Haiku 5.5 Finally Fills Anthropic's Small Model Gap — With a Trap

The breakthrough. Anthropic just released Haiku 5.5, a small model that finally fixes the glaring hole in their lineup. The previous Haiku 4.5 was so bad and expensive that its own user base actively avoided it, going so far as to build workarounds using OpenAI models for cheap tasks. This release changes that.

The pricing trap. The headline pricing hides a sharp catch. The cheap rates only apply up to 100,000 tokens of context. Exceed that threshold and costs quintuple to $2.50 per million output tokens. This is intentional. Anthropic designed the pricing to steer users toward the model's intended high-volume, low-latency use cases like classification, summarization, and database queries rather than long-running agentic work. The model output speed reaches 100-200 tokens per second, making it viable for customer support and rapid categorization.

The performance trade-off. Early testing reveals a sharp divergence depending on how you use the model. With reasoning enabled, Haiku 5.5 performs well for its price tier, beating Gemini 2.0 Flash on coding benchmarks. With reasoning disabled, it falls apart. Users report worse pass rates, higher failure rates, and slower performance compared to competing models like GPT-4o Mini. This matters because some workflows deliberately disable reasoning for speed, and Haiku 5.5 is not trained to handle that mode effectively.

The real use case. The smartest application is not selecting Haiku as your primary coding model. It is using it as a cheap sub-agent that smarter models like Opus or Sonnet delegate tasks to. Anthropic demonstrated this convincingly: Opus 5.5 alone solved an egg-drop simulation in 3 minutes at $0.47 across 25 attempts. Opus commanding 10 Haiku 5.5 sub-agents solved it in under 1 minute at $0.14 across 86 attempts. The principle holds for any task where brute-force parallel attempts beat single smart attempts.

The hidden gem. Alongside the Haiku launch, Anthropic halved the cache-read price for Sonnet 5.5 from $0.20 to $0.10 per million tokens. This is a significant change for anyone doing agentic work where cache reads can represent 30-50% of the total bill. Combined with the new $200 monthly API credit for Pro subscribers, the pricing structure is becoming genuinely competitive with OpenAI's ecosystem.

The bottom line. Haiku 5.5 is finally a small model worth using within the Anthropic ecosystem, but its value is contextual. It will not replace Opus for complex reasoning or coding. It is a superb tool for cheap, fast, parallel sub-tasks that smarter models orchestrate. Teams that treat it as a general-purpose replacement will be disappointed. Teams that embed it as a delegable worker within an agentic architecture will find real cost savings.