Cold Open — Google ships Nano Banana 2 Lite and Omni Flash to builders

🎧 This is the print twin of today's Cold Open episode. Listen to today's episode.

Wednesday, July 1, 2026. We scanned 2,778 items off the overnight wire; three made the front page — and the one on top is the rare launch you can wire into a pipeline before lunch.

The lead · Google ships Nano Banana 2 Lite and Omni Flash to builders

Google DeepMind put two generative-media models into developers' hands today, and the theme is the same for both: speed and cost, knocked down far enough to build with. Nano Banana 2 Lite (gemini-3.1-flash-lite-image) is the new floor of the Nano Banana family — text-to-image in about 4 seconds at $0.034 per 1K-resolution image, positioned as a drop-in swap for the original Nano Banana. Alongside it, Gemini Omni Flash (gemini-omni-flash-preview) comes to developers for the first time: video generation and conversational editing from text, image, and video inputs, priced at $0.10 per second of output — the same as Veo 3.1 Fast.

"Nano Banana 2 Lite is our fastest, most cost-efficient image model in the Nano Banana family yet, built for high throughput, speed and scale." — Google DeepMind

Both are available today in Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform, with Nano Banana 2 Lite also rolling out across consumer surfaces including AI Mode in Search, the Gemini app, NotebookLM, and Google Photos. The launch also clarifies the family ladder: Lite for speed, Nano Banana 2 as the generalist workhorse, and Nano Banana Pro (Gemini 3 Pro Image) for the jobs where reasoning and control matter more than latency.

Why it matters

This is a builder's launch, not a benchmark flex. When an image model returns in ~4 seconds at three-and-a-half cents a shot, it stops being a special-occasion API call and starts being something you put inside a loop — draft thumbnails, generate placeholder art, iterate on a layout a hundred times without watching a budget. Omni Flash does the same thing for video at 10 cents a second: a full minute of generated, conversationally editable video for about six dollars. The constraint being removed here is not capability; it is the latency and cost that kept generative media out of high-volume pipelines. The practical move is to pick your point on the ladder — Lite when you are iterating, Pro when the output ships.

The fine print

Two caveats before you re-plan around it. First, the quality-versus-speed-versus-cost comparisons are Google's own Elo charts against competitors — read them as vendor numbers until someone independent runs the same prompts. Second, "Lite" earns its low latency by trading away some quality, and Omni Flash is still a preview; the consumer-surface rollout is gradual, so what you can call from the API today is ahead of what every product surface exposes.

Sources: deepmind.google

02 · Anthropic's Claude Science pulls agents into the lab

Anthropic announced Claude Science, an AI workbench for research that lets scientists converse with agents in natural language to run their work end to end. The builder-relevant hook is how it plugs into hardware: Claude Science integrates the NVIDIA BioNeMo Agent Toolkit, which — in NVIDIA's own words — "packages NVIDIA-accelerated capabilities as callable skills, enabling Claude Science to select the appropriate tool, prepare valid inputs and execute the workflow" against NVIDIA compute deployed anywhere.

Why it matters. Watch the phrase "callable skills." The agent-skills pattern that started in coding tools is now the interface for wrapping domain compute — GPU-accelerated models and microservices exposed as things an agent can pick up and run inside a scientist's natural-language session. That is the same skills abstraction leaving the IDE and showing up in the wet lab. Caveat: this is a partner announcement, so treat the workflow claims as a launch narrative until independent labs report real-world speed-ups.

Sources: blogs.nvidia.com · anthropic.com

03 · OpenAI used core-dump "epidemiology" to kill an 18-year-old bug

OpenAI's engineering team published how it debugged rare crashes in Rockset — the C++ real-time data system OpenAI acquired in 2024 — by treating core dumps like an epidemiologist treats case data. Instead of inspecting a few crashes by hand, they "had ChatGPT write a script that downloaded a prefix of each core file, extracted the registers, filtered known false positives, and automatically labeled the crash," then ran it across every production core dump from the previous year. Clean data made the correlations pop: what looked like one weird bug was two separate populations — a software return-to-null bug spread across regions, and misaligned-stack crashes that all traced back to a single physical machine with bad hardware. The software half ran deep: C++ exception unwinding in their binary resolved to GNU libunwind rather than the libgcc implementation they expected, a plumbing detail roughly 18 years in the making.

Why it matters. This is the AI-in-the-debugging-loop story with a real payload — an agent writing the analysis pipeline that turns an intractable pile of crashes into a labeled dataset a human can reason over. It is also a quiet reminder that the frontier stack rides on decades-old systems plumbing, and that "look at it at population scale" beats "stare harder at one core dump."

Sources: openai.com

Also on the radar

• Benchmarks — GeneBench-Pro: OpenAI's 129-question, research-level computational-biology benchmark; its strongest model, GPT-5.6 Sol, tops out at 28.7% (31.5% with Pro mode) — up from below 5% when the original GeneBench began.
• Adoption — How ChatGPT adoption has expanded: new OpenAI Signals data on usage growing across regions, languages, and capabilities.
• Thesis — Why Specialization Is Inevitable: the case that the model landscape fragments into specialists rather than converging on one generalist.
• Infra — NVIDIA GQE: GPU-accelerated query engines aimed at the memory- and I/O-bandwidth walls that bottleneck data systems.

Trends in dev tools

What moved in the tools engineers actually ship with.

• From copilots to "software factories." On Latent Space, Warp CEO Zach Lloyd argues software factories are the next phase of coding — the frame shifting from a single assistant in your editor to a managed fleet of agents producing software as a pipeline.
• Agents get benchmarked on legacy migration. IBM Research's ScarfBench measures AI agents on enterprise Java framework migration — evals meeting the unglamorous, high-value work of modernizing old code, which is where a lot of real budget lives.
• Orchestration is becoming its own benchmark category. ClawArena-Team benchmarks subagent orchestration and dynamic workflows in language-model agents — a sign that "how well do the agents coordinate" is now a measurable axis, not just "how good is the model."
• Cost-per-token is the new infra metric. NVIDIA makes the case that inference economics now hinge on token cost, not peak chip specs — useful vocabulary if you are the one defending an AI feature's unit economics.

Popular skills

The agent-skills wave — packaging expertise an agent loads on demand — kept spreading and, notably, kept getting formalized.

• Skill descriptions are load-bearing. A Single Rewrite Suffices shares empirical lessons from production skill descriptions — evidence that one careful rewrite of the words in a skill file can meaningfully change how well an agent uses it.
• Skills move into vision AI. NVIDIA's Into the Omniverse walks three workflows for improving vision-AI agent accuracy with synthetic data and fine-tuning — the skills pattern applied to agents that read the physical world.
• Skills are getting an identity. The Decomposition Is the Fingerprint proposes per-component identity for agent skills — provenance and verification for the reusable folders agents pass around, which is what you want before you trust a skill you didn't write.

AI fun fact

To show off Gemini Omni Flash, Google's launch demo has someone perform four "digital magic tricks" — pulling a 3D balloon word out of her phone, pouring water from the screen into a glass — with a small inset clip in the corner revealing the plain footage the model actually dressed up. It is a neat, honest way to demo generative video: the trick and the reveal, in the same frame. (deepmind.google)

Sources: deepmind.google · blogs.nvidia.com · anthropic.com · openai.com · huggingface.co · latent.space