Marshmallow on the Grill: Claude's Leaked Models Push Opus-by-Sept-30 to 57¢
Leaked Claude models Marshmallow and Melon pushed the Opus-by-Sept-30 prediction market from 36¢ to 57¢. Here is what the signal means and what can still go wrong.
TL;DR
- Two new Claude models codenamed Marshmallow and Melon surfaced in developer apps and Discord around August 24, signaling active external testing at Anthropic.
- The prediction market for a next Claude Opus release by September 30 jumped +21¢ in 24 hours to 57¢, repricing from a long-shot to a coin-flip-plus.
- The community's leading interpretation is that Marshmallow equals Claude Opus 5.1 — a polish cycle on the July 24 Opus 5 flagship.
- Five weeks remain until the hard September 30 deadline; the risk is real that testing extends, Anthropic holds for Q4, or Marshmallow turns out to be something other than a production Opus release.
A codename is not a ship date. But the market has decided Marshmallow is close enough to treat seriously — and it moved the number accordingly.
What the Market Says
At press time the Polymarket contract "Will the next Claude Opus model be released by September 30, 2026?" sat at YES 57¢ and NO 43¢, captured August 26, 2026 at 10:30 UTC. Twenty-four hours prior, YES was trading at 36¢. That is a +21¢ single-day move on $749 of volume — not deep liquidity, but a clear directional reprice on a specific catalyst.
The resolution date is October 1, 2026, meaning any Anthropic Opus release must be publicly live by midnight September 30. A model that ships October 1 resolves NO. That hard stop matters more than it might appear: Anthropic has never, to this observer's knowledge, coordinated a launch to match a Polymarket deadline.
The Case
The catalyst is a pair of model identifiers — claude-marshmallow-eap and claude-melon-eap — that began appearing in third-party developer applications and on Discord around August 24. Times of AI reports that neither model has received any official announcement from Anthropic, and that the dominant community reading is Marshmallow as Claude Opus 5.1. Universe of AI notes that the food-themed naming follows the same internal convention as earlier experimental identifiers, including a prior experiment called Horchata. Melon's tier placement is less settled; the working hypothesis is a Haiku-level update, though the evidence for that is thinner.
The eap suffix is the operative detail. Early Access Program designations at Anthropic have historically corresponded to external testing phases that run approximately two to four weeks before a public API release. If that clock started around August 24, the math lands inside the September 30 window — barely.
AI Tools Review provides the relevant base rate: Anthropic has shipped five major Claude generations in eighteen months, and point updates have historically followed major versions by four to eight weeks. Claude Opus 5 launched July 24, 2026. Eight weeks from that date is September 18. The September 30 deadline is ten weeks out. On a pure cadence argument, the market's 57¢ is not obviously wrong.
The economic case for a fast point release is also reasonable. Opus 5 was positioned as frontier-performance-at-mid-tier-pricing, a genuinely competitive product claim. A 5.1 patch cycle — reliability fixes, context-window tuning, latency improvements — is the standard way Anthropic has historically defended that claim against model drift and competitive pressure. Point releases don't require the same training overhead as a generational jump. Development teams working on incremental updates from an already-trained base can move faster than the headline release cadence implies.
So: credible signal, plausible timeline, defensible base rate. The market moved from imminent-but-unlikely to plausible-within-weeks, and 57¢ is a fair summary of that view — not a conviction bet, but not a fade either.
Risks
The honest case for the NO side is not trivial.
Leaks are not roadmaps. The appearance of claude-marshmallow-eap in a developer app tells you Anthropic is testing something with that identifier. It does not tell you that the model will pass internal quality bars, that the release is scheduled for this quarter, or that "Marshmallow" maps to a production Opus variant rather than an experimental cost-optimized build. Anthropic runs multiple parallel research tracks. Some of them do not ship.
Opus 5 is two months old. That is an extremely short interval by historical standards for a follow-on Opus release. Anthropic may judge that the more valuable use of the next five weeks is stabilization, training data curation, and post-launch feedback incorporation — none of which produces a new model identifier by September 30.
September 30 is a hard and arbitrary stop. Anthropic does not time product releases to prediction-market deadlines. A model that is 95% ready on September 30 and ships October 3 resolves this contract NO. The deadline creates asymmetric risk for YES holders that doesn't exist in the underlying development timeline.
Marshmallow could be something other than a production Opus release. If it is a cost-optimized inference variant, a safety-tuned research model, or a regional deployment build, the market resolves NO regardless of whether Anthropic ships it. The community interpretation is plausible, not confirmed.
Volume is thin. $749 in 24-hour volume is enough to move a low-liquidity market by 21 cents without requiring strong conviction from the moving party. The price is a signal; it is not an oracle.
At 57¢, YES traders are paying roughly fair value for a coin-flip with a hard deadline and a soft catalyst. That is a defensible position. It is also a position that can go to zero on an October 1 press release.
Prices captured at press time and are not live. Not financial advice.
Every price in this piece was captured Aug 26, 2026 10:30 UTC. Odds move; the analysis may not age with them. Not financial advice.