Latest / UpNext AI / Avataar’s Low-Cost Video AI, OpenAI’s Ona Deal, and Verifiable Science Agents | UpNext AI – June 12, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Friday, June 12th, 2026, and here's what matters in AI today. We start in India, where TechCrunch reports that Avataar AI has launched a new video model called Varya, aimed squarely at a market where cost and local relevance may matter as much as raw model quality. The headline number is price. Avataar says its distilled model will cost about half a cent per second of video generation, or $0.005 per second on its hosted service. TechCrunch says that compares with roughly ten cents or more per second for other video models, which would make this about a twenty-fold gap. The company says Varya is built for India’s scale and for local context, including things like festivals, food, clothing, and architecture. According to TechCrunch, Avataar started from Alibaba’s Wan 2.2 and used distillation to compress it into a leaner model tuned for its own use cases. The other big point here is speed. TechCrunch reports that Varya runs in four steps instead of fifty for Wan 2.2, and on an NVIDIA H200 GPU it can generate a five-second 720p clip in 45 seconds, compared with 1,230 seconds for Wan 2.2. That matters because this is really a story about AI adoption constraints. TechCrunch ties the launch to India’s broader AI push, including the India AI Mission, a roughly $1.2 billion initiative that gives selected startups access to subsidized compute if they release models publicly. Avataar was one of the startups selected. The same report says India is aiming to attract $200 billion in AI investment by 2028. So the takeaway is straightforward: Avataar is betting that in a video-first market, the unlock is not just model capability. It’s getting video generation cheap, fast, and culturally specific enough to be used at population scale. From there, to enterprise infrastructure. OpenAI says it plans to acquire Ona in order to expand Codex with secure, persistent cloud environments, enabling long-running AI agents across enterprise workflows. That wording is doing a lot of work. The important part is persistent environments. Instead of an agent handling a short task and disappearing, the idea is that it can keep state, stay inside a secure environment, and work across longer business processes. Based on OpenAI’s announcement, this is less about a flashy consumer feature and more about the plumbing required to make agents useful inside real companies. If that works, it could make Codex fit more naturally into enterprise workflows where tasks are not one-shot requests but ongoing chains of work. We do not have deal terms in the provided materials, and this is coming from OpenAI’s own announcement. But strategically, the direction is clear: OpenAI is pushing beyond chat and short task completion toward durable agent infrastructure for enterprise use. For the research section, a paper from earlier today introduces EpiBench, a verifiable benchmark for AI agents working on epigenomics analysis. The reason this stands out is that a lot of agent benchmarks still depend on fuzzy judgments about whether a system was helpful. EpiBench instead tests whether an agent can make well-defined analysis decisions from realistic workflow states and return deterministic outputs that can actually be checked. And the paper includes real performance data. The benchmark covers 106 evaluations across several epigenomics workflows, and across 5,088 valid trajectories from 16 model-harness pairs, no system passed a majority of attempts. The best reported result was GPT-5.5 with Pi at 45 percent, followed by GPT-5.5 with OpenAI Codex at 39.9 percent. The researchers also say many failed runs still included parts of the correct answer. In other words, the agents often found the right files or computed useful intermediate results, but still broke down when the task required deeper, assay-specific scientific judgment. If epigenomics is unfamiliar, think of it as studying chemical markers on DNA and related material that affect how genes are used. The bottom line is that verifiable science-agent benchmarks are getting better, but this paper suggests today’s systems are still well short of reliably handling even short-horizon scientific decisions on their own. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ... Anthropic is partnering with TCS to scale enterprise AI deployments. TechCrunch reports that TCS will create a business unit focused on deploying Anthropic’s models to its customers, which is another sign that frontier model companies are leaning harder into large implementation partners. Google has released DiffusionGemma, an open model that, according to The Decoder, generates text from noise instead of producing it word by word. The key idea there is a different generation approach than the usual token-by-token method. Amazon is rolling out a software update for Echo Hub. The Verge reports that the device, which launched in 2024, is getting a more customizable home screen, plus Ring AI video search and Alexa Plus summaries for camera events. And Jeff Bezos’ AI startup Prometheus has reportedly closed a $12 billion funding round at a $41 billion valuation, according to The Decoder. The report also says the company launched last November with $6.2 billion. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. If you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back Monday with what's up next!