Two unlisted Claude model identifiers surfaced in developer tools and on Discord late last week, and the timing lines up with a separate report that Anthropic is quietly benchmarking a frontier system well past its current flagship. Put those two threads together and you get the story spreading across tech outlets this week: Anthropic may be sitting on not one but two unannounced Claude models, and a double release could land before the year is out.
Neither model has a public system card, a pricing page, or a confirmed ship date. But the pattern is hard to ignore. Developers spotted new API identifiers in the wild, Anthropic’s own August risk report quietly named an internal system that beats its newest public model, and industry chatter about a possible $100 billion Anthropic valuation has put every scrap of roadmap information under a microscope. Here is what has actually been confirmed, what remains speculation, and why a double release would be a genuine departure from how Anthropic has shipped models all year.
The Leak: Two New Model IDs Show Up in the Wild
The story starts with a pair of strings that Anthropic never announced. According to Times of AI, two new identifiers, claude-marshmallow-eap and claude-melon-eap, appeared in third-party developer apps and on Discord around August 24, 2026. Neither name traces back to an official Anthropic announcement. There is no model card, no listed release date, and no entry on the public API pricing page for either one.
The “eap” suffix is the detail developers latched onto first. It typically signals an Early Access Program build, the same kind of pre-release tagging that has preceded past Claude launches. That is not proof of an imminent release on its own, but it is the strongest signal yet that Anthropic is running two separate model tracks in parallel rather than one.
claude-marshmallow-eap
claude-melon-eapPer Times of AI’s reporting, the most common theory among developers who spotted the strings is that Marshmallow corresponds to a Claude Opus 5.1 update and Melon maps to a Claude Sonnet 5.1 refresh. A smaller group of developers floated a different read: that Melon could represent an entirely new Haiku-class model rather than a Sonnet point release. Anthropic has not confirmed either theory, and the company has a track record of shipping codenames that do not map cleanly onto the tier they end up shipping in.
Model 2: The Internal System Anthropic Didn’t Mean to Reveal
The second thread predates the Marshmallow and Melon leak by about ten days and comes from a very different source: Anthropic itself. On August 14, 2026, the company published a 186-page company-wide risk report, and buried inside it was the first public acknowledgment of a system referred to only as “Model 2,” reported BigGo Finance. The report described Model 2 as more capable than Claude Mythos 5, Anthropic’s strongest publicly released model at the time, and stated explicitly that the company has no current plans to release it externally.
GizChina’s coverage of the same disclosure notes that no release date has been set and frames Model 2 as a research and acceleration tool being used internally rather than a product in the pipeline for near-term launch. Anthropic’s own system cards page does not list Model 2 at all, which tracks with the company’s insistence that the system is not on a public roadmap.
Separate reporting adds a number to the story. A Seeking Alpha piece on Anthropic’s reported ambitions for a $100 billion IPO said the company’s researchers are actively benchmarking Model 2, describing it as widely speculated to become a future Claude Mythos-line release. Per that report, early internal evaluations show Model 2 scoring 62.8% on Anthropic’s CoBench v2 development benchmark, compared with 50.3% for Claude Mythos 5. That is a meaningful jump on an internal yardstick Anthropic uses to measure AI research and development capability, though it is worth stressing this is an unreleased, unaudited internal number, not a published benchmark result.
Why the “Double Release” Framing Is Catching On
Individually, either story would be a minor footnote. A pair of unlisted API strings is not unusual for a company that ships as often as Anthropic does, and a research lab quietly testing a stronger internal model is standard practice at every major AI lab. What has pushed this into breaking-news territory is the overlap: two unexplained model identifiers surfacing within days of a disclosure that Anthropic is running an internal system that beats its current flagship, all inside the same two-week window in August 2026.
That overlap is why outlets including thewincentral.com and Seeking Alpha have been treating the “two new models” angle as the story worth chasing this week, rather than either leak in isolation. None of the reporting confirms that Marshmallow, Melon, and Model 2 are the same thing, or that all of it lands on the same release date. It is entirely possible Marshmallow and Melon are minor point releases while Model 2 stays internal for months longer. But the clustering of signals is exactly the kind of pattern that has preceded real Anthropic launches before, most recently the “Honeycomb” leak cycle that predicted Claude Sonnet 5, Fable 5, and Opus 5 weeks ahead of their actual ship dates.
The Honeycomb Precedent: Anthropic’s Last Leak Cycle Played Out Exactly Like This
This would not be the first time stray identifiers turned into a real Claude release inside a few weeks. According to a running roadmap tracker maintained by AI Tools Review, a set of leaked codenames dubbed the “Honeycomb” cycle correctly predicted three separate launches earlier this summer. Claude Sonnet 5 shipped on June 30, Fable 5 and Mythos 5 were restored to availability on July 1, and Claude Opus 5 launched as the new general-purpose flagship on July 24. All three matched names that had circulated as unconfirmed leaks in the weeks prior.
That same tracker noted that, as of early August, one loose thread remained: a possible Claude Fable 5.1 update that had not shipped and had no system card, pricing page, or Anthropic announcement to back it up. The Marshmallow and Melon identifiers that surfaced August 24 could be an extension of that same unresolved thread, or a fresh cycle entirely. Anthropic does not comment on unreleased products, so outside observers are left triangulating from API strings, developer forum posts, and whatever the company lets slip in unrelated documents like its risk reports.
Claude’s 2026 Release Cadence So Far
To understand how unusual a double release would be, it helps to see how tightly spaced Anthropic’s confirmed 2026 launches have already been. The company has shipped a new model roughly every six to eight weeks this year, alternating between the Opus, Sonnet, and Mythos lines rather than releasing everything at once.
| Model | Confirmed Release Date | Tier / Line | Status |
|---|---|---|---|
| Claude Opus 4.6 | February 4-5, 2026 | Opus | Superseded |
| Claude Sonnet 4.6 | February 17, 2026 | Sonnet | Superseded |
| Claude Opus 4.8 | May 2026 | Opus | Active |
| Claude Sonnet 5 | June 30, 2026 | Sonnet | Active, default Sonnet |
| Claude Fable 5 / Mythos 5 | July 1, 2026 | Mythos line | Active |
| Claude Opus 5 | July 24, 2026 | Opus | Active, current flagship |
| Claude Marshmallow / Melon (eap) | Unconfirmed | Unknown (Opus/Sonnet suspected) | Leaked identifiers only |
| Model 2 | No release planned | Unknown, Mythos-line suspected | Internal only |
Notice the gap between Opus 5 shipping on July 24 and the Marshmallow and Melon identifiers appearing exactly a month later on August 24. If Anthropic is preparing 5.1-class updates to both Opus and Sonnet at the same time, that would break from its recent one-model-at-a-time pattern and mark the first genuinely simultaneous dual launch of the year. Readers who want the full history of how this cadence has evolved can check our earlier coverage of why Anthropic’s August already looked different from every prior month before this latest leak even surfaced.
What the Risk Report Actually Says About Model Safety
The same August risk report that named Model 2 also raised Anthropic’s own internal misalignment rating from “very low” to “low,” a change the company attributed to the increasing capability of its models rather than any specific incident. The redacted version of the report, published on Anthropic’s own site, includes model-by-model figures for a metric the company tracks internally, showing rates of 1.4% for Claude Opus 4.7, 0.27% for Claude Opus 4.8, and 2.7% for Claude Fable 5. Anthropic has not published an equivalent figure for Model 2, and the company’s Transparency Hub does not yet reference it by name.
That distinction matters for the double-release story. A rising internal risk rating, paired with a stronger unreleased model, is consistent with a company that is being more cautious about what it ships next rather than rushing two models out the door. It is possible that caution is exactly why Marshmallow and Melon, if they are indeed close to release, have not been formally announced even after their identifiers leaked.
Benchmark Data Points Fueling the Speculation
The numbers driving most of this week’s coverage come from CoBench v2, Anthropic’s internal benchmark for measuring AI research and development capability. It is not a public leaderboard metric like MMLU or SWE-bench, which makes the reported scores harder to independently verify, but it is the same benchmark Anthropic itself cited in the risk report that first named Model 2.
| System | CoBench v2 Score | Source | Availability |
|---|---|---|---|
| Claude Mythos Preview | 54.8% | Anthropic risk report (via AIBase) | Limited |
| Claude Mythos 5 | 50.3% | Anthropic risk report / Seeking Alpha | Public |
| Model 2 (internal) | 62.8% | Seeking Alpha | Not released |
| Claude Marshmallow / Melon (eap) | Not disclosed | Times of AI | Leaked identifiers only |
Two things stand out. First, Mythos Preview scoring higher than the publicly shipped Mythos 5 on this particular benchmark is a reminder that internal test builds do not always translate cleanly into shipped performance, which is a useful caveat for anyone assuming Model 2’s 62.8% figure will hold once, or if, it becomes a real product. Second, there is no published benchmark data at all for Marshmallow or Melon, since neither has a model card. Any performance claims circulating about those two identifiers right now are speculation, not measured results.
The $100 Billion Backdrop
None of this is happening in a vacuum. Reports about Anthropic exploring a roughly $100 billion valuation through a potential IPO have circulated alongside the Model 2 disclosure, and it is not hard to see why the timing matters. A frontier lab preparing to raise capital or go public has every incentive to demonstrate that its research pipeline extends well past whatever it currently sells. Model 2 scoring meaningfully higher than the public flagship on an internal benchmark serves that narrative, whether or not the company intends to ship it soon.
That context also explains some of the skepticism worth applying here. A “we have something better internally” disclosure is a common move for AI labs navigating funding conversations, and it does not by itself guarantee two consumer-facing models are about to ship. Readers who want more background on how Anthropic has been positioning its market opportunity this year can see our coverage of Anthropic’s $30 trillion market claims tied to AI-run lab robots, which covers a separate but related pattern of the company talking up long-term addressable market ahead of major funding milestones.
Competitive Pressure Is Part of the Story Too
Anthropic is not shipping models in isolation. DeepSeek’s V4-Flash-Vision-Exp release was reported to come close to matching Claude Opus 4.8 on multimodal agent benchmarks, and every major lab, OpenAI, Google, and now several well-funded Chinese labs, is on a release cadence measured in weeks rather than quarters. A double release, if it happens, would fit a broader industry pattern where labs increasingly ship an efficient mid-tier model and a frontier flagship close together to cover both the cost-sensitive and capability-maximizing ends of the market at once.
That pattern is visible in how Anthropic’s existing lineup is already priced and positioned across tiers, something we broke down in detail in our comparison of Claude Opus 5 against Mythos 5 and Gemini 3.1 Pro. If Marshmallow becomes an Opus 5.1 refresh and Melon becomes a Sonnet 5.1 or new Haiku-tier model, it would extend that same two-speed strategy rather than break from it.
What Developers and Enterprises Should Actually Do Right Now
For teams building on the Claude API, the practical advice is unglamorous: nothing changes today. Neither Marshmallow nor Melon has a documented API endpoint, a pricing tier, or a migration path, and Anthropic has not published guidance for either. Building against unconfirmed model identifiers is a good way to end up debugging a 404 the moment Anthropic rotates or deprecates a leaked EAP string, which is a routine practice for early access builds that never reach general availability.
Teams evaluating which current model to standardize on can get a clearer picture from existing, shipped comparisons rather than speculation. Our breakdown of Claude Fable 5 against Opus 5 and GPT-5.6 Sol covers the pricing and context-window tradeoffs across today’s confirmed lineup, and remains the more actionable reference point until Anthropic formally announces whatever ships next.
What’s Confirmed vs. What’s Still Rumor
It is worth separating the pieces of this story cleanly, because they carry very different levels of certainty.
- Confirmed: Two unlisted identifiers, claude-marshmallow-eap and claude-melon-eap, appeared in developer tools and on Discord around August 24, 2026, per Times of AI.
- Confirmed: Anthropic’s August 14 risk report disclosed an internal system called Model 2 that outperforms Claude Mythos 5 on the company’s CoBench v2 benchmark, per BigGo Finance and Anthropic’s own published report.
- Confirmed: Anthropic has stated it has no current plans to release Model 2 externally.
- Unconfirmed: Whether Marshmallow and Melon correspond to Opus 5.1 and Sonnet 5.1, or to a new Haiku-class model.
- Unconfirmed: Whether Marshmallow, Melon, and Model 2 are related systems at all, or three unrelated internal projects.
- Unconfirmed: Any release date for any of the three systems.
Historical Context: How Often Leaks Turn Into Real Launches
Anthropic’s track record with leaked identifiers is mixed but leans toward eventual confirmation. Earlier in 2026, a data leak first revealed the existence of Claude Mythos months before its formal release, an episode covered at the time by Seeking Alpha. The Honeycomb cycle discussed above turned three separate leaked names into three shipped products within about a month. Not every leaked string ships, and Anthropic has quietly dropped or renamed internal projects before without ever making them public. But the base rate for this specific company, once an identifier reaches developer tooling rather than staying confined to internal Slack channels, has trended toward eventual release rather than permanent cancellation.
Market Impact: Why This Matters Beyond Claude Users
A double release, if confirmed, would land at a moment when Anthropic’s competitive position is already under scrutiny from multiple directions: reported IPO plans that would put a hard number on the company’s valuation, rising benchmark scores from DeepSeek and other rivals, and a self-reported increase in its own model risk rating. Enterprises evaluating multi-model strategies, including those comparing Claude against Perplexity for research and retrieval workloads, will want to watch whether Anthropic ships a genuinely simultaneous Opus and Sonnet refresh or staggers the releases as it has done all year. A true double release would be a signal that Anthropic has enough compute and engineering headroom to run two release pipelines in parallel, which matters as much for procurement planning as it does for raw model capability.
Predictions: What Happens Next
Based on the confirmed pattern of leaks preceding launches this year, here is what is likely to happen over the coming weeks.
- Anthropic will likely confirm at least one of the two leaked identifiers, Marshmallow or Melon, within the next several weeks, following the same lag seen between the Honeycomb leaks and their eventual ship dates.
- A simultaneous, same-day release of both an Opus 5.1 and Sonnet 5.1-class model is possible but not yet the base case. Anthropic’s 2026 pattern has favored staggered releases spaced weeks apart rather than same-day dual launches.
- Model 2 is unlikely to reach public availability in the near term. Anthropic’s own statement that it has no plans to release the system, combined with the company’s rising internal risk rating, points toward continued internal-only use for the foreseeable future.
- Expect further leak-driven coverage before any official Anthropic announcement, given how closely developer communities are now monitoring API identifiers after the Honeycomb cycle proved predictive.
- Any formal Anthropic announcement will likely arrive alongside updated system cards and pricing, consistent with how Opus 5 and Sonnet 5 were both announced with same-day availability details across the Claude API, AWS Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
The Bottom Line
Two things are true at once here. Anthropic almost certainly has more than one unreleased Claude-related system in some stage of development right now, that much is corroborated by its own risk report and by independently spotted API strings. But “two models exist internally” and “a coordinated double release is imminent” are not the same claim, and only the first one currently has solid sourcing behind it. Until Anthropic publishes a system card, a pricing page, or a release note for Marshmallow, Melon, or Model 2, the responsible read is that this is a fast-moving, multi-outlet story built on real signals, not yet a confirmed product announcement.
Frequently Asked Questions
What are claude-marshmallow-eap and claude-melon-eap?
They are two unofficial API identifiers that surfaced in third-party developer apps and on Discord around August 24, 2026, according to Times of AI. Neither has a public model card, pricing listing, or confirmed release date from Anthropic.
Is Anthropic really preparing two new Claude models?
The evidence points that way but stops short of confirmation. Two separate identifiers leaked within days of Anthropic’s own risk report disclosing a stronger internal system called Model 2. That combination is why multiple outlets are treating a double release as plausible, though Anthropic has not made an official announcement.
What is Model 2?
Model 2 is the internal name Anthropic used in its August 2026 risk report for a system it says outperforms Claude Mythos 5 on its internal CoBench v2 benchmark, scoring 62.8% versus 50.3% for Mythos 5. Anthropic has stated it has no current plans to release Model 2 externally.
Will Model 2 become Claude Mythos 6?
That is speculation reported by Seeking Alpha, not a confirmed naming decision from Anthropic. The company has not stated that Model 2 will ever ship publicly, under that name or any other.
What is the “Honeycomb” leak cycle mentioned in this story?
Honeycomb is the name a roadmap-tracking site used for a set of leaked codenames earlier in 2026 that correctly predicted the launches of Claude Sonnet 5, Fable 5, Mythos 5, and Opus 5 within weeks of the leaks appearing.
How does this connect to Anthropic’s reported IPO plans?
Reports of Anthropic exploring a roughly $100 billion valuation surfaced around the same time as the Model 2 disclosure. A frontier lab preparing for a major fundraising step has an incentive to show investors it has capability in reserve beyond its current public models, which is part of why the timing of these disclosures has drawn extra scrutiny.
Should developers build against the leaked model identifiers now?
No. Neither identifier has a documented, stable API endpoint. Early access strings can be rotated or removed without notice, and Anthropic has not published any integration guidance for either name.
When will Anthropic officially confirm what’s next?
There is no confirmed date. Based on the gap between the Honeycomb leaks and their eventual launches earlier this year, an official announcement, if one comes, would most plausibly arrive with a full system card and pricing details rather than a piecemeal reveal.