Claude Haiku 5.5 Pricing: How Much Cheaper Is It Than Haiku 4.5?
Claude Haiku 5.5 costs 10 cents per million input tokens, about 75% less than Haiku 4.5. Here's the full pricing breakdown versus Sonnet 5.5.
Claude Haiku 5.5 costs 10 cents per million input tokens and 50 cents per million output tokens for prompts up to 100k tokens, which Anthropic says is roughly 75% cheaper than Haiku 4.5’s pricing. That makes it Anthropic’s cheapest model to date, positioned as a daily-driver option for developers who need Claude-level reasoning without Sonnet or Opus-level bills. For comparison, Sonnet 5.5 charges about 20 times more per input token, according to Anthropic’s published rates.
TL;DR
- Haiku 5.5 input pricing sits at 10 cents per million tokens for prompts under 100k tokens, down from Haiku 4.5’s $15 per million, a roughly 75% cut.
- Output pricing lands at 50 cents per million tokens, which keeps the input-to-output ratio similar to previous Claude models even as the base price drops.
- Prompt length matters: the discounted rate applies to prompts up to 100k tokens, and Anthropic notes that around 90% of real-world requests to the previous Haiku fell under that threshold.
- Sonnet 5.5 still costs roughly 20 times more per input token than Haiku 5.5, which makes Haiku the obvious fallback when Sonnet usage hits rate limits.
- Cache reads are cheaper too, which matters most for coding agents that repeatedly reread the same context window during multi-step tasks.
- Benchmark performance tracks the price drop in the right direction: Haiku 5.5 beats Haiku 4.5 on every measured category and even edges out some competing models in categories like computer use, though it still trails Sonnet 5.5 on agentic coding.
- Real-world testing showed the model catching a genuine logic bug in a multi-tier application stack, suggesting the price cut hasn’t gutted practical usefulness for everyday debugging and coding work.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How much does Claude Haiku 5.5 actually cost?
Anthropic lists Haiku 5.5 at 10 cents per million input tokens and 50 cents per million output tokens, for prompts up to 100k tokens. That’s the headline number, and it’s what makes this release notable: Haiku 4.5 charged $15 per million tokens under the same conditions, so the new model represents roughly a 75% reduction according to Anthropic’s own figures.
Context length matters here. The discounted rate applies specifically to prompts under 100k tokens. Anthropic has said that band covers about 90% of the requests that previously went to Haiku 4.5, so most real usage, especially chat interfaces, coding assistants, and short-context tool calls, falls squarely inside the cheap tier.
Cache pricing has also come down. Cache reads are cheaper than fresh input tokens across Claude’s lineup, and that discount carries extra weight for agentic coding tools that keep rereading the same files, system prompts, or project context on every turn. If your workflow involves an agent looping over a codebase or a long-running task with a static context window, cache pricing can end up mattering more than the headline input rate.
How does Haiku 5.5 pricing compare to Sonnet 5.5?
Sonnet 5.5 charges around 20 times more per input token than Haiku 5.5. That gap is wide enough to change how teams should think about model selection, not just for cost reasons but for availability. Anthropic’s higher-tier models, Sonnet and Opus, carry tighter rate limits, and heavy users frequently hit throttling within a given usage window.
Swapping appropriate workloads over to Haiku 5.5 doesn’t just save money, it also stretches the usable budget within Anthropic’s rate-limit windows. If you’re running a mixed workload where some tasks need Sonnet-level reasoning and others don’t, routing the simpler, shorter tasks to Haiku 5.5 frees up Sonnet quota for the work that actually needs it.
This tiered approach isn’t unique to Anthropic. Most frontier labs now ship a small/cheap model alongside their flagship, explicitly so developers can route traffic by task difficulty rather than defaulting everything to the most expensive option. What’s notable about this release is the size of the price cut relative to the previous Haiku generation, not just the existence of a cheap tier.
Is Claude Haiku 5.5 worth it compared to Haiku 4.5?
On price alone, yes, the math is straightforward: a 75% reduction in input cost with output pricing following a similar pattern. The harder question is whether performance held up.
Based on hands-on testing covered in a recent review, Haiku 5.5 scored above Haiku 4.5 across every benchmark category shown, and in several categories it outperformed competing models as well. The most dramatic jump was in computer use tasks, where the model reportedly went from completing a small fraction of tasks to around three-quarters of them. That’s a meaningful jump for a model whose main selling point is being cheap and fast rather than maximally capable.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
In a practical test, the model was pointed at a real three-tier application (Postgres backend, Flask API layer, Nginx frontend, containerized with Docker and Redis caching) and asked to find a bug without being told what it was. It correctly identified that a dashboard was averaging milk production across all animals in a herd, including non-producing males and calves, which silently understated the productivity of actual milk-producing animals by nearly half. The model found the bug, fixed it, and also caught a connection leak and a divide-by-zero error along the way. That’s a solid result for a model priced at a fraction of its predecessor’s cost.
Where does Haiku 5.5 fall short?
The same testing found real limits. When asked to build a full 3D game environment from scratch, including modeling assets and assembling a playable build, Haiku 5.5 produced something that ran but didn’t fully deliver on the brief. Core interactions like trampoline physics didn’t register correctly, even though other elements (jump animations, a bonus “ninja warrior” course) worked and exceeded what was asked. The task reportedly cost around $2 to complete over roughly 25 minutes, which is cheap for an agentic build of that scope, but the output quality showed the gap between Haiku and larger models on long, multi-step creative and coding tasks.
Benchmark data backs this up directly: Sonnet 5.5 still leads Haiku 5.5 across every category tested, and the gap is narrowest on knowledge work, reasoning, and computer use, but widest on agentic coding, where Sonnet’s performance is described as nearly double Haiku’s. The takeaway is consistent with how Anthropic appears to be positioning the model: fast everyday work and shorter tasks, not extended autonomous coding sessions.
Multilingual and nuanced reasoning tests painted a similar middle-of-the-road picture. The model handled instruction-following across multiple languages correctly but produced fairly flat, literal output without much depth. On a couple of visual reasoning tasks involving ambiguous images, it got one right (correctly identifying which of two trucks was braking based on liquid movement inside, citing internal baffles as the reasoning) but hedged on another where a clearer read was arguably available.
Should you switch to Claude Haiku 5.5?
If your usage currently sits on Haiku 4.5 for cost reasons, there’s little reason not to move to 5.5 given the price drop and the improved benchmark scores. If you’re currently on Sonnet 5.5 and running into rate limits, testing Haiku 5.5 for your lighter-weight tasks is a reasonable way to extend your effective usage window without paying for capability you don’t need on every call. The clearest limitation is long, autonomous coding work, where the benchmark gap to Sonnet remains widest and the real-world game-building test showed unresolved issues even after a multi-step agentic run.
Frequently Asked Questions
How much does Claude Haiku 5.5 cost per million tokens?
Input tokens cost 10 cents per million, and output tokens cost 50 cents per million, for prompts up to 100k tokens. This applies to the vast majority of real-world requests, since Anthropic reports about 90% of previous Haiku traffic used prompts under that length.
Is Claude Haiku 5.5 cheaper than Haiku 4.5?
Yes. Anthropic says Haiku 5.5 costs roughly 75% less to run than Haiku 4.5, whose input price was $15 per million tokens compared to Haiku 5.5’s 10 cents.
How does Haiku 5.5 compare to Sonnet 5.5 in price?
Sonnet 5.5 costs about 20 times more per input token than Haiku 5.5. That makes Haiku the more practical option for high-volume or latency-sensitive workloads where Sonnet-level reasoning isn’t required.
Is Haiku 5.5 good enough for coding tasks?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
It performs well on bug-finding and shorter coding tasks, as shown by a test where it correctly diagnosed a data-averaging bug in a multi-tier application. It performs less reliably on long, autonomous coding projects, where benchmark results show Sonnet 5.5 holding a significant lead.
Does Haiku 5.5 support cheaper caching?
Yes, cache read pricing has also been reduced, which benefits coding agents and tools that repeatedly reread the same context across multiple steps in a task.
