Haiku 4.5 Is Here—And It’s a Beast at Coding

Key Concepts

• Vio 3.1: A new release with enhanced capabilities for video generation via API, including multi-image input, longer video creation (up to 1 minute), and the use of first/last frames.
• Gemini API: The interface for interacting with Google's Gemini models, now integrated with Vio 3.1.
• Cloud Haiku 4.5: The latest model from Anthropic, characterized by its smaller size, affordability, and speed. It is positioned as a "workhorse" model.
• Anthropic: A company developing AI models, including the Claude series.
• Sonnet 4.5: A previous model from Anthropic, used as a benchmark for Haiku 4.5.
• Sweet Benchmark: A benchmark that Anthropic reportedly prioritizes.
• Gemini 2.5 Pro: A model from Google, compared against Haiku 4.5 on benchmarks.
• Gemini 3: An anticipated upcoming model from Google.
• Claude: Anthropic's suite of AI models.
• Plan Mode (Claude): A feature where Sonnet 4.5 handles planning and Haiku 4.5 handles execution for complex tasks.
• Computer Use Agent: An AI agent capable of performing actions on a computer.
• Evy: A company that conducted testing on Haiku 4.5.
• GPT 5 Mini: A model from OpenAI, used for pricing comparisons.
• Gemini 2.5 Flash: A model from Google, used for pricing comparisons.
• Token Efficiency: The ability of a model to process information using fewer tokens.
• Sora 2: A video generation model, used for pricing comparisons with Vio 3.1.
• Frontier Labs: Refers to leading AI research and development companies.

Vio 3.1 and Gemini API Enhancements

Vio 3.1 has been released with new features for its Gemini API integration. Key advancements include:

• Multi-Image Driving: The ability to generate video using multiple input images.
• Extended Video Length: Users can now create longer videos, up to one minute in duration.
• First and Last Frame Utilization: The capability to use the first and last frames of a sequence for video generation.

The API structure for Vio 3.1 is highlighted as being user-friendly for developers. The new model allows for the provision of reference images, enabling the combination of visual elements based on user descriptions. Additionally, the "extend screens" feature uses the final seconds of a previous video to further extend the generated content. The ability to provide specific first and last frames facilitates transitions between defined start and end points.

Vio 3.1 Pricing

• Default Version (with audio): 40 cents per second.
• Vio 3.1 Fast (with audio): 15 cents per second.

This pricing is noted as being relatively more expensive than Sora 2, which starts at 10 cents per second for its lower version and can go up to 50 cents per second for higher resolutions.

Cloud Haiku 4.5: A New Workhorse Model

Anthropic has released Cloud Haiku 4.5, a model positioned as a significantly more affordable and faster alternative to previous Claude models.

Key Features and Performance

• Size and Cost: Haiku 4.5 is considerably smaller and approximately one-third the cost of comparable models like Sonnet 4.5.
• Speed: It offers more than twice the speed of Sonnet 4.5.
• Benchmark Performance:
  • On the Sweet Benchmark, Haiku 4.5 surpasses the previous Sonnet 4.5 model.
  • It also outperforms previous state-of-the-art models on other key benchmarks.
  • Notably, it beats Gemini 2.5 Pro on a number of key benchmarks.
• "Workhorse" Positioning: Anthropic aims to portray Haiku 4.5 as a model suitable for tasks that do not require the full intelligence of models like Sonnet 4.5, emphasizing speed, token efficiency, and cost-effectiveness.
• Plan Mode Integration: When used in plan mode within Claude, Haiku 4.5 automatically leverages Sonnet 4.5 for planning complex tasks and then uses Haiku 4.5 for execution. This is seen as a good balance for simpler tasks or when complex tasks can be broken down into subtasks.

Pricing Comparison

Haiku 4.5 is priced at $1 per million input tokens and $5 per million output tokens. This is compared to:

• GPT 5 Mini: 25 cents per million input tokens and 2.5 cents per million output tokens.
• Gemini 2.5 Flash: Similar pricing to GPT 5 Mini.

This indicates that Haiku 4.5 is significantly more expensive on a per-token basis compared to GPT 5 Mini and Gemini 2.5 Flash, despite its superior capabilities according to benchmarks. The speaker notes that Anthropic has not reduced pricing with model upgrades, unlike other companies.

Real-World Application: Computer Use Agent Demonstration

A demonstration was conducted using a computer use agent with Haiku 4.5 to perform a task: navigating to the IRS website and downloading the W9 form.

Step-by-Step Process and Observations

1. Task Assignment: The agent was instructed to "go to the IRS website and download the W9 form for me."
2. Action Execution:
   • The agent navigated to the IRS website.
   • It took screenshots after each action, with a two-second delay to ensure action completion.
   • It searched for forms and instruction links.
   • It then attempted to find the search box.
3. Challenges and Delays:
   • The agent initially seemed to try using a direct URL, which failed.
   • It then reverted to searching for the form.
   • The process was observed to be relatively slow compared to Gemini's computer use agent, though faster than Sonnet 4.5 for this specific task.
   • The agent took approximately 1 minute and 40 seconds before asking for confirmation to download the file.
   • After confirmation, it opened the file, completing the task in about 2 minutes and 30 seconds.
4. Comparison with Sonnet 4.5:
   • The same task was repeated with Sonnet 4.5.
   • While Sonnet 4.5 also completed the task, it took approximately 3 minutes and 30 seconds.
   • This suggests that Haiku 4.5, despite its perceived slowness in the initial steps, was faster overall for this particular computer use task compared to Sonnet 4.5.

Evy's Testing Results

A company named Evy conducted tests on Haiku 4.5 for a "wipe check and topic cooked" scenario. Their findings indicated that Haiku 4.5 performed similarly to Sonnet 4.5 but was significantly faster and cheaper.

Synthesis and Conclusion

The YouTube video highlights two significant AI releases: Vio 3.1 with enhanced Gemini API capabilities for video generation and Anthropic's Cloud Haiku 4.5, a new, cost-effective, and fast language model.

Vio 3.1 introduces features like multi-image input, longer video generation (up to 1 minute), and the use of first/last frames, with pricing at 40 cents/second for the default version and 15 cents/second for the fast version.

Cloud Haiku 4.5 is presented as a powerful "workhorse" model, offering substantial speed and cost improvements over Sonnet 4.5, while maintaining competitive or superior performance on key benchmarks, including outperforming Gemini 2.5 Pro in certain areas. Despite its affordability and speed, its per-token pricing is higher than models like GPT 5 Mini and Gemini 2.5 Flash. A demonstration of a computer use agent with Haiku 4.5 showed it completing a task faster than Sonnet 4.5, though with some initial delays.

The speaker concludes that while model capabilities are converging among leading labs, the focus is shifting towards the cost-effectiveness and application-specific utility of these models. The choice between models will increasingly depend on the specific use case, budget, and desired balance between speed, intelligence, and price.