Two unannounced Claude models just surfaced in Anthropic's API — and one may already beat Opus 5

Anthropic's next lineup has leaked the old-fashioned way: developers spotted two never-announced models, claude-marshmallow-eap and claude-melon-eap, sitting in the company's API. Multiple developers confirmed both codenames through API access and Discord, and early testers report that Marshmallow's conversational performance already edges out Opus 5 — Anthropic's current flagship, released barely a month ago. Neither model carries a Fable-class designation, so community consensus points to them as the next rungs on the Opus and Sonnet ladders — with TheWinCentral speculating "Marshmallow" is an early Claude Opus 5.1 and "Melon" a Claude Sonnet 5.1.

The "-eap" suffix marks both as Early Access Program builds: internal test stages opened to only a small circle of developers. That framing matters, because codenames that appear in a public API are neither final nor committed — they can be renamed, merged, or cancelled entirely. But the circumstantial evidence for an imminent refresh is strong. Opus 5 shipped July 24 and Sonnet 5 on June 30; a new pair entering EAP barely a month later would push Anthropic's iteration cadence to roughly a major update every two weeks this year. Even the internal naming scheme has shifted — previous leaks used animal codenames like "Fennec" for Sonnet 5.5, while this pair switches to foods, which some read as a sign the models belong to a fresh development branch.

The timing lands at a delicate moment. Rivals are cutting prices to court cost-sensitive enterprise buyers, Anthropic's IPO paperwork is reportedly weeks away, and its own users have been grumbling about silent quality changes — a controversy over Claude Code quietly remapping "high" reasoning effort blew up just yesterday. Leaking your next-generation models into the wild is usually an accident; leaking them this loudly, right before a public listing, looks almost like a rehearsal.

What would it take for you to switch from your current Claude tier — a smarter Marshmallow, or a cheaper Melon? Tell us in the comments.

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Read more

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute