Claude Haiku 5.5: Anthropic Says the Cheapest Tier Is Coming, Weeks After a Leak Said It Was Dead

For about two weeks in September, the working assumption among people who track the vendor's roadmap was that the cheapest Claude tier was being retired. The leak — a post on X from an account known for the vendor's roadmap claims — said the 5.5 generation would restructure the lineup: Claude Fable, Claude Opus and Claude Sonnet would all get 5.5 versions, Fable would move to a new pre-training base, and the Haiku line would exit. The company never confirmed any of it. Then on September 22, 2026, alongside the launch of Claude Opus 5.5, the company said the opposite in one sentence: "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks." Claude Haiku 5.5 is therefore announced, undated, and unlaunched — and it exists specifically because the rumor about it was wrong. The current shipping small model, Claude Haiku 4.5, is still what you would call today.

What follows is what is actually established, what is inference, and what nobody knows yet. The distinction matters more than usual here, because most of what is circulating about Haiku 5.5 is extrapolation from the tier above it.

What is actually confirmed, and by whom

The confirmed set is small, and it is worth keeping it small.

• The announcement — Anthropic's own launch post for Claude Opus 5.5 states: "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks." That is the whole of the direct evidence. It is a sentence in someone else's launch post, not a Haiku 5.5 announcement.

• The scope of the promise — the same post says the two models will carry "many of the same improvements to performance, efficiency, and safety" as Claude Opus 5. That is a category list, not a specification. No benchmark, no price, no context window, no model ID, no date.

• The tier position — Haiku is the fastest and cheapest of Anthropic's three current tiers, below Sonnet and Opus. That has been true since the line was introduced and is not in dispute.

• The timeline — "the coming weeks," which places the earliest plausible arrival in late September or October 2026 and says nothing about the latest.

Everything else in circulation is either the leaker's claims, which Anthropic has not endorsed and which the announcement contradicts on the central point, or inference drawn from Claude Opus 5.5. Treat both as unverified.

The rumor, specifically, and why it was plausible

The exit claim was not stupid. It had a real basis, and understanding it tells you something about how Anthropic's small tier has been treated.

Claude Haiku 4.5 shipped on October 15, 2025, at $1 per million input tokens and $5 per million output, with a 200K-token context window and a 64K output cap. It was the fastest model in the lineup and the cheapest by a wide margin, and it was positioned as matching an older Sonnet generation on coding and agent tasks at roughly a third of the cost.

Then the tier went quiet. Between Haiku 4.5 in October 2025 and the Opus 5.5 launch in September 2026 — eleven months — every other Claude tier moved at least twice. Opus went 4.5, 4.6, 4.7, 4.8, 5, 5.5. Sonnet went 4.5, 4.6, 5. Fable arrived as a new top tier entirely. Haiku did not move at all. In that same window Anthropic's retirement commitment for Haiku 4.5 lapsed to "not sooner than October 15, 2026" — a date that, as of this writing, is about three weeks away.

An eleven-month gap on a tier that used to refresh roughly every six months, plus a retirement clock running out, is exactly the shape a product line takes on its way out. That is what the leak read. It was a reasonable read. It was also wrong, and the correction came from the vendor rather than from a second leaker.

What Haiku 5.5 will almost certainly inherit from Claude Opus 5.5

This section is inference. It is labelled as such because Anthropic has confirmed none of it, and because the analogy to the tier above is the only source of signal available.

Anthropic's promise of "many of the same improvements" gives a starting list. Claude Opus 5.5 shipped with four properties a reader can check against the current small model:

• Price direction — Claude Opus 5.5 is $4 input / $20 output per million, down from Claude Opus 5's $5 / $25, with cache reads cut from $0.50 to $0.20. A 20% cut on the flagship makes a matching cut on the small tier plausible. It is not a commitment, and nothing in the announcement says it.

• Adaptive thinking — Claude Opus 5.5 does not accept thinking: {"type": "disabled"} or a fixed budget_tokens; both return a 400. Thinking is always on and effort is the control. Claude Haiku 4.5 today is the opposite: it uses the older manual {"type": "enabled", "budget_tokens": N} form and has no effort parameter at all.

• Context window — Claude Opus 5.5 keeps a 1M-token context with 128K max output. Claude Haiku 4.5 is 200K with 64K out. A jump to 1M is the single biggest thing that would change what the small tier is good for, and it is also the one nobody should assume.

• Safeguards — Claude Opus 5.5 can return a refusal stop reason across a wider category set, with biology requests routed to Claude Opus 5 and most cybersecurity requests to Claude Opus 4.8. Whether the small tier inherits the same classifiers, or a lighter set, is unknown.

If you are planning a migration on the assumption that Haiku 5.5 is a cheaper Claude Opus 5.5 with a smaller context window, you are planning on an analogy. The analogy is the best available, and it is not evidence.

Why the tier matters more than the flagship does

Flagship releases get the coverage and the small models get the traffic, which is a mismatch worth naming.

Claude Haiku 4.5 is what runs in the places where a model call happens hundreds of thousands of times a day and a per-token price is the entire product decision: classification, extraction, routing, summarisation, the cheap first pass in a cascade, the sub-agent that reads files so the expensive model does not have to. Anthropic's own framing for it is the fastest model with near-frontier intelligence, and the latency profile backs that up — independent measurement has it at roughly 86 output tokens per second with a 0.65-second time to first token, faster on both counts than the median non-reasoning model in its price band.

Against that, Haiku 4.5 is not a frontier model and does not pretend to be. Its independent Artificial Analysis Intelligence Index score is 15, ranked 28th of the 61 models on that board and flagged as estimated. That is the number that makes the 5.5 refresh interesting rather than routine: if the small tier picks up anything like the capability step Claude Opus 5.5 took, it moves the price-performance floor for high-volume work, and a lot of production pipelines are built on exactly that floor.

The practical question for a reader today is not whether to wait. It is what the wait costs. If you are building on Haiku 4.5, the eleven-month gap is already the answer to whether the tier is stable enough to depend on — the vendor has just told you it is. If you are holding a migration for Haiku 5.5, you are holding it for an undated model with no published specification, which is a worse position than shipping on what exists and moving when the new one lands.

Running the small tier you have, next to the ones above it

The reason a Haiku refresh is worth watching from a routing seat rather than a model-review seat is that the interesting decision is rarely which model to use — it is which model to use for which call. A cascade that sends the easy 80% to a small model and escalates the rest is the standard cost play, and it only works if both tiers are reachable from the same place with the same credentials.

Claude Haiku 4.5 is already on OrcaRouter at Anthropic's list price, and so are Claude Opus 5 and Claude Sonnet 5 — one API key across all three, with 0% markup, meaning the provider's list price is passed through rather than marked up. When a vendor cuts a price, that cut is live on our side the same day, which matters more on the tier you call a million times than on the one you call a hundred. For the cascade specifically, the routing DSL lets you compose the tiers into a single call so the escalation logic lives in the route rather than in your application, and automatic failover means a bad afternoon at one provider does not take the pipeline down with it.

Two honest caveats. Claude Opus 5.5 itself is not on OrcaRouter — the newest tier is not always the newest route, and we would rather say so than imply otherwise. And nothing about Haiku 5.5 can be routed, because it does not exist yet. The models that are callable today are the ones listed above.

What to watch for, in order

When Haiku 5.5 does arrive, four facts will settle almost every open question, and they will arrive in a predictable order.

• The model ID and the context window, from Anthropic's model overview page — a dateless ID in the claude-haiku-5-5 form would match the 4.6-generation-onward convention, and a 1M-token window would be the headline.

• The rate card. The tier's whole reason to exist is its price. A number above $1/$5 would be a repositioning; a number at or below it is a straight improvement.

• Whether thinking is adaptive-only. If Haiku 5.5 rejects budget_tokens the way Claude Opus 5.5 does, that is a breaking change for every integration that currently sets it, and it is the kind of thing that only shows up when you run it.

• Independent scores. The tier's vendor benchmarks have historically been generous relative to what independent boards measure, so the first Artificial Analysis placement is worth more than the launch table.

Until those exist, the accurate statement is the boring one: Anthropic has said Claude Haiku 5.5 is coming in the coming weeks, it has not said what it is, and the leak that said it was cancelled was wrong. That last part is the news. Everything else is a schedule.