Claude Opus Next: Polymarket Traders Just Paid 86 Cents on Leaked Codenames and One Engineer's Tweet

TL;DR - Polymarket's Claude Opus release contract moved from 57¢ to 86¢ in three calendar days, a 51% price surge driven entirely by unconfirmed leaks and a single employee post. - Two internal model identifiers — claude-marshmallow-eap and claude-melon-eap — leaked into the AI community around Aug 24–26, sparking speculation about an imminent Opus 5.1 launch. - Anthropic has made no official announcement, posted no pricing, and published no API documentation for any successor to Opus 5 as of Aug 29. - The resolution contract hinges on "general public" availability, a phrase that has historically done a lot of heavy lifting in frontier AI releases.

Polymarket traders pushed the odds of a new Claude Opus model shipping before October to 86 cents on the back of two leaked codenames and one engineer's candid acknowledgment that the current model talks too much. The question is whether that confidence is earned — or whether the market just bought a rumor at a markup.

What the Market Says

At 86¢ as of 2026-08-29 10:37 UTC, the market is pricing an 86% probability that Anthropic releases its next Claude Opus model to the general public before October 1, 2026. Three days earlier, on Aug 26, the same contract sat at 57¢. The 29-cent move — on 24-hour volume of $732 — represents one of the sharper short-term dislocations in this contract's history, and it happened without a single word from Anthropic's communications team.

The NO side, at 14¢, implies roughly a one-in-seven chance the deadline is missed or the release doesn't meet the contract's "general public" threshold. Given how much is still unconfirmed, 14¢ looks like a bargain that the market isn't currently interested in.

The Case

The catalyst breaks into two parts.

Part one: the codenames. Around Aug 24–26, two early-access model identifiers surfaced across X posts and secondary AI community coverage, subsequently reported by atoms.dev on Aug 27: claude-marshmallow-eap and claude-melon-eap. The community consensus — and it is community consensus, not Anthropic documentation — maps Marshmallow to an Opus 5.1 candidate and Melon to a Sonnet 5.1 candidate. The shared eap suffix is being read as "early access program" or "evaluation access program," signaling that at least some external users are touching these models.

The atoms.dev report is worth quoting directly on the limits of what is actually known:

"The strongest leak signals are claude-marshmallow-eap and claude-melon-eap. Community reports generally describe Marshmallow as stronger than Melon, with some users saying it is more pleasant to use than Opus 5. No official launch post, API model ID, pricing page, system card, rate-limit notice, or downloadable weights were verified for Opus 5.1 at the research cutoff [Aug 27]."

That is not a launch signal. That is a sighting.

Part two: the engineer post. On Aug 26, Anthropic engineering lead Boris Cherny posted publicly on X responding to criticism of Claude Opus 5's tendency toward long, padded responses. His words: "Opus excels at long-running work and coding but has quirks... verbosity is top of mind for the team." He offered a short-term workaround — claude /config outputStyle=concise — and said nothing about release timelines, new model versions, or product roadmaps.

The market read this as confirmation that something is coming. That interpretation is possible. It is also a significant inferential leap. Acknowledging a known problem and offering a config-level workaround is exactly what a team does when the fix is not imminent.

Still, traders connected the dots: leaked internal identifiers plus an engineer publicly signaling active work on the product's biggest complaint equals a launch that is close. Community speculation in late August converged on a "this week" window. The contract price followed.

The baseline case for YES is not unreasonable. Anthropic's release notes, accessed Aug 29, show Claude Opus 5 shipped on July 24, 2026 at $5 per million input tokens and $25 per million output tokens. A September delivery of an incremental successor would fit roughly a monthly cadence. The EAP identifiers suggest models in some form of structured external evaluation, which is a further stage along the pipeline than pure internal development. If Anthropic is running an early-access program in late August, a general launch in September is at least a coherent timeline.

The 86¢ price says the market finds that timeline not just coherent but highly probable. That is a different claim.

Risks

The NO case at 14¢ deserves a fair hearing, because the resolution language is where this contract gets complicated.

1. "General public" is not a trivially satisfied condition. The market resolves YES only if the next Claude Opus model is "made available to the general public." Anthropic's standard playbook for frontier model launches is to gate initial access to Claude Max subscribers or enterprise tiers, then roll out to API and lower tiers over subsequent weeks. If Marshmallow ships on September 15 but only to Max subscribers and enterprise API customers, a reasonable resolution judge could find that the general public threshold has not been crossed. Traders buying at 86¢ may be underweighting this specific failure mode.

2. EAP does not mean imminent launch. Early-access identifiers can represent partner experiments, red-teaming programs, or evaluation pipelines that never reach a standard product release. The community inference that eap equals "almost out the door" is plausible but unverified. Internal identifiers have leaked from Anthropic before without yielding immediate public products.

3. The verbosity fix could be a config update, not a new model. Cherny's workaround is a configuration flag, not a teaser. If Anthropic addresses verbosity through system-prompt engineering or a behavior patch to the existing Opus 5 API endpoint, there is no new Opus model to release. The contract would resolve NO, and traders at 86¢ would have paid a steep premium for a product problem that got solved with a settings change.

4. Safety evaluation can override cadence. Anthropic has delayed releases on safety grounds before. A roughly monthly release cadence is a data point, not a contractual commitment. If internal red-teaming on Marshmallow surfaces something concerning in September, the company has demonstrated willingness to hold a model. One month's delay past September 30 resolves this contract NO regardless of how close the model was.

5. Volume context. This contract is moving on $732 in 24-hour volume. At that liquidity level, a handful of informed traders — or a handful of enthusiastic speculators — can shift the price materially. The 86¢ figure reflects the marginal buyer's conviction, not a deep, liquid consensus. Price discovery in thin markets should always carry a mental asterisk.

The honest synthesis: something real appears to be in Anthropic's pipeline. The codenames are credible signals of internal activity at a stage beyond pure development. The engineer's post confirms the team is actively addressing Opus 5's most publicly criticized behavior. A September release of an incremental Opus model is a plausible scenario.

Whether it is an 86% probability scenario is a different question. The contract's resolution language, Anthropic's tiered launch history, and the genuine ambiguity of what "EAP" means in production terms all argue for a somewhat lower number. The market moved fast on thin volume in response to unconfirmed signals. That is not necessarily wrong — markets do discover information through inference — but it is a posture that rewards scrutiny before following.

At 14¢, NO is not a comfortable fade either. The underlying momentum is real. But the gap between "something is coming" and "something is coming to the general public before October 1" is exactly where this trade lives, and 86 cents does not leave much room for the known unknowns.

Prices captured at press time and are not live. Not financial advice. Independent publication - not affiliated with Polymarket, Banana Gun, or any venue.