finally a good small model

Summary

The video reviews Anthropic’s Haiku 5.5, arguing that it is the company’s first small model that is genuinely useful—especially as a fast, inexpensive helper for larger models.

Product and pricing

Anthropic positions Haiku 5.5 for high-volume, cost-sensitive work such as summaries, context compaction, database queries, classification, browser use, and customer support. It is also intended to work as a sub-agent alongside Sonnet or Opus.

The creator says Haiku 5.5 costs roughly 75% less to run than Haiku 4.5. The cited prices per million tokens are:

• Input: $0.10
• Output: $0.50
• Cache reads: $0.01
• Cache writes: $0.125

The low pricing applies to contexts up to 100,000 tokens. Above that threshold, the creator says prices increase fivefold. He recommends setting Claude Code’s Haiku auto-compaction window to 100,000 tokens when predictable costs matter.

The video also notes that Sonnet 5.5’s cache-read price was reduced from $0.20 to $0.10 per million tokens, making it more economical for agentic workloads.

Performance and comparisons

In the creator’s benchmark suite, Haiku 5.5 scores substantially better than Haiku 4.5, including moving from a zero score on Terminal Bench to nearly 40%. He also reports strong coding and computer-use results relative to some competing models.

Reported generation speeds range from 100–200 tokens per second; the creator saw about 180 tokens per second in his own use.

The creator cautions that benchmarks do not represent every task. His cost-versus-performance analysis generally favors cheaper models such as GPT-6 Luna or 6.1 Soul for some API workloads, while Haiku’s main advantage is fitting naturally into Anthropic-based workflows.

He views OpenAI’s 6.1 Soul as a useful independent auditor for Opus-generated code, while Haiku is better suited to delegated, lower-cost work.

Recommended uses and practical lessons

Best fit: Let a stronger model decide when to delegate repetitive or exploratory work to Haiku—for example, searching a large codebase, categorizing information, gathering data, or trying many candidate solutions in parallel.

Less suitable: Using Haiku as the primary coding agent for complex work. The creator argues that code generation itself is often cheap; gathering context, testing, verification, and recovering from mistakes are the more expensive parts. A weaker model can become costly if it struggles and repeatedly fails.

Parallel-agent example

In an Anthropic egg-drop demo, Opus working alone took about 3.5 minutes, tried 25 designs, and cost $0.47. Opus coordinating 10 Haiku sub-agents finished in under a minute, tried 86 designs, and cost $0.14. The example illustrates Haiku’s value when many attempts are more useful than one especially thoughtful attempt.

Creative demos and hands-on tests

A community video-generation example reportedly cost $0.60 and took under 20 minutes. Front-end design demos looked promising, particularly with Anthropic’s design skill enabled, but results varied.

In hands-on tests, a generated “Fish Slap” game had noticeable gameplay and control problems and cost about $1; many requests exceeded the 100,000-token pricing threshold. A PR-review task found potentially useful items but also relied on GitHub status information and stalled after hitting API rate limits.

The creator uses these examples to argue for having a stronger model orchestrate Haiku rather than relying on Haiku alone.

Other product notes

• Anthropic offers monthly API credits to some Claude subscribers. The creator says 5x-plan users receive $100 and 20x-plan users receive $200 per month.
• The Claude Python and TypeScript SDKs are adding support for computer use and browser use.
• Sponsor segment: CodeRabbit is presented as a code-review and security tool. The creator highlights pull-request security feedback, model-assisted fixes, and repository-wide scans that take roughly 30–60 minutes.

Overall assessment

The creator’s conclusion is that Haiku 5.5 is not a replacement for Opus or a general-purpose coding agent, but it makes Anthropic’s model lineup more useful for inexpensive, fast delegated work. He recommends using it selectively through agents while reserving harder planning and coding tasks for stronger models.

Main speakers and sources

• Theo, the video’s host and reviewer
• Anthropic, through Haiku 5.5 release information and cited employee comments
• Artificial Analysis, for benchmark and cost-performance data
• Community examples, including demos credited to Arshan and Dara
• CodeRabbit, the video sponsor