Claude Haiku 5.5 Launches at $0.10 Per Million Tokens: The Pricing Trap Behind the Price War

Anthropic has released Claude Haiku 5.5, the smallest, fastest, and cheapest model in the Claude family to date. Input costs just $0.10 per million tokens, the context window has been expanded to 1 million tokens, and it introduces an adaptive thinking mechanism for the first time. But behind the fanfare of this price war lies an easily overlooked pricing structure issue.

How Aggressive Is the Pricing?

Haiku 5.5's core pricing is $0.10 / $0.50 per million input/output tokens—provided your single prompt doesn't exceed 100K tokens. Once you cross that line, the price jumps directly to $0.50 / $2.50, a 5x increase on the input side.

Compare that with other models released around the same time: Meta's Muse Wonder is priced at $4 per million tokens, 40x Haiku 5.5's base tier. The two products launched just five days apart—a proximity that's hard to dismiss as coincidence.

Anthropic officially states that Haiku 5.5 is 75% cheaper than the previous-generation Haiku 4.5. The model ID is claude-haiku-5-5, and it's already available on the API, Bedrock, Google Cloud, and Copilot.

More Than Just Cheap: Three Substantive Upgrades

Adaptive Thinking

Haiku 5.5 is the first model in the Haiku series to support adjustable thinking effort. It automatically determines how much reasoning a task requires, and this is enabled by default. Developers can also manually trade off between intelligence, cost, and latency—a level of control granularity that hasn't been common on previous small models.

What Is Test-Time Compute

Adaptive thinking is a specific implementation within the broader category of test-time compute. Traditional models invest a fixed amount of computation per request, whereas test-time compute lets the model dynamically allocate a "thinking budget" before generating an answer—less computation for simple questions, more for complex ones. Common implementation approaches include chain-of-thought expansion, multi-path sampling with voting, and difficulty-based self-evaluation. The value of this mechanism lies in trading controllable additional compute for improved accuracy, at the cost of increased latency and token consumption. Haiku 5.5 brings this capability down to small models, meaning even the budget tier can automatically scale reasoning depth based on task difficulty rather than always taking the shortest path.

1 Million Token Context

A direct jump from Haiku 4.5's 200K tokens to 1 million, with the output limit also raised from 64K to 128K. Processing entire codebases, long documents, or large-scale conversation histories in a single pass becomes feasible in the budget tier.

Designed for High-Throughput Scenarios

Anthropic's positioning for Haiku 5.5 is clear: classification, summarization, information extraction, real-time customer service, voice agents, in-app assistants, and serving as a fast sub-agent model for Opus and Sonnet. What these scenarios share is that they don't require frontier reasoning capabilities, but are extremely sensitive to response speed and invocation cost.

What Is a Context Window

The context window refers to the total number of tokens a model can process simultaneously in a single invocation, covering both the input prompt and the output. It determines how much information the model can "see" at once—codebases, long documents, and multi-turn conversation histories are all constrained by this limit. The larger the window, the more it can support tasks like cross-file analysis, whole-book summarization, and long-term memory. But note that expanding the window doesn't proportionally improve effective utilization—models tend to lose attention on the middle portions of very long contexts (the "lost in the middle" phenomenon). At the same time, as this article reveals, large windows are often tied to higher billing tiers, and practical usability depends on whether the cost is acceptable.

The Pricing Trap Nobody Mentions

Here you need to look carefully at the pricing card:

• Under 100K tokens: $0.10 / $0.50
• Over 100K tokens: $0.50 / $2.50

This means the "$0.10 model" only holds when a single prompt doesn't exceed 100K tokens. When you actually use that 1 million context window—the biggest selling point at launch—the input cost effectively increases 5x.

There's a structural tension between the cheap unit price and the large context. This isn't concealment—tiered pricing is itself an industry convention—but every headline saying "Haiku 5.5 costs just $0.10" omits this precondition. Before undertaking long-context tasks, it's worth running the numbers first.

Also on the same day of launch, Anthropic halved the cache read price for Sonnet 5.5, reducing the overall cost of agent-based workflows by about 20%. If you're doing long-context-intensive tasks, comparing Haiku 5.5 against the price-reduced Sonnet 5.5 side by side may be more cost-effective than simply picking the cheaper one.

What Is Tiered Pricing

Tiered pricing is a billing method where service providers divide usage into different price brackets, with the unit price jumping once a certain threshold is crossed. Unlike linear billing (usage multiplied by a fixed unit price), marginal costs under tiered pricing change discontinuously with usage. In the LLM API space, common tiering dimensions include: context length per request, monthly cumulative usage, rate limit tiers, and more. For developers, the key is identifying where the "price cliff" is—once a single prompt crosses the threshold, unit costs can rise severalfold, fundamentally changing the cost model for long-context tasks. Understanding tiered structures is a prerequisite for capacity planning and model selection.

When to Use It and When Not To

Scenarios suited for Haiku 5.5:

• High-throughput classification, extraction, and summarization tasks
• Sub-agent invocations within large workflows (completing reasoning steps at low cost)
• Latency-sensitive voice agents and real-time customer service
• All scenarios currently using Haiku 4.5—strictly better and 75% cheaper

Scenarios not suited:

• Frontier reasoning tasks—that's Opus 5.5's territory, at 40x the price
• Ultra-long context tasks without running the numbers first
• Scenarios requiring the strongest safety guarantees—Haiku 5.5 has built-in blocking for certain high-risk cybersecurity requests

The Market Is Stratifying—Architecture Thinking Needs to Keep Up

Muse Wonder priced at $4/$20, Haiku 5.5 priced at $0.10/$0.50, released within five days of each other. This isn't just a difference in pricing strategy between two companies—it's the entire model market bifurcating into two tiers:

• High-end tier ($4+/million tokens): frontier reasoning, memory capabilities, long-duration coding tasks—Muse Wonder and Opus 5.5 occupy this range
• Throughput tier ($0.10/million tokens): fast, cheap, good enough—Haiku 5.5 is the representative of this tier

For developers, the sensible approach isn't choosing one or the other but doing routing and distribution: 90% of calls that don't need frontier intelligence go to Haiku 5.5, while the remaining 10% requiring frontier reasoning get routed to Opus or Muse Wonder. Anthropic positioning Haiku 5.5 as a sub-agent model is itself hinting at this architecture.

The key to cost optimization has shifted from "which model to choose" to "how to distribute requests."

What Is Model Routing

Model routing refers to setting up a dispatcher at the application layer that dynamically selects a backend model based on request characteristics (such as task type, complexity, latency requirements, and cost budget). A typical approach is to first use a lightweight classifier or rules to judge request difficulty, then send simple requests to a cheap model and complex requests to a high-end model—or even do cascading, trying a small model first and escalating to a large model when confidence is insufficient. The benefit of routing comes directly from the price gap: under the 40x gap described in this article, as long as 90% of traffic can be correctly directed to the cheap tier, overall costs can drop dramatically. The challenge lies in the routing judgment's own accuracy—misjudgments lead to quality degradation or extra overhead.

Key Takeaways

• Claude Haiku 5.5 is priced at $0.10/$0.50 per million tokens (under 100K tokens), with a 1 million context and adaptive thinking
• Above 100K tokens, the price increases 5x—calculate costs clearly before using the large context
• 75% cheaper than Haiku 4.5, now available on the API, Bedrock, Google Cloud, and Copilot
• The practical strategy is routing: throughput tasks go to Haiku, frontier tasks to Opus or Wonder
• Same-day Sonnet 5.5 cache read price halved, cutting agent workflow costs by about 20%

One-line Summary

Claude Haiku 5.5 enters the market at a low $0.10 price, but the 5x price jump above 100K tokens plants a long-context cost trap.

Paragraph Summary

Anthropic released Claude Haiku 5.5, entering the small model market with aggressive pricing of $0.10/$0.50 per million tokens, a 1 million token context, and an adaptive thinking mechanism, at 75% cheaper than Haiku 4.5. However, its pricing has a structural trap: the low $0.10 price only applies when a single prompt doesn't exceed 100K tokens, and once you use the full 1 million context—the core selling point—input costs increase 5x to $0.50. The article points out that the model market is stratifying into a high-end reasoning tier ($4+) and a throughput tier ($0.10), and the sensible strategy for developers isn't choosing one or the other but doing model routing—sending the vast majority of simple requests to Haiku while routing the few frontier reasoning tasks to Opus or Muse Wonder. On the same day, Sonnet 5.5's cache read price was halved, so long-context tasks require a comprehensive comparison before selection.