For years, AI followed one simple rule: a smarter model costs more money, and better video takes more time. Over the past two weeks, that rule broke on seven fronts at once – with tools that are stronger, faster and far cheaper to run.

The old rule: more power costs more
Big players and small start-ups showed tools over the past fortnight that are stronger, more precise and faster – while costing far less to run.
Here are the seven most important releases from November 2026, with a practical take on each for Swiss SMEs.
1. GPT-6.2 Astra and Sol: better results, far lower cost per task
On 17 November 2026, OpenAI introduced two new models: GPT-6.2 Astra and GPT-6.2 Sol. They are not in the free chat. They are available in ChatGPT Work and Codex on the Plus, Pro, Enterprise and Edu plans. Developers can use them through the API. [1]
- Astra handles the hardest engineering and analysis work.
- Sol is the efficient workhorse for everyday tasks.

Astra in numbers (OpenAI's figures)
- Agents' Last Exam tests complex, multi-step professional work. Astra scores 5.2 points higher than GPT-6 Astra there. It does the same work for less than a third of the cost.
- API prices drop by 20%: USD 8 per million input tokens and USD 40 per million output tokens.
- Against Claude Fable 5.2, OpenAI reports a 5.5-point lead on the same benchmark at about 5% of the cost. For context: the Fable results leave out nine tasks that ran on a fallback model.
Sol in numbers (OpenAI's figures)
- On LifeVal, OpenAI's test for demanding everyday tasks, Sol beats the former flagship GPT-6 Astra. It costs about a seventh as much.
- In coding (FrontierCode 2.0), Sol trails the new Astra by just 0.5 points. Per task, it costs about a fifth as much.
- API price: USD 2 per million input tokens and USD 10 per million output tokens.
- Pro 200 subscribers get the “Ultrafast” mode for the first time, starting with Sol.
Better documents: Astra was trained to produce cleaner slides, spreadsheets and documents. It follows the conventions of law, sales, finance and accounting. The goal is less generic text and less of the typical “AI look”. [1]
Cybersecurity: According to OpenAI, both models make a significant jump in cyber capabilities. They reach the “Critical” level in OpenAI's own Preparedness Framework. OpenAI has tightened its safeguards as a result, which also limits some legitimate security research. [1]
Why it matters: Lower prices plus higher capability shift the balance of the market. Advanced automation is no longer a privilege of big corporations; it becomes an affordable everyday work tool. One caveat: all figures come from OpenAI. Independent tests are still pending.
2. Voice models that pick up sarcasm and emotion
Open models for real-time speech are moving fast. The line between a synthetic voice and natural human expression is blurring.

ZONOS3-Mini (Zyphra): The model has 800 million parameters, but only about 100 million are active at each step. The weights take up about 1.8 GB. ZONOS3-Mini clones a voice from a short sample. Zyphra says you can steer age, accent, emotion and pace down to single words. It supports 43 languages, including German and Polish. It runs locally, even right in the browser. The licence is Apache 2.0. [2]
Haru-4B (higurashi, Tokyo): Haru listens and speaks at the same time, like a phone call. Its developer reports a response time of about 65 milliseconds on a MacBook. Haru turns audio straight into audio, skipping the detour through text so tone and emotion don't get lost. German is one of its nine primary languages. The licence is Apache 2.0. Note: Haru is a base model for research, not a finished assistant. [3]
Sophon-Realtime: This open model also works in full duplex – it keeps listening while it talks. It pays attention not only to what you say, but how you say it. It recognises emotion, emphasis, irony, laughter and sighs, and laughs or sighs itself when that fits. [4]
- Someone sighs and says gloomily, “Everything's fine.” The model hears the tension and asks whether they'd like to talk about it.
- Someone says sarcastically, “I love Mondays.” The model catches the irony and laughs.
3. Real-time video: Apple breaks the speed barrier
OpenAI cuts costs. Apple goes after the biggest pain point of video generators: speed. The result is FastFrames, built together with Cornell University. [5]
The researchers took the pipeline apart and found a surprising bottleneck. More than half the time did not go into the frame content itself, but into support work: decoding the finished image and constantly refreshing the model's working memory.
- A 480p video frame takes under 50 ms.
- On an Nvidia B200, the small model needs just 36 ms per frame.
- For camera-controlled generation, the process runs up to 16 times faster.
In practice, you can steer a generated video world live with your arrow keys or mouse. One note for business use: the weights are under an Apple research licence, not a permissive commercial one. [5]
4. Compute where it counts: 3D and video
Why should an AI spend power and compute on a blank wall? Classic generators treat every pixel the same. The boring wall costs as much as the main character's face. A new wave of tools allocates resources on purpose.

Budget3D turns a single photo into a 3D object. You choose a budget of compute units (tokens). The model merges low-information regions into coarser units, starting with empty space. The compute goes into the details. Generation gets faster and needs less memory. A turbo version needs only 5 steps. Budget3D is an open-source community project. [6]
MPEG-Diffusion applies the same idea to video. Important parts of the frame are generated in high resolution, the rest in lower resolution, and the system blends both layers seamlessly. At 960p, the paper reports video 4.8 times faster. It stays sharp where the human eye actually looks. Compared with full resolution, “no difference” was the most common answer from test viewers. [7]
5. ReShoot 4D: change the camera move after the shoot
ReShoot 4D moves key camera decisions from the set to the edit. The system analyses an existing video and rebuilds the person and the surroundings as a 3D scene, frame by frame, with smooth motion. [8]
You then plan a new camera path in a virtual editor – such as an orbit around the actor, a dynamic close-up or a new angle – and can even adjust depth of field and aperture. A connected video model renders the new shot. ReShoot 4D is free and open source. For video production, a bad framing or missing shot can often be fixed in the edit without a reshoot. [8]
6. AI against disease: NovaAtom-Lite and AlphaGenome 2
Better algorithms are driving big progress in drug research, too.

NovaAtom-Lite (Geodesic Intelligence): Geodesic Intelligence is only about five months old, founded by Quanquan Gu, formerly an AI scientist at ByteDance. NovaAtom-Lite predicts the 3D structure of molecular complexes, such as how an antibody binds to a protein. On the strict FoldBench benchmark, Gu says the model got 79.1% of these interfaces right, compared with 47.9% for AlphaFold 3. These figures come from the company itself. [9]
AlphaGenome 2 (Google DeepMind): The model predicts how single changes in DNA affect specific cell types. In a test across 28 immune cell types, AlphaGenome 2 found 30% of the causal variants that change gene activity, nearly doubling the 16% found by the first version. AlphaGenome 2 is a research tool, not a diagnostic system. [10]
7. Robots learn from us
For decades, people had to learn how to operate machines. They memorised shortcuts and fought complex interfaces. Now machines are learning from us:
- Flash ALOHA (Stanford): A two-armed robot folds shorts, clears a table or catches a ball after just 50 human demonstrations per task. It succeeds more than 80% of the time at close to human speed. The project is open source. [11]
- Nvidia SOMA Retargeter: This open-source tool maps recorded human motion onto the proportions and joints of a humanoid robot, currently supporting the Unitree G1. [12]
What does this mean for Swiss SMEs?
| Area | Model / Tool | Key gain (vendor figure) | Licence & SME use |
|---|---|---|---|
| Text, code & docs | GPT-6.2 Astra & Sol | USD 2 / 10 (Sol) or USD 8 / 40 (Astra); < 1/3 cost per task | Commercial via API, ChatGPT Work & Codex |
| Local voice AI | ZONOS3-Mini / Haru-4B / Sophon | In-browser (~1.8 GB) or ~65 ms full duplex & irony | Apache 2.0 (ZONOS3-Mini, Haru-4B); voice consent required |
| Real-time video | Apple FastFrames | < 50 ms per 480p frame, up to 16× faster | Apple research licence (not freely commercial) |
| 3D & video editing | Budget3D, MPEG-Diffusion, ReShoot 4D | 5-step 3D, 4.8× faster 960p video, camera moves in edit | Budget3D & ReShoot 4D open source |
| Medicine & research | NovaAtom-Lite & AlphaGenome 2 | 79.1% FoldBench; 30% causal DNA variants | Research tools only, not diagnostic systems |
| Robotics | Flash ALOHA & Nvidia SOMA | > 80% success after 50 demos; motion retargeting | Open source (research & prototyping) |
Automation becomes affordable: If a top model does the same task for less than a third of the cost, new workflows start to pay off – such as preparing quotes, checking documents or summarising data. What was too expensive a year ago may make sense today. Run the numbers again.
Voice AI runs locally: Models like ZONOS3-Mini run in the browser or on your own device so audio never has to leave the device, simplifying Swiss FADP data protection. When cloning a voice, though, one rule applies: only with the person's consent.
Read the licence: Apache 2.0 allows commercial use. FastFrames is under a research licence. Check the small print before you use a model in your business.
Vendor numbers are a starting point, not proof: Test new models on your own tasks before you switch.
Final thought: who learns faster?
These two weeks show a deep shift. For a long time, we adapted to machines. Now machines adapt to us. They get faster, cheaper and better at understanding us.
That leaves one question: in this new relationship, who will learn faster – the human or the machine?
Assess where AI genuinely pays off in your business
Whether it is a controlled AI assistant (RAG) over your own documents or visibility on Google and in AI answers: we separate hype from measurable value. The full Website Analysis costs CHF 2'450 once; ongoing SEO and GEO support starts from CHF 1'500 per month with at least 15 documented working hours per month in app.glasbox.ch.
Book an introductory call