---
title: "Anthropic Ships Opus 5.5 and Announces Sonnet 5.5, Which Is Not the Same Thing"
description: "The mid-tier Claude is coming “in the coming weeks,” but claims about its speed and price are already on the shelf."
author: "Nate Ledger"
published: 2026-09-28T19:13:03Z
modified: 2026-09-29T03:27:00Z
url: https://rews.cc/a/anthropic-ships-opus-5-5-and-announces-sonnet-5-5-which-is-n-68adeb
language: en
tags: ["anthropic", "openai", "ai", "llm", "productivity", "tech"]
publisher: "Rews (https://rews.cc)"
---

# Anthropic Ships Opus 5.5 and Announces Sonnet 5.5, Which Is Not the Same Thing

*The mid-tier Claude is coming “in the coming weeks,” but claims about its speed and price are already on the shelf.*

By Nate Ledger · September 28, 2026 · https://rews.cc/a/anthropic-ships-opus-5-5-and-announces-sonnet-5-5-which-is-n-68adeb

## In brief

- Anthropic released Claude Opus 5.5 on Sept 22, 2026, and said Sonnet 5.5 and Haiku 5.5 would follow “in the coming weeks”
- Opus 5.5 is priced at $4 per million input tokens and $20 per million output, down from $5/$25 for Opus 5
- Anthropic says Opus 5.5 is over 30% faster than Opus 5 and roughly 40% cheaper on typical workloads
- Online claims about Sonnet 5.5 speed, token costs and safeguards have no official source; no model card or pricing exists
- Lower latency and token costs matter most for agentic systems that call a model dozens of times per task

Here is a sentence that is true: on September 22, 2026, Anthropic released Claude Opus 5.5, its new top-of-the-line model, confirmed the pricing, and said that the smaller Sonnet 5.5 and Haiku 5.5 would follow. Here is a sentence that is also true, and that does a surprising amount of work these days: a lot of people on the internet apparently read that announcement and came away believing Sonnet 5.5 exists, has a price, and posts benchmark numbers.

It does not. This is worth dwelling on, because the modern AI industry has developed a genuinely strange information structure. A company announces a product. The product is not available. Nonetheless, within days, the product acquires a full spec sheet online — speed figures, cost figures, safety features — assembled from who knows where and attributed to no one in particular. The announcement is real; the spreadsheet is fan fiction. And because the fan fiction looks exactly like the documentation that will eventually exist, you can only tell them apart by checking whether Anthropic actually published anything.

So let’s separate the two piles. Pile one, the confirmed facts, and they all concern Opus 5.5. It is live. It costs $4 per million input tokens and $20 per million output tokens, per Anthropic’s official announcement, down from $5 and $25 for Opus 5. Anthropic characterizes typical workloads as roughly 40% cheaper overall. It generates output more than 30% faster than Opus 5, an Opus-to-Opus comparison that says nothing about any other model. And it ships with safety measures previously reserved for Anthropic’s most capable systems, covering cybersecurity and biology-related prompts.

Pile two, the unconfirmed claims, all concern Sonnet 5.5. Anthropic’s September 22 statement said Sonnet 5.5 and Haiku 5.5 would arrive “in the coming weeks” — Reuters used the same forthcoming framing — and as of now there is no official release page, no model card, no API pricing document. Windows Forum coverage notes that unofficial reports have floated possible launch dates, all unconfirmed. Meanwhile, reports circulating online attribute to Sonnet 5.5 a 30% speed improvement over Sonnet 5, a lower token-burn rate, and cost-adjusted advantages over Opus 5.5 on agentic coding tasks. None of those numbers appear in Anthropic’s announcement or in Reuters’ coverage. Claims that Sonnet 5.5 will carry cybersecurity safeguards comparable to Opus 5 are similarly floating around without an identifiable primary source.

You can see the shape of the error, if error is even the right word. Anthropic confirmed a real number — more than 30% faster — for one model, and somehow that exact number got photocopied onto a different model that doesn’t exist yet. This is not a conspiracy; it’s just how attention works. The announcement happens, the details blur, and within a news cycle the blur has converged on something specific and wrong. Specificity, it turns out, is easy to fake. Sourcing is not.

Why does anyone care enough to pre-invent the specs? Because the mid-tier model is arguably the economically important one, and the reason is arithmetic. Anthropic has consistently positioned the Sonnet tier as the workhorse for coding, document creation, autonomous agents and general office tasks. Positioning is not a benchmark, but the logic is sound: imagine — purely as an illustration — a coding agent working through a multi-step software problem. It might call the model dozens of times before it’s done. Each call costs tokens and takes time, so a modest improvement in latency or token efficiency doesn’t just help once; it compounds across every call in the chain. Shave a little off each of thirty calls and you’ve shaved a lot off the task.

Which means that if Sonnet 5.5 genuinely is faster and thriftier than its predecessor, that’s a real event for anyone building on it. It also means the incentive to say so early — to be the first account with the numbers, real or otherwise — is strong and constant. Whether the model actually delivers depends on independent testing that, as of now, has not happened, because the thing being tested does not yet exist.

The backdrop makes this more than a footnote. Reuters reported continued competitive pressure among major AI developers after Anthropic’s Claude 5.5 family announcement, and the race for faster, cheaper models capable of running autonomous agents shows no sign of letting up. In a market moving this quickly, developers and buyers are making tooling decisions on real deadlines, and the gap between “announced” and “shipped” is where bad decisions get made. The tells are boring and reliable: an official Anthropic model page, confirmed API pricing, a published model card, and independent benchmarks on latency, token consumption and coding performance.

Until those exist, the correct posture toward Sonnet 5.5 specs is the one you’d take toward any product that isn’t on the shelf: interesting, not actionable. The impressive part, honestly, is that the product’s reputation got released ahead of schedule. Delivery of the model itself remains, as they say, in the coming weeks.