Two unreleased Claude models, codenamed Marshmallow and Melon, have leaked into developer testing — and early results contain the kind of output LLMs were not supposed to produce. Testers report one-shot generation of complex 3D spatial layouts and architectural scenes, plus conversational polish they rate above Opus 5. The catch: both models burn so many thinking tokens that testers repeatedly hit their maximum-token limits. If the reports hold, Anthropic is moving the frontier fight from training-phase scale to inference-phase reasoning — and from language to spatial intelligence.
Two Leaked Claude Models Are Changing the Odds
Neither model is confirmed by Anthropic. What leaked is hard evidence of an internal push: claude-marshmallow-eap and claude-melon-eap were spotted in API traffic on August 21, 2026, between 00:45 and 00:57 UTC, when the marshmallow-ht-eap identifier was called 57 times. The same model then appeared in Claude Code as a "Custom model" with a 1,000,000-token context window — the longest Anthropic has fielded in public testing channels.
Developers who got access report three things that matter. First, both models are unusually strong at spatial reasoning: 3D reinforcement-learning tasks and architectural layout planning, handled one-shot, with gravity, topology, geometric and load-bearing constraints respected on the first pass. Second, conversation quality feels better than Opus 5, the flagship Anthropic released on July 24, 2026, to a lukewarm reception. Third, they consume compute at an alarming rate for a frontier lab: testers say the models chew through thinking tokens until they trigger <max-tokens> limits mid-response.
What Leaked: Codenames, Traffic, and First Test Results
The food-based naming continues a tradition: in July 2026, Anthropic briefly tested claude-horchata-eap ("horchata", a Mexican drink). Marshmallow is reported slightly stronger overall than Melon; both sit below the top-tier Fable line. There are no first-party API endpoints yet, and access appears limited to red-teamers and internal testers — which is typical of a model being readied for rollout.
Two readings of the leak are circulating. One: these are Opus 5.1 builds, an emergency response to Opus 5's weak reception — an attempt to forcibly raise the reasoning ceiling with heavy thinking tokens. Two: they are the next Sonnet or Haiku generation, a faster and cheaper class that absorbed the new spatial capability. The 1M-context Custom model appearance in Claude Code is a strong hint that Anthropic wants these tested inside real coding workflows, not just benchmark suites.
Why 3D Was the Hard Part — and Why Thinking Tokens Matter
For LLMs, spatial imagination has been a chronic weak spot: the physical coordinates of a chair inside a 3D room, gravity rules for stacked objects, or circulation and load-bearing planning for a building — these consistently broke language models. What vision was to earlier language models, spatial reasoning is becoming to frontier models: the next modality to conquer, turning text intelligence into world-model capability (for related mechanics, see our overview of test-time compute costs).
Thinking tokens are the second half of the story. Instead of a fast-talking probability sampler, these models spend hidden tokens on deep self-derivation, chain-of-thought construction and logical trial-and-error before answering. That is the same economics family as how frontier models are now being evaluated: the value shifts from what was learned at training time to how much genuine thinking happens at inference time — and that thinking is billed per token.
What It Means: The Reasoning Economy and the Spatial Frontier
Take the reports at face value and two structural moves are visible. First, competition is moving from training-phase compute to inference-phase reasoning. The scarce resource stops being the data center that trained the model; it becomes the reasoning budget every query is allowed to spend. Expect pricing and latency to reorganize around thinking tokens, with new billing lines appearing in agent workloads.
Second, the spatial frontier unlocks application categories that text alone could not reach. A game developer describes "a medieval fortress with three defense towers and a hidden underground passage" and receives a 3D layout in one shot. An architect asks for dozens of structurally valid concept layouts given terrain and lighting — generated in minutes. Digital twins, metaverse environments, autonomous-driving simulation and robot spatial planning all sit on the same capability. Anthropic's cadence also matters: with Opus 5 landing softly and an IPO on the horizon, the leaked timing reads like a fast "revenge release" — Anthropic cannot afford a weak generation.
What to Do Now
1. Watch for official endpoints. Neither model is confirmed; an announcement is likely within weeks, and the Custom model listing is the tell that rollout testing has started.
2. If you run Claude Code, check the Custom model list — the 1M-context entry is where these builds have been appearing.
3. Add spatial tasks to your eval suite now — room-layout, floor-plan and object-placement tests. 3D reasoning will be the differentiator of the next model generation.
4. Budget thinking tokens in agent workloads. Inference costs are shifting from training-scale amortization to per-query reasoning; track token burn the same way you track GPU hours.
5. Plan for the class refresh. The Fable tier remains the quality ceiling, but a faster, spatially capable mid-tier could reset the defaults for coding, data and agent products.
FAQ: Marshmallow, Melon, and Thinking Tokens
What are Claude Marshmallow and Melon?
Codenames for two unreleased Anthropic models — claude-marshmallow-eap and claude-melon-eap — first spotted in API traffic on August 21, 2026, and in Claude Code as a Custom model with a 1,000,000-token context. Testers report one-shot 3D spatial layout generation and conversation quality above Opus 5, still below the Fable tier.
What is a thinking token?
A token a model spends during inference on internal reasoning — chain-of-thought construction and logical trial-and-error — before producing the final answer. Testers say Marshmallow and Melon burn so many that they repeatedly hit maximum-token limits, evidence that the competition is moving from training-phase scale to inference-phase reasoning.
When will these models be officially released?
No date has been announced. Speculation ranges from an Opus 5.1 refresh — a response to Opus 5's weak reception — to the next Sonnet or Haiku generation. The Custom model appearance in Claude Code suggests rollout testing is already underway.
