Last week, Anthropic released Claude Haiku 5.5, and the pricing caught my attention: $0.10 per million input tokens for requests under 100K context. That's 75% cheaper than Haiku 4.5.
But price cuts alone don't matter if performance tanks. So I ran some tests.
The Numbers
Here's what changed:
| Model | Input (≤100K) | Output (≤100K) | OSWorld 2.1 |
|---|---|---|---|
| Haiku 4.5 | $1.00 | $5.00 | 15.7% |
| Haiku 5.5 | $0.10 | $0.50 | 72.4% |
That's not a typo. The cheaper model scores 4.6x better on OSWorld's offline subset.
For context above 100K tokens, pricing jumps to $0.50/$2.50 per million—still 50% below Haiku 4.5's rates.
Anthropic also cut Sonnet 5.5 cache-read prices from $0.20 to $0.10 per million tokens. If you're running agentic workflows with repeated context, that's roughly 20% off your total bill.
Performance Beyond Benchmarks
I tested Haiku 5.5 on three real-world tasks:
1. Code summarization (500+ line files)
- Haiku 4.5: Missed edge cases, needed manual cleanup
- Haiku 5.5: Caught 90%+ of issues on first pass
2. Multi-turn customer support routing
- Haiku 4.5: Confused on ambiguous intents ~15% of the time
- Haiku 5.5: Dropped to ~3% error rate
3. Document classification (10K+ token docs)
- Both models performed similarly, but Haiku 5.5 was 2x faster
The new adjustable effort setting is interesting. You can tell the model to spend more or less compute per task. For simple classification, I set effort to "low" and got responses in under 200ms. For complex reasoning, "high" effort took 2-3x longer but caught nuances the lower setting missed.
When to Use What
Haiku 5.5 makes sense for:
- High-volume classification and routing
- Summarization tasks where speed matters
- Agent subtasks that run frequently
- Cost-sensitive applications with tight margins
Sonnet 5.5 or Opus 5.5 for:
- Complex multi-step reasoning
- Long-form content generation
- Tasks where accuracy trumps cost
The Aggregator Advantage
Here's the thing: every few months, a new model drops with better pricing or performance. If you're hardcoding to a single provider, you're stuck re-integrating every time.
We built RollTok to solve this. One API key, one endpoint, access to Haiku 5.5, Sonnet 5.5, Opus 5.5, GPT-6 variants, Gemini models—whatever fits your task. When Anthropic cuts prices, you get the savings automatically. When a new model launches, you can test it without changing your codebase.
For teams running multiple workloads, that flexibility adds up. You're not locked into one vendor's roadmap.
Bottom Line
Haiku 5.5 isn't just cheaper—it's genuinely better for most high-volume tasks. The 75% price cut makes it viable for applications that were previously too expensive to automate.
If you're building anything that runs LLM calls at scale, the math has shifted. Test it against your current setup. You might find the cost-performance tradeoff looks very different than it did six months ago.
Want to test multiple models without managing separate integrations? RollTok gives you one API for 50+ models. Get started here
Top comments (0)