In ten days at the turn of the quarter, the cost of making media with a model stopped being a line item worth arguing about. On June 30 Google shipped two models that put image generation at roughly three cents per thousand and editable video at about a dollar per ten-second clip. On July 8 OpenAI demoed a voice model that listens and talks at the same time. Three headlines, one question for anyone building: which of these can I actually put in a product, and which is still a keynote?

The answer splits cleanly. Two of the three have an API and a price. One has neither. That line — not the demos — is the thing to plan around.

Images: a rounding error now#

Google's Nano Banana 2 Lite (API name gemini-3.1-flash-lite-image) generates images at about $0.034 per 1,000 on the standard paid tier — and roughly half that, ~$0.017 per 1,000, on the batch tier — at about four seconds per image. Google's own framing puts it around 2.7× faster than Gemini 3.1 Flash Image, which is the number that matters if you're generating inside a request rather than overnight.

At those prices, the calculus flips. Per-image cost stops being a reason not to generate thumbnails, variations, avatars, or on-the-fly illustration; the constraint moves back to product design and moderation, where it belongs. If you were caching aggressively or gating image generation behind a paywall purely on cost, that assumption is now stale.

Video: cheap and editable#

The more interesting release is Gemini Omni Flash (gemini-omni-flash-preview), at about $0.10 per second of output — roughly $1 for a 10-second clip. Video generation crossing under a dollar a clip is notable on its own, but the price isn't the whole story.

Omni Flash does conversational, multi-turn editing — you refine a clip in plain English across a session that keeps history, instead of re-prompting from a blank slate.

That's the difference between a one-shot novelty and something you can wire into a product. Iterative, session-based editing means a user (or an agent) can say "make it slower, drop the caption, warmer light" and get a coherent revision — the workflow a real editing surface needs. It's a preview, so treat the numbers and limits as provisional, but the shape is buildable today.

Voice: the catch#

Then there's GPT-Live, OpenAI's July 8 headliner. It's genuinely a step change: a full-duplex architecture, meaning it listens and speaks simultaneously rather than taking turns. No more waiting for the model to detect that you stopped talking; you can interrupt, and it can backchannel, the way people actually converse. It ships as GPT-Live-1 (paid) and GPT-Live-1 mini (free), with a backend reasoning model picking up the hard tasks.

Here's the part the launch coverage buries: at launch, GPT-Live lives only inside ChatGPT. There is no announced API pricing, and the developer API is waitlist-only. You can experience full-duplex voice this week. You cannot build your product on it this week. For planning purposes, that makes GPT-Live a signal about where voice is going — not a dependency you can take.

If you need real-time voice you can ship, the buildable option today is Cartesia, whose Sonic-3.5 (TTS) and Ink-2 (STT) landed June 16 with sub-90ms time-to-first-audio and native turn detection — fast enough for live conversation, and behind a real API. It's not simultaneous listen-and-speak, but it's the difference between a roadmap and a running feature. (For how the buildable voice stacks actually stack up, see Cartesia vs ElevenLabs vs Kokoro, and for what full-duplex changes once GPT-Live opens up, our companion decision piece.)

What a founder should actually do#

The pattern across all three is the same discipline: design around what has an API, not what has a keynote.

The through-line of 2026 keeps repeating: capability arrives at a demo, and buildability arrives some weeks later behind an API and a price. The founders who win the gap are the ones who can tell the two apart on the day of the announcement.