Claude Opus 5.5 is here — what it actually means for your bill and your workflow

If you use Claude — through the app, through Claude Code, or through an API key in something you've built — the model most people call 'Claude' just quietly got better and cheaper on the same day. Anthropic released Claude Opus 5.5 on 22 September 2026, and if you were already on Opus, you're on it now without doing anything. This is what actually changes, in plain terms.

Anthropic sells three main tiers, in order of price and (usually) capability: Haiku, the fast cheap one; Sonnet, the middle ground; and Opus, the flagship. Above Opus sits Fable, a separate, considerably more expensive tier for the hardest work. The useful news in this release is that Anthropic itself is saying the gap between Opus and Fable has mostly closed for everyday tasks.

Two different things are changing at once, and it's worth separating them. First, the sticker price per million tokens dropped: input tokens (what you send Claude) fell from $5 to $4, output tokens (what Claude sends back) fell from $25 to $20 — a straightforward 20% cut. Second, Anthropic says the model also does the same work using fewer tokens overall, because it's more efficient at reaching the answer. Multiply a smaller per-token price by fewer tokens needed, and you get Anthropic's headline claim of 40% lower cost on typical workloads. The first part you can verify yourself by checking your own invoice. The second part is Anthropic's own estimate — worth testing on your actual use case rather than assuming it applies uniformly.

One number that matters more than it sounds: 'cache reads' — the cost of Claude re-reading context it's already seen, which happens constantly in any long back-and-forth or coding session — fell from $0.50 to $0.20 per million tokens, a 60% cut. If you use Claude for anything that runs over many turns (a long coding session, an agent working through a task, a long chat), this is the number that will move your actual bill more than the headline input/output prices.

Both, on Anthropic's own testing. It reports Opus 5.5 doing noticeably better than Opus 5 on coding and technical tasks, giving one example of a very large code migration finished fast.

It also reports the model writing more plainly: putting the key point first instead of burying it in paragraphs, and using less jargon, which several of Anthropic's business customers specifically praised. That's a real, noticeable difference if you've found Claude's answers wordy in the past — though it's also the kind of subjective improvement you'll only really judge by using it yourself for a week.

Anthropic says the smaller Sonnet 5.5 and Haiku 5.5 models are coming 'in the coming weeks' with similar efficiency gains — worth waiting for if you're on one of those tiers for cost reasons rather than because you need Opus-level capability. This launch also didn't happen in isolation: about 90 minutes later, OpenAI released updated GPT-6 Sol and Luna models at roughly half the price of its previous generation, and xAI had released its own Grok 4.7 the day before. If you build on more than one provider's API, it's a genuinely good week to re-run your cost comparisons — three labs moved on price within 48 hours, and the numbers you had saved from last month are probably out of date.

The practical takeaway, stripped of the launch-day noise: if you already use Claude Opus, you're now on a better, cheaper model automatically, with nothing to configure. If you were paying extra for Fable, it's worth re-testing whether you still need to. And if cost is the main thing you care about, watch for Sonnet 5.5 and Haiku 5.5 — the same efficiency gains, at tiers that were already cheaper to begin with.