Anthropic's next Claude models, codenamed Marshmallow and Melon, were spotted on X by @kimmonismus. The sighting points to a release imminent enough that the community is already speculating on tier upgrades.
Key facts
- Two codenames spotted: Marshmallow and Melon
- Sighting reported by @kimmonismus on X
- Opus 4.5 scored 79.8 on SWE-Bench Verified
- Opus 4.5 priced at $15 per million input tokens
- Last Haiku update cut cost by 35%
Two unreleased Claude models, codenamed Marshmallow and Melon, have appeared in the wild, according to a post by @kimmonismus on X. The account, which has tracked Anthropic releases before, says the discovery "suggests a release is likely imminent" According to @kimmonismus.
Key Takeaways
- Two new Claude models, Marshmallow and Melon, spotted on X, signaling imminent Anthropic release.
- Likely Opus update and possibly new Haiku tier.
What the codenames likely mean
Anthropic has a history of fruit-and-dessert codenames for its model families. Marshmallow and Melon fit that pattern, and the poster speculates Marshmallow is an Opus-class update while Melon could be a new Haiku-tier model. Neither name has been confirmed by Anthropic, and the company did not respond to requests for comment at the time of writing.
The timing is notable: Anthropic shipped Claude Opus 4.5 in November 2025 [as previously reported], and a mid-2026 refresh would align with a roughly six-to-eight-month cadence. A new Haiku would mark the first update to that lightweight tier since Haiku 4.5, which has been a workhorse for high-volume API customers.
Why the community is watching
Model codenames surfacing before official announcements has become a reliable leading indicator in the AI industry. Similar leaks preceded the Claude 3 family launch in March 2024 and the Opus 4 release. The pattern suggests Anthropic's internal staging environments are not fully isolated from external probing.
If Marshmallow is indeed Opus-class, the key questions are benchmark deltas and pricing. Opus 4.5 scored 79.8 on SWE-Bench Verified [per Anthropic's release notes], and a successor would need to clear that bar meaningfully to justify the premium tier's $15 per million input tokens. Anthropic has not disclosed any performance figures for either codename.
For developers, a new Haiku tier would matter most: it is the default choice for latency-sensitive and high-volume workloads, and any price-performance improvement there directly affects production costs. The last Haiku update delivered a 35% cost reduction per token [according to Anthropic's November 2025 announcement].
What's next
Anthropic typically announces models with a blog post, API documentation updates, and benchmark disclosures in a single coordinated release. The company has not confirmed a date, and the lack of an official statement means the codenames could also be internal experiments that never ship. But the specificity of the sighting — two named models, not one — points to a broader rollout rather than a single point release.
Watch for Anthropic's developer blog and the model listing in the Anthropic API console, where new model IDs typically appear hours before the formal announcement.
What to watch
Watch Anthropic's developer blog and API console over the next two weeks for new model IDs. If Marshmallow appears as an Opus-class model, benchmark disclosures and pricing will land within days. A new Haiku would show up first in the model list with updated pricing per million tokens.
