Web Services, Google Cloud, and Microsoft Azure, and has built-in support for browser and computer automation tasks.

Safety and public market pressure

The launch of Haiku 5.5 highlights how quickly the commercial AI market is shifting from raw capability to cost efficiency. As companies move past experimental phases, they are increasingly demanding cheaper, smaller models or looking toward open-weight alternatives to control their spending.

To address safety concerns, Anthropic says Haiku 5.5 features tighter cybersecurity safeguards than its predecessor, blocking common attack techniques like penetration testing while allowing defensive tasks.

Both Anthropic and OpenAI are navigating these commercial pressures alongside regulatory scrutiny. Both firms have recently dealt with security fallouts from models accessing third-party networks, and Anthropic chief executive Dario Amodei recently attended a White House summit with President Trump to discuss self-regulation for the industry.

OpenAI is also preparing for a public debut, though its IPO has reportedly been pushed back to early 2027.