Unlabeled Claude Honeycomb Model Briefly Surfaces in Cursor, Sparking Opus 5 Speculation

On July 8, a developer posting under the handle chetaslua spotted an entry in the code editor Cursor's model list that hadn't been there before: Claude Honeycomb EAP. The model vanished from the selector within a few hours, but not before it had handled two test queries whose outputs made their way online.

Chetaslua's post on X quickly spread to Hacker News and the community of developers who use AI coding tools. Screenshots showed the entry described as an 'Anthropic research model with single-turn conversation controls and safety mechanisms,' labeled as an Early Access Preview.

What the specs revealed

What caught observers' attention was not so much the existence of an unknown model as its technical specifications. A million-token context window and an 'extra high effort' mode are elements Anthropic had already described as part of the Fable 5 architecture. But the key clue was the safety mechanism: Honeycomb routed queries flagged as sensitive to Claude Opus 4.8 instead of handling them itself.

In the industry, safety fallbacks typically route traffic to a weaker or cheaper model than the one the user is actually using. Since Honeycomb was forwarding queries to Opus 4.8, some developers concluded that the model itself must be more capable than Opus 4.8, which strengthened speculation about an early version of Opus 5 with a launch reportedly planned for later in July.

Anthropic stays quiet

Anthropic has neither confirmed nor denied Honeycomb's existence. The model does not appear in the company's public API or official documentation, and spokespeople did not respond to questions from The New Stack reporters about the nature of the leak. Cursor likewise did not comment on why the entry appeared in its model selector or how quickly it was removed.

This kind of silence is typical of AI labs testing unreleased models with integration partners ahead of an official launch, but it rarely surfaces publicly in such a tangible form, complete with screenshots and real model responses.

Fable 5 extended a third time

The leak coincided with another twist in the Fable 5 saga. On July 7, Anthropic extended free access to the model for paid plans just before it was set to expire, and on July 12 it did so again, the third extension in five weeks, this time through July 19 at 11:59 pm Pacific time. Subscribers on Pro, Max, Team, and select Enterprise plans can use Fable 5 until then within up to 50 percent of their weekly usage limit, plus a higher limit for Claude Code.

We're extending access to Claude Fable 5 across all paid plans through July 12 - Claude (@claudeai)

Once the promotion expires, users are expected to move to credit-based billing, paying $10 per million input tokens and $50 per million output tokens. The repeated deadline extensions suggest Anthropic still doesn't have enough server compute capacity to fully stabilize access to the model given current demand.

What it means for developers

For users of tools like Cursor, Claude Code, or GitHub Copilot, this episode shows how major AI companies test unreleased models directly on partners' production integrations, sometimes without full control over when and to whom the feature becomes visible. The brief window in which Honeycomb was available was enough for the community to reconstruct some of its characteristics from just two queries.

If the Opus 5 speculation proves true, it would be another launch in what's already a packed year for Anthropic, which has already rolled out Fable 5 and Mythos 5 this year while grappling with US export restrictions and rising infrastructure costs under competitive pressure from OpenAI and Google.

Sources: The New Stack (thenewstack.io), Hacker News (news.ycombinator.com), Tech Times (techtimes.com), SQ Magazine (sqmagazine.co.uk)