The AI world has a favorite pastime, and it is reading tea leaves in model pickers. This week the tea leaves say Claude Opus 5, so let me do what I always do with rumors: separate what is actually known from what people want to be true.

My rule on this site is no proof, no progress. So we will do the proof first.

What is actually confirmed: nothing

There is no Claude Opus 5. As I write this, Anthropic has not announced one. No pricing, no model card, no blog post. Every "Opus 5" headline you have seen traces back to secondhand sightings, not to Anthropic.

Here is what the sightings actually are. In July, a research-preview entry matching an unreleased Anthropic model briefly showed up in a coding tool's model picker, and a matching model string appeared in Google Vertex quota catalogs. Leak coverage attaches the codename "Honeycomb" to it, along with a rumored million-token context and finer per-conversation controls. A rumored July 20-21 launch window already came and went with nothing. And in the strangest data point of the season, Polymarket bettors put 88% odds on an Anthropic model debut landing today, July 23. People are literally betting money on a launch Anthropic never promised. I checked before publishing this: still nothing.

Catalog strings and picker entries are real signals, to be fair. Companies stage infrastructure before launches, and these are the same breadcrumbs that preceded other releases. But a breadcrumb is not a loaf. Vendors test things that never ship, rename things at the last minute, and stage entries weeks early.

What Anthropic has actually shipped this year

The confirmed 2026 timeline explains why everyone expects something, so here it is.

The newest Opus is Opus 4.8, released May 28. Then on June 9, Anthropic went over the top of its own lineup with a new tier entirely: the Mythos class, launching Claude Fable 5 and announcing Claude Mythos 5, with a million-token context and double Opus pricing. Three days later a US export directive forced Fable 5 offline after researchers reported a safeguard bypass, and Anthropic brought it back July 1 with new safety classifiers. Mythos 5 remains restricted to select organizations.

Squint at that timeline and the rumor logic writes itself. The Mythos class sits above Opus as the restricted, expensive frontier tier. The Opus line is the workhorse tier normal companies actually build on, and it has been sitting at 4.8 while OpenAI shipped the whole GPT-5.6 family and xAI shipped Grok 4.5. There is an obvious hole in Anthropic's lineup shaped exactly like an Opus 5. That is why the rumor will not die. It makes too much sense.

Why it matters if it is real

If an Opus 5 ships, the interesting question is not the benchmark bar chart. It is where Anthropic prices the workhorse.

Right now the market is splitting into two lanes. The frontier lane, Fable 5 and Sol, is where the crowns live and where the tolls are steep. The volume lane is where the actual agent economy runs, and it is getting cheap fast. I wrote yesterday about running my own agents on Grok 4.5 and GPT-5.6 Luna for exactly that reason. An Opus 5 that lands frontier-smart at workhorse prices would put real pressure on both lanes at once. An Opus 5 priced like a trophy changes nothing for people like me.

There is also the backdrop nobody at Anthropic is enjoying: the White House publicly accused Moonshot AI this week of building Kimi K3 by distilling Anthropic's Fable model. Whatever Anthropic launches next, it launches into a world where its own outputs are the thing competitors get accused of stealing, and where its last big release got yanked by the government inside of a week. If the Opus 5 rollout feels slow and careful, that recent history is probably why.

My honest read

I think the breadcrumbs are real and something is coming. I also thought that about three other rumored launches over the years that never happened, so that and five dollars gets you a coffee.

Treat every Opus 5 leak as fiction until it shows up in Anthropic's own docs. That is not me being a grump, that is just the standard: model strings are not products, prediction markets are not press releases, and the only launch date that counts is the one the company publishes. When it is real, you will not need a rumor site to tell you. Until then, keep building on what actually exists today, because what exists today is already pretty great.