The next tier of frontier models is leaking out sideways, before anyone will admit it exists.

In the span of a week, two unreleased flagships tipped their hand by accident. Anthropic's next Opus, hiding behind the codename Honeycomb, flashed up inside the Cursor coding editor and vanished within hours. Then a promotion page for Moonshot's Kimi K3 briefly went live on the company's own platform, pointing to a launch that is essentially now. Neither company has officially confirmed a thing.

They are the two poles of the same race: one closed, premium, and American, the other open, cheap, and Chinese, both chasing million-token context and long-horizon agents. Here is what leaked, what is real, and what is still just hype.

On July 8, developer @chetaslua caught a model labeled "Claude Honeycomb EAP" sitting in Cursor's model picker, grabbed screenshots, and watched it disappear within hours after just two prompts. The listing described it as an Anthropic research model in Early Access Preview, with per-turn controls, safety fallbacks, a 1-million-token context window, and an "extra high effort" reasoning mode.

Leak details via The New Stack and WinCentral

The detail that set the community off was the safety behavior: Honeycomb routes sensitive prompts down to Claude Opus 4.8 rather than handling them itself. That is exactly how Fable 5, Anthropic's Mythos-tier model, behaves, and if a model falls back to 4.8, most developers read that as a sign it sits above 4.8, not below. Combine that with a spec sheet that mirrors Fable 5's architecture, and a conspicuous gap in Anthropic's lineup, Sonnet 5 and Fable 5 have shipped while Opus is the one tier still on its old number, and the theory hardened fast: Honeycomb is an early preview of Claude Opus 5, possibly launching by the end of July.

Before you rearchitect anything, the skeptics have a point. There is no claude-opus-5 API ID, no system card, no pricing, and no benchmarks, and Anthropic has said nothing about the name, the specs, or the timeline. The cleanest guess from people tracking Anthropic's catalog is a model priced around $5 per million input tokens and $25 per million output, slotting between Sonnet 5 and Fable 5. Treat every number here as provisional until it shows up in Anthropic's own docs.

The timing is not random. OpenAI's GPT-5.6 Sol and xAI's Grok 4.5 both launched this month undercutting Fable 5 on price, and Opus is the one flagship tier Anthropic has not refreshed. An Opus 5 with a 1-million-token context would be the company's answer, and an included-in-subscription Opus 5 would be its strongest lever to keep developers from defecting over Fable's $10 and $50 per million pricing. For now, the smart move is to prep an eval and a model router so you can drop it in the day it is real, and keep shipping on Opus 4.8 in the meantime.

Anthropic was not the only lab to leak a flagship this week. A page titled "Kimi K3 launch limited-time recharge campaign" briefly went live on Moonshot AI's own Kimi platform, dated to start July 15 China time, before it was swapped out for a redirect. Screenshots spread on X, where developer @kimmonismus translated the promo, 10 to 30 percent bonus credits on top-ups running through August 11.

Leak details via Times of AI and Technosports

The specs circulating are bigger than a pricing bump. Community reports and insider briefs point to a roughly 2.5-trillion-parameter Mixture-of-Experts model with a 1-million-token context window, built on a new architecture aimed squarely at long-horizon agent work rather than a straight scale-up of the K2 series. Moonshot's K2 family already runs 12-hour autonomous coding sessions and 300-agent swarms as open weights at roughly a tenth of Claude's price, so a bigger K3 on the same open and cheap playbook is exactly what has US-lab watchers nervous. One widely shared post put it bluntly on X: the Sonnet tier was already dead, and K3 could make the Opus tier obsolete too.

The honest read: the July timing is a strong signal, the parameter and context figures are plausible but unconfirmed, and there is still no official model card, no weights, no API ID, and no independent benchmarks. The leaked launch window has arrived without an official release, and Moonshot's platform still lists K2.6 and K2.7 Code as its current models. Three things will turn this leak into fact: an official Moonshot post with a model card, a Hugging Face weights drop, and third-party SWE-bench numbers. Until then, K2.7 Code is the Kimi model you can actually use.

The stories moving fast beyond today's headlines:

Two unreleased frontier models leaked in one week, both by accident, and neither lab will confirm anything. Anthropic's next Opus surfaced inside Cursor on July 8 under the codename Honeycomb, an Early Access Preview model with a 1-million-token context, an extra-high-effort reasoning mode, and a safety fallback that routes sensitive prompts down to Opus 4.8, a detail that convinced developers it sits above 4.8 and is likely Claude Opus 5, possibly launching by month-end. There is still no API ID, system card, price, or benchmark, and the strongest guess is a $5 and $25 per million model between Sonnet 5 and Fable 5. Days later, a Kimi K3 promotion page briefly went live on Moonshot's own platform pointing to a July 15 China-time launch, with leaks describing a roughly 2.5-trillion-parameter, 1-million-token, new-architecture model aimed at long-horizon agents, on the same open-weight and low-price playbook that has US labs nervous. But that window has arrived with no model card, no weights, and no benchmarks, and Moonshot still lists K2.6 and K2.7 Code as current. The takeaway for builders is the same on both sides: prep an eval and a router now, keep shipping on Opus 4.8 and K2.7 Code, and treat every leaked spec as provisional until it lands in an official doc.

Stay building. 🤖