Nano Banana 2.1 pushes native 4K and price cuts, Jev hits a $7.5B valuation, dexterous hands close the tactile loop

The yardstick for this round of AI competition is changing tracks: no longer just whether a model can write beautiful prose or generate stunning images, but whether it can be slotted into a production line, called reliably by software, and actually held in a robot's hand. From falling image-generation costs, to decision models becoming capital's new favorite, to dexterous hands writing touch into the control loop, three seemingly unrelated developments point at the same thing — deliverability. ^[1][2][3]

A new yardstick for image models: native 4K and half the price

Google released its new-generation image model Nano Banana 2.1 in the early hours, with upgrades focused on visual design, mask editing, subject consistency and a photorealistic, natural look, headlined by native 4K photographic-quality images plus high-frequency editing needs such as freeform retouching and one-click background replacement ^[1]. According to data from the US benchmarking platform Arena, compared with the previous generation Nano Banana 2, the new model improves by an average of about 61.7 points across three capabilities — text-to-image, multi-image editing and single-image editing — ranking fourth in text-to-image, 80 points above the previous generation and behind only GPT ^[1]. Price is another variable: generation for 1K and 2K images is halved outright, and 4K is about 25% cheaper ^[1]. The hands-on cases in the evidence also point to 'usable': user Christopher Fryant showed native 4K output with natural gradients in pupil color and curvature, curled lash tips, and detail in skin texture and the furrow that forms between the brows when frowning; user @归藏's Chinese promotional poster test showed clean Chinese rendering, attractive typography and one-click 4K generation ^[1].

Faster, but the watermark bug and image quality remain weak spots

Better capability doesn't mean there are no weak spots. User Zane compared ChatGPT Images 2.5 and Nano Banana 2.1 generating the same scene on a train platform, concluding the new model is indeed faster but still loses to GPT on texture and overall finish ^[1]. Tests by users Stav Zilbershtein and snow turned up a watermark bug: the initial edit is accurate and good, but watermarks keep stacking, and as edits pile up the image can get worse and blurrier, with noise, color drift and texture distortion appearing in areas that should not have changed ^[1]. Nano Banana 2.1 is now live on the Gemini App, Google AI Studio and other platforms, built on Gemini 3.6 Flash, with 50 free credits per day; developers can also access the paid API under the model name gemini-nano-banana-2.1 for text-to-image, image-to-image, poster generation and 4K upscaling ^[1]. Separately, Heather Cooper fed the new model's frames into Veo 3.1 and turned them into an HD short inside Google Flow — a workflow of 'generate key frames — preserve the subject's style — make the frames move' can now be completed within a single platform ^[1].

Jev: a model that doesn't chat, only judges, lands a $7.5 billion valuation

Unlike the iterative logic of image models, TypeSafe AI's Jev takes a subtractive route: it strips out the generative functions of conventional large models and keeps only one thing — making judgments, and beyond yes/no questions it also chooses among options, with answers closer to yes, no or a given candidate ^[3]. Because the new architecture is so cheap that it is hard to meter, the company simply made output tokens permanently free, charging just $0.042 (about 0.3 yuan) per million input tokens ^[3]. Capital markets reacted far faster: parent company TypeSafe AI announced an $870 million (about 6.308 billion yuan) Series A at a post-money valuation of $7.5 billion (about 54.375 billion yuan), with a16z, Sequoia Capital and DCVC participating, and a16z partner Martin Casado joining the board to take part in the next stage of strategic decisions ^[3]. And it was only on September 15 this year that the company came out of stealth, showing up with its first System One model Jev, a $40 million seed round and a $200 million valuation ^[3]. Founder and CEO Diogo Almeida previously worked at Google Brain and OpenAI, is one of the core authors of the InstructGPT paper and worked on RLHF, ChatGPT and GPT-4; the other two co-founders are COO Sasha Sheng from Meta/FAIR and serial entrepreneur CTO Erik Gafni ^[3].

The growth curve and the rivals that arrived within four weeks

Jev's growth numbers are the most direct footnote to this round. Vercel data shows that within 24 hours of Jev being connected to the AI Gateway, it was adopted by nearly 13% of paying teams — more than double any previous new model release, making it the fastest-adopted model in the gateway's history; a16z disclosed that just 3 days after launch Jev was processing 1 trillion tokens ^[3]. By the time the Series A was announced, TypeSafe AI said about one-third of Fortune 500 companies had started using or integrating Jev, saving customers millions of dollars in production costs; Diogo Almeida also noted daily volume had topped a trillion tokens, with overnight traffic still climbing steadily ^[3]. The window on this track isn't long: Microsoft chairman and CEO Satya Nadella entered with Microsoft-Decision-1, post-trained on Qwen3.5-9B and purpose-built for routing, classification, priority judgment, verification and workflow control ^[3]; on October 6, that is 21 days after Jev's launch, OpenAI opened its Decisions API to public testing, powered by GPT-6 Luna, turning text and images into three kinds of structured answers — Predicate, Choice and Score ^[3]; earlier still, on October 1, Cloudflare had already launched the open-source decision models Clef and Clef-flash, which not only are compatible with the Jev API but put Jev directly on the comparison table ^[3].

Dexterous hands move beyond degrees of freedom: Sharpa spreads touch across the upper body

On the embodied-intelligence side, Sharpa unveiled three products at IROS: D01, an integrated tactile-sensing dexterous manipulation robot; W02, a new-generation fully tactile ultra-compact lightweight dexterous hand; and AE01, a high-fidelity exoskeleton tactile data glove ^[2]. D01's key numbers all revolve around dexterous manipulation: an arm payload-to-weight ratio close to 1:1, maximum end-effector speed above 10.5 m/s, communication frequency of 1000 Hz, repeat positioning accuracy of 0.2 mm, and a spherical wrist joint designed at 1:1 human scale; electronic skin covers the whole body with tactile sensing over the upper body, a tactile sampling rate of 100 Hz, a force-sensing range of 0.1–20 N and force resolution of 0.2 N ^[2]. High-speed motion means the robot can quickly complete dynamic tasks such as handoffs and tool operation, the high payload leaves working headroom during movement, and 0.2 mm accuracy pushes its capability toward fine manipulation ^[2]. W02 makes a counterintuitive trade-off: the previous generation W01 had 22 active degrees of freedom, while W02 cuts that to 21, based on application research showing that degree of freedom is used less often in common grasping and in-hand manipulation and contributes limited value to tasks; the whole hand is about 30% smaller than W01 and lighter still, reducing forearm load and making it easier to fit into workspaces designed around human hand size ^[2]. It also achieves a 'zero grasp radius', handling small-diameter objects such as chopsticks and thin wire more stably; the fingertips use high-resolution visual tactile sensors with a force-sensing range of 5 mN–30 N and spatial resolution of 1 mm, while the rest of the palm is covered by electronic skin with a force-sensing range of 0.1–20 N and spatial resolution of 5 mm ^[2].

From 'touching' to 'remembering': writing contact into the control loop

Beyond hardware, Sharpa puts its emphasis on the loop that follows contact. Its CraftNet model has three layers: System 2 handles task understanding and long-horizon planning, System 1 handles pre-contact motion planning, and System 0 enters the contact phase, processing tactile feedback and fine motion at higher frequency and feeding state information back to System 1, so upper-layer actions can readjust when execution fails or the physical state changes ^[2]. The AE01 data glove uses 22 encoders to precisely capture the operator's natural hand movements and map them to the robot in real time, while feeding the robot's tactile state back to the operator, forming a bidirectional interface that puts the human motion trajectory, the robot's contact feedback and the mapping between them into a single capture process ^[2]. The latest research, WM-Craftnet, trained a World Synesthesia Model with temporal memory: inputs include wrist depth, touch, proprioception and previously executed actions, compressed through a Dreamer-style recurrent state-space model into time-varying latent states that then serve as policy context for judging the object's current geometry, contact and motion state — what the model learns is not just 'what I'm touching now' but, combined with the past, 'why it has become this way now' ^[2].

The common thread: from demo performance to deliverable capability

Put the three threads together and the common trend is that AI's competitive focus is shifting from 'demo performance' to 'being called reliably'. Image models are trading Arena scores, native 4K and half the price for a place in designers' workflows, while exposing engineering details such as watermark stacking that degrades quality ^[1]; decision models simply abandon conversation, entering software systems through free output and high-frequency background calls, but Microsoft, OpenAI and Cloudflare all arriving within weeks shows the barrier isn't the idea — it's whether you can grab the developer ecosystem first ^[3]; robots are adding touch and force feedback beyond vision, putting speed, payload, precision and contact sensing into one loop and trying to use world models to remember contact experience ^[2]. What's worth watching next: whether image models can balance edit consistency against cost, how much room decision models have to differentiate under pressure from several giants, and whether the tactile data collected by dexterous manipulation can truly settle into reusable capability.