Claude Opus 5 Next Week? The Honeycomb Leak, Explained

On July 8, 2026, an unannounced Anthropic model called Claude Honeycomb EAP appeared in Cursor's model picker and vanished within hours. It is the first artifact in this entire cycle worth taking seriously: a screenshot of a shipping product's UI, not a screenshot of a tweet. There is still no confirmed Claude Opus 5 release date, and Anthropic has said nothing on the record, but the evidence base has changed shape.

Last verified: July 15, 2026. This is a developing story built on unverified reports. We will update this page when Anthropic makes an official announcement, or when any rumor here is confirmed or disproven.

The Honeycomb leak: what actually surfaced

On July 8, 2026, developer @chetaslua posted screenshots of a model nobody had seen before. It was not in a config file or a leaked deck. It was sitting in Cursor's production model picker, rendered, selectable, with real metadata attached. It was called Claude Honeycomb EAP. Within hours it was gone.

The listing text was specific enough to be useful: "Anthropic research model with per-turn controls and safety fallbacks. Early access preview." Next to it, two attributes: "1M context window" and "Version: extra high effort". The picker rendered the entry as "Honeycomb EAP 1M Extra High", with a "Long-running Preview" toggle and an "Upgrade for access" link beside it.

The company it kept matters as much as the listing itself. Honeycomb EAP appeared below Cursor Grok 4.5 High Fast, Composer 2.5 Fast, Opus 4.8 1M High, GPT-5.5 1M Medium, Fable 5 1M High, Sonnet 5 1M High, Sonnet 4.6 1M Medium, and Codex 5.3 High Fast. That is the ordinary roster any Cursor user sees, with exactly one entry in it that does not exist publicly.

The sighting was picked up by The New Stack, which framed it as Anthropic declining to talk about what developers found inside Cursor, and by TechTimes. There is also a Hacker News thread on the model being spotted and then pulled. Anthropic has not confirmed it, denied it, or commented on it. The Honeycomb model string does not appear anywhere in Anthropic's public API or documentation.

That combination is what makes this different from the rest of this cycle. The claim is checkable, it was surfaced by a named developer, it lives on a specific UI surface that thousands of people use, and it was seen by more than one person before it disappeared.

Why Honeycomb is probably Opus 5

The reasoning below is largely @pankajkumar_dev's reading of the screenshots, and it is worth engaging with because it argues from the artifact rather than from vibes. It is still a theory.

The 1M context window

Anthropic has an unreleased Haiku in the pipeline that is broadly expected at 300K context. Honeycomb listed at 1M. If you assume the context tier is not decorative, that pushes Honeycomb out of the small-model slot and into the frontier slot, where Opus 4.8, Fable 5 and Sonnet 5 already sit at 1M. This is the load-bearing part of the argument.

The fallback chain

The listing mentions "safety fallbacks", and @pankajkumar_dev's reading is that the chain steps down to Opus 4.8. Fallbacks conventionally step down in capability, not up. If that reading is right, Honeycomb sits above Claude Opus 4.8, which is currently the flagship. There is no tier above the flagship that is not a new flagship.

"Extra high effort" and the long-running toggle

"Version: extra high effort" reads as a reasoning tier, sitting a step above the "High" suffix on Opus 4.8 and Fable 5 in the same picker. The "Long-running Preview" toggle is the more interesting tell: it points at long-horizon agent work, which is the direction the whole category is moving and the thing a 1M window plus per-turn controls would exist to serve.

Put together, the picture is a frontier-tier, long-context, high-effort Anthropic model in early access with a named partner. That is what a pre-launch flagship looks like from the outside.

What it is not: confirmation. Anthropic has not confirmed the name, the tier, or that Honeycomb is Opus 5 at all. Research models get codenames that never survive to launch. Honeycomb could ship as Opus 5, ship under a different name, ship as something that is not a flagship, or never ship. All of those have precedent.

So is Opus 5 landing next week?

Possibly. Nobody has actually said so. That is the honest answer, and the distinction is the whole point of this section.

The end-of-July target that is circulating is @pankajkumar_dev's inference from the leak, not a statement from Anthropic and not something the Cursor listing said. Next week falls inside end of July, so next week is consistent with the leak. It is not what the leak claims, and no source has claimed it independently.

Cadence supports near-term. Anthropic shipped Opus 4.8 on May 28, Fable 5 on June 9, and Sonnet 5 on June 30: three releases in roughly five weeks. A company moving at that rate with a frontier model already in partner early access is not a company that is many months out.

The caveat that belongs right here: early access models appearing in Cursor have preceded public launches before, and they have also been pulled and never shipped. An EAP listing proves the model exists and is being tested with a partner. It does not prove a launch, and it certainly does not prove a date. Evaluation happens on models that get shelved too.

Land it plainly: the odds went up, the date did not get confirmed. If you needed to bet, near-term is now the better bet than it was on July 7. If you need to plan, you still have nothing to plan against.

What Anthropic has actually confirmed

The verified record is short, and Honeycomb is not on it.

• Claude Opus 4.8 launched May 28, 2026 and is the current flagship.
• Claude Fable 5 launched June 9, 2026.
• Claude Sonnet 5 launched June 30, 2026.
• Anthropic's blog stated Fable 5 is included for up to 50% of weekly usage limits, "after which it will be available via usage credits."
• Anthropic described Fable 5 demand as "very high, and difficult to predict."
• The Fable 5 subscription deadline moved from July 7, to July 12, and now to July 19, 2026 at 11:59:59 PM PT.

Worth disambiguating, because the naming confuses this whole conversation: there is no Claude 5 release date either. "Claude 5" is how people informally refer to the next Opus, but Anthropic has never used that name. The Claude 5 generation that actually shipped is Fable 5 and Sonnet 5. An unannounced Opus is not "Claude 5", and searching for one will mostly return speculation about the other.

There is no Opus 5 announcement. Not on the models page, not in the API, not in a press release. Honeycomb has no official existence at all: it is a thing developers saw, which is a different category from a thing Anthropic said.

The other release-date rumors

Next to a real screenshot of a real UI, the X chatter looks thinner than it did a week ago. It is still worth cataloguing, because a lot of it is being recycled as if it corroborates Honeycomb, and it does not.

Note in particular that @nusagain and @pankajkumar_dev posted near-identical wording. That is one claim propagating through two accounts, not two independent confirmations. This is the single most common way a rumor gets promoted to a fact in this cycle.

The Vertex AI sighting

@marmaduke091 posted that "claude-opus-5 is on vertex, coming very soon". A week ago that was the closest thing to hard evidence in the rumor set. Honeycomb now occupies that position, and the Vertex claim looks weak by comparison.

It is still uncorroborated. No screenshot has been produced, no other developer has reported the same listing, and the latest Opus on Vertex remains 4.8. Between two artifact-class claims, one has a screenshot, a UI surface others could check, and press coverage. The other has a sentence.

The mechanism is real, though, and that is why the claim is not zero. Both claude-sonnet-5@20260203 and anthropic-claude-opus-4-7 surfaced on Vertex before launch, the latter on a quota page on April 16, 2026. Cloud partner listings genuinely do leak Anthropic models early. That makes Vertex a good place to watch. It does not make an unsourced report of a Vertex listing into evidence.

The rumored pricing

Four accounts, four incompatible numbers. @ChrissGPT says much cheaper than Fable 5. @Conor_D_Dart expects pricing unchanged from Opus 4.8. @bindureddy floats a worst case of 2x Opus 4.8. @AlexFinn frames Opus 5 as the "affordable version of Fable".

They cannot all be right, and none of them cites a source. The spread runs from "much cheaper than the current premium tier" to "double the current flagship", which is another way of saying the range is uninformative.

Honeycomb does not help here. The Cursor listing showed context, effort tier, and safety behavior. It revealed nothing about price. An "Upgrade for access" link tells you the model was gated, not what it will cost.

If your interest in Opus 5 pricing is really an interest in your bill, the lever you control is token spend on the models you already have. Our notes on OpenClaw cost control with Anthropic cover that, and none of it depends on an unannounced model.

The rumored performance claims

@ChrissGPT claims Opus 5 beats GPT-5.6 Sol and Terra on SWE Pro and comes close to Fable on 3D and UI work. @borisplavljanic says it is close to Sol on both performance and token usage. @Conor_D_Dart expects it slightly behind Fable.

These are benchmark placements for a model nobody outside the early access program can run. They are unfalsifiable by construction: you cannot check them, and when the model ships they will be quietly replaced by whatever the real numbers are. Treat pre-release benchmark claims as entertainment.

Sam Altman's public claim that GPT-5.6 Sol matches Fable 5 at roughly a quarter of the cost belongs in the same bucket, with an extra discount applied: it is an interested party marketing his own product. If you want to test that yourself rather than take it on faith, you can run GPT-5.6 in Claude Code via Claude Code Router and compare on your own workload. For a broader cross-model view, our Grok 4.5 vs Claude vs GPT coding comparison uses models you can actually run today.

Honeycomb's listing does corroborate one thing, narrowly: capability tier. A fallback chain stepping down to Opus 4.8 implies Honeycomb sits above Opus 4.8. That is a claim about ordering, not about SWE Pro scores. It says nothing about any specific benchmark against any specific competitor.

What happens to Fable 5 on July 19

The rumor, in its strongest form, is that Fable 5 is being removed from subscriptions and made API-only on July 19, because Opus 5 is replacing it. @LoopOnChain and @borisplavljanic both pushed versions of this. It is contradicted on four counts.

First, Anthropic's own wording. Fable 5 is included for up to 50% of weekly usage limits, "after which it will be available via usage credits." Available via usage credits is not removed, and it is not API-only. It is a billing change.

Second, the Claude Code lead engineer said on X: "While it will come off subscriptions after July 7th, we aim to restore Fable as a standard part of our subscriptions as soon as capacity allows." That is the opposite of a permanent removal. BleepingComputer reported the correction directly.

Third, the stated driver is capacity, not a launch. Anthropic called Fable 5 demand "very high, and difficult to predict." That is a serving constraint. It has no dependency on whether another model ships.

Fourth, the deadline has slipped twice: July 7, then July 12, now July 19. A date tied to a hard product launch does not wobble like that. A date tied to capacity headroom does. This is also not new behavior: the pull-and-restore pattern is documented, and we wrote it up when Fable 5 came back after being pulled the first time.

@jonnygravity asked the reasonable question underneath all this: where does Opus 5 fit if Fable is the premium tier? Nobody has answered it, including Anthropic, and Honeycomb does not answer it either. Honeycomb's existence does not rescue the rumor. A capacity-driven billing change and a model launch remain two different things that happen to be near each other on a calendar.

How to read AI model leaks

Honeycomb is a useful worked example, because it is the first thing in this cycle that clears the bar. What makes it a signal: a named developer attached his identity to it, there is a screenshot, it lives on a specific UI surface (Cursor's picker) that other people can and did check, the listing text is oddly specific rather than round, and multiple outlets independently reported it. You can be wrong about what it means and still be right that it happened.

Compare that to the noise patterns in the table above. Confident tone with no artifact. Round numbers with no source ("2x the price", "10 days"). Benchmark placements for a model nobody can run, which cannot be checked and therefore cannot be wrong. Two accounts posting identical wording, then being cited as two confirmations.

Claimed NDAs are their own category. @Josh67Joshua said he signed an NDA "with claude" and will post his thoughts six hours after release. The claim refutes itself: announcing that you signed an NDA is a thing people under NDA do not do, and the promised payload arrives only after the information stops being worth anything.

@jimmycrossusa's "temu fable" snark is not analysis, but it is a useful skeptic signal, and the ratio of that kind of reply to breathless agreement is a decent live read on how much a given thread deserves.

What to expect

Honeycomb is the first thing this cycle that is actually checkable, and it points to something real: a capable, long-context, high-effort Anthropic model that is far enough along to be in partner early access, sitting above Opus 4.8 in whatever internal ordering produced that fallback chain. Combined with three releases in about five weeks, near-term is a reasonable expectation.

Be precise about what that expectation is worth. There is no date. There is no price. There is no confirmed name. There is no announcement. "Probably soon" is not a plan, and the gap between "the odds went up" and "it ships next week" is exactly where every bad take in this cycle lives.

The falsifiable thing to watch is narrow: Anthropic's docs, the changelog, a model string appearing in the public API, or a cloud partner listing on Vertex or Bedrock. Those either exist or they do not. Anthropic's news page is the primary source, and an announcement there will settle in one paragraph everything the last three weeks of X threads could not.

Practically, for this week: keep using Fable 5 through July 19, and understand that after that it bills via usage credits rather than disappearing. Budget for that, not for a removal. Do not restructure your stack, your evals, or your spend around a model that has no name, no price, and no ship date. If Honeycomb turns out to be Opus 5 and lands at the end of the month, adopting it a week late will cost you nothing. Rebuilding around a model that gets pulled will cost you a lot.