Anthropic has released Claude Haiku 5.5, a small model designed to own the cost-sensitive end of the AI inference market. The pricing is aggressive: $0.10 per million input tokens for prompts under 100K tokens, down roughly 75% from Haiku 4.5's $1.00 rate. Cache reads drop to $0.01 per million tokens at the lower tier. For 90% of existing Haiku workloads — the ones under 100K tokens — this is a dramatic repricing of what it costs to run AI at scale. The performance story is real but bounded. Haiku 5.5 posts 72.4% on OSWorld 2.1 (computer use), 45.9% on Humanity's Last Exam (no tools), and 39.2% on Terminal-Bench 4.0 (agentic coding). These numbers beat Haiku 4.5 by wide margins and outperform GPT-6 Luna on several benchmarks. But Anthropic is careful to frame the model as a workhorse, not a thinker — it's for summaries, classification, database queries, and subagent tasks, not the complex multi-step coding where Sonnet 5.5 and Opus 5.5 still dominate. The simultaneous 50% cut to Sonnet 5.5 cache reads is the subtler move. Because cache reads dominate token consumption on agentic tasks, this translates to roughly 20% cheaper agentic work on Sonnet. Anthropic is effectively creating a two-tier pricing architecture: Haiku for volume, Sonnet/Opus for complexity, with cache pricing as the knob that tunes the boundary between them. The API credit bundles for Max and Team subscribers — $100-$500/month depending on tier — are a classic platform play. Give developers free compute to build on your platform, and you create switching costs before they've written a check. This is Anthropic borrowing from AWS's playbook: subsidize experimentation, harvest lock-in. The adjustable effort setting, new for Haiku-class models, deserves attention. It lets users slide between cost and intelligence on a per-request basis, which means the effective price of any given task becomes a function of how much you're willing to spend on quality. This is granular price discrimination dressed up as user control — and it's smart. It lets Anthropic capture more of the demand curve without publishing a dozen SKUs. On safety, Anthropic reports improved alignment relative to Haiku 4.5 with lower willingness to cooperate with misuse, but applies somewhat less restrictive cybersecurity safeguards than Sonnet 5.5 — permitting a wider range of defensive tasks while still blocking offensive techniques. Biology safeguards remain consistent across the model family. The competitive implication is straightforward: Anthropic is trying to make Haiku 5.5 the default choice for any high-volume AI task where cost previously pushed developers toward open-source or self-hosted alternatives. At $0.10/M input tokens, they're pricing below many developers' cost of running open models on their own infrastructure. Whether this is generative competition or a margin-destroying land grab depends entirely on whether the resulting ecosystem creates net new value or just shifts workloads from one provider's balance sheet to another's.
