Every model router I've seen in the last two years is built on the same assumption. Small models are cheap and a bit dim. Big models are smart and expensive. You send the easy stuff down one pipe and the hard stuff down the other, and you spend a lot of engineering time deciding which is which.
Claude Haiku 5.5 just made that line a lot blurrier.
On the latest Artificial Analysis Intelligence Index, Anthropic's smallest new model scores 43. Claude Opus 4.8 scores 42. Kimi K3, the model a lot of teams moved to precisely because it was cheaper than the Western flagships, scores 44. Same tier, within a point of each other.
The bills are not in the same tier at all.
The one-page version: view the infographic
The numbers
Artificial Analysis also publishes what it costs to run each model through its index tasks. That's a better comparison than list price per token, because it bakes in how many tokens a model actually burns to get to an answer. Reasoning-heavy models can look cheap per token and still cost a lot per task.
Here's how the three line up. The cost figures are my readings off AA's log-scale chart, so treat them as approximate.
Model AA Index Cost per task vs Haiku 5.5
Kimi K3 (max) 44 ~$2.00 ~9x more
Claude Haiku 5.5 43 ~$0.22 baseline
Claude Opus 4.8 42 ~$4.10 ~18x moreAnd here's the raw API view from OpenRouter's side-by-side comparison:
Haiku 5.5 Kimi K3
Input / 1M tokens $0.10 $0.69
Output / 1M tokens $0.50 $13.00
Cached input / 1M $0.01 $0.45
Throughput (p50) 141 tok/s 49 tok/s
Latency (p50) 1.51 s 1.66 s
Context window 1M tokens 1.05M tokens
Max output tokens 128K 1.05MOutput tokens are where agent and reasoning workloads spend most of their money, and that's where the gap is widest: $0.50 against $13, roughly 26x. Haiku is also close to three times faster on throughput, which matters as much as price once a user is waiting on the other end.
For a sense of how far the floor has moved: Haiku 4.5 lists at $1 in and $5 out per million tokens on Anthropic's own pricing. The OpenRouter listing for 5.5 is a tenth of that on both sides.
What this changes
Your routing thresholds are probably wrong now. If you tuned a router last year, the cutoff for "this needs the big model" was set against a much weaker small tier. Some of what you're sending to a flagship today would now get the same answer from Haiku at a fraction of the cost. The only way to know is to replay real traffic through it and compare.
Volume workloads get a new default. Extraction, classification, summarisation, RAG answer synthesis, document triage, support drafting. These are the jobs that run millions of times a month, where a 10x price drop is the difference between a feature that pays for itself and one that finance keeps asking about.
Speed becomes a feature, not a tax. At ~141 tok/s, there's room to add a verification pass or a second sample and still come in under the latency of a single slower call.
What it doesn't tell you
I want to be careful here, because "Opus-level for 5% of the cost" is the kind of line that gets screenshotted without the footnotes.
There's no coding or agentic data yet. OpenRouter's benchmark panel shows AA coding and agentic scores for Kimi K3 (76 and 50), and none for Haiku 5.5. Vals AI has no scores for it either. K3 is strong exactly where Haiku is still unmeasured. If your workload is multi-step tool use or real code changes, the composite index isn't enough to go on.
It's brand new. OpenRouter's activity chart shows about 62B tokens through Haiku 5.5 in the last 30 days, almost all of it from the last two days, against 6.6 trillion for K3. Production behaviour at scale, rate limits and provider stability are still unknowns.
It isn't the cheapest model at this level. On AA's cost-versus-intelligence chart, Xiaomi's MiMo-V2.6-Pro scores 46 at a lower cost per task. Haiku 5.5 isn't on the Pareto line. What it is: the cheapest way to get this tier of intelligence from a major Western lab, with Anthropic's tooling, prompt caching and the provider spread that comes with it. For a lot of enterprise buyers, that's the comparison that actually matters.
Your prices may differ. The OpenRouter figures above are for one provider route on one day. Check the current rate card for the provider you'll actually use.
How I'd test it
- Pull a week of real traffic from whichever model you use today, with outputs.
- Replay it through Haiku 5.5 with the same prompts.
- Score both on the thing your users care about, not a generic benchmark. Exact-match for extraction, a rubric for drafting, task success for anything agentic.
- Compare cost per successful task, not cost per token.
- Move workloads over one at a time, starting with the highest-volume, lowest-risk one.
If Haiku holds up on your data, the savings are large enough to change what you can afford to build. If it doesn't, you've lost a day of compute.
That's the easiest A/B test you'll run this quarter.
Sources: Artificial Analysis Intelligence Index v4.3.2 and its Intelligence Index vs Cost per Task chart (7 Oct 2026), OpenRouter's Claude Haiku 5.5 vs Kimi K3 comparison (8 Oct 2026), and Anthropic's API pricing for Haiku 4.5. Related: Adding agents to software you already run.