A Swiss lab says it has built a European AI model near the frontier, and a smaller one that beats Google's Gemma 4 31B on a single GPU. On October 27, 2026, Giotto put two models, Alto and Solo, behind a public app and published benchmark tables comparing them with models from OpenAI, Anthropic, Google, Moonshot AI and Mistral.
The tables are striking, including a 100% AIME 2026 score for Alto. They are also entirely the company's own. This post lays out what Giotto claims, what is confirmed (a launch happened, the company and its partners exist), what is not (the scores), and how to test the claims yourself. The company's earlier June 16, 2026 press release gives the strategic context.
TL;DR: claimed vs verified
| Item | Status |
|---|---|
| Giotto, Lausanne-based, launched Alto and Solo on Oct. 27 | Shown on the company site and its X account; consistent with its June release |
| Alto hosted on European cloud, free in app stores | Company claim; not independently tested by us |
| Solo runs on a single GPU | Company claim; the GPU is not named on the page we read |
| Benchmark scores vs GPT, Claude, Gemini, Kimi, Mistral | Company-published only; no third-party reproduction found |
| Open weights | None found |
| Pricing for self-host or appliance | Not published |
What did Giotto actually announce?
Giotto calls itself "The Smartest Sovereign AI," with the tagline "Intelligence You Control. Hosted in Europe." It sells a reasoning model plus the "operating system" around it: assistant, documents, code, integrations, agents and infrastructure monitoring. Per its FAQ there are three deployment paths: Giotto cloud in a dedicated European environment, a download to run on your own GPUs, or a certified appliance with Giotto preinstalled.
Alto is the flagship, "our most powerful model yet." Solo is "the smartest model running on a single GPU." Per Giotto's announcement posts, Alto is free to the public via Google Play and the App Store, while Solo is aimed at organizations' own infrastructure. The giotto.ai page itself links an App Store download and a "Use Giotto for Free" sign-up but does not spell out pricing.
Context matters. In the June 2026 release, CEO Aldo Podestà framed the product as an alternative to "dependency on foreign platforms," citing recent restrictions on access to advanced AI systems such as Mythos and Fable. The release said access would go first to institutional partners and public authorities and to individuals from July. Other announcements surfaced in search include a cooperation with RUAG AG (April 2026) and a Groupe Mutuel collaboration (September 16, 2026), both via press releases.
The benchmark tables, and how to read them
Scores below are as printed on giotto.ai, where each value precedes its model label. The page marks rival effort settings in brackets.
Alto
| Benchmark | Alto | Rivals listed by Giotto |
|---|---|---|
| AIME 2026 | 100.0% | GPT 5.5 (xhigh) 100.0%; Claude Opus 4.8 (max) 100.0%; Gemini 3.1 Pro (preview) 98.3%; Kimi K3 96.7%; Kolibri-1 (en) 96.0% |
| GPQA-Diamond | 91.9% | GPT 6 Astra (max) 96.1%; Claude Fable 5.1 (max) 93.7%; Gemini 3.8 Flash (high) 95.3%; Kimi K3 (max) 93.5%; Kolibri-1 (en) 84.3% |
| HLE | 52.0% | GPT 6 Astra (max) 54.7%; Claude Fable 5.1 (max) 59.1%; Gemini 3.8 Flash (high) 47.8%; Kimi K3 (max) 46.9%; Mistral Large 4 35.0%; Kolibri-1 (en) 21.5% |
| Math-500 | 99.80% | GPT 5 (high) 99.40%; Claude Sonnet 4 99.10% |
Solo
| Benchmark | Solo | Rivals listed by Giotto |
|---|---|---|
| AIME 2025 | 100.0% | GPT-OSS-120B 97.9%; Nemotron 3 Nano 30B-A3B 99.2%; Ministral 3 14B 85.0% |
| AIME 2026 | 96.6% | Gemma 4 31B 89.2% |
| GPQA-Diamond | 88.3% | Gemma 4 31B 84.3%; GPT-OSS-120B 80.1%; R1-Distill-Qwen-32B 62.1%; Ministral 3 14B 71.2% |
| HLE | 37.5% | Gemma 4 31B 26.5%; GPT-OSS-120B 19.0%; Nemotron 3 Nano 15.5%; R1-Distill-Qwen-32B 5.0% |
| Math-500 | 99.80% | R1-Distill-Qwen-32B 94.30% |
Three honest observations:
- Alto does not top every row by Giotto's own numbers. On GPQA-Diamond and HLE the page lists GPT 6 Astra and Claude Fable 5.1 above it, and on AIME it ties. The claim is "near frontier," not "best."
- Comparison settings differ. Rivals appear at "max" or "xhigh" effort, while the page does not state Alto's reasoning budget. Different test-time compute makes cross-vendor rows hard to compare.
- Math saturation. 100% on AIME and 99.8% on Math-500 sit at the ceiling of those benchmarks, where contamination and answer-format effects matter most. HLE and GPQA are the more informative rows.
We treat other vendor tables the same way; our Muse Wonder vs GPT-6.1 Sol vs Opus 5.5 comparison says which numbers come from which source.
What is not known
- No independent evaluation. We searched for third-party results and found none. One early reply on X said the author was looking forward to "more benchmark results."
- No model card or weights. We found no open-weight release, parameter count, context window or training-data statement.
- Solo's hardware. "A single GPU" could mean a 24 GB consumer card or an 80 GB data-center card. The page does not say.
- Pricing and data handling. The FAQ lists questions on cost, hardware ownership and where data sits for each option, but the answers did not appear in what we read.
- Residency guarantees. "Hosted in Europe" is a location claim, not a certification. Ask for contract terms and the hosting provider.
Why the sovereign pitch is landing now
The timing is part of the story. Giotto's June release tied its launch to a perception that access to frontier systems "can no longer be taken for granted." Whatever one thinks of that framing, it matches a wider European procurement mood: governments and regulated firms want a model they can run on infrastructure they control, with a contract that names where data lives. We have tracked the same impulse in Sweden's government offices blocking ChatGPT and Claude on work computers and in Ecosia dropping Mistral's open-weight models, which shows that even sovereign-leaning buyers shop on cost and quality, not only on flag.
For a buyer, that creates a specific test. A sovereign model has to be good enough that choosing it is not a sacrifice. That is why the benchmark rows matter more than the branding: if Solo really does beat Gemma 4 31B on HLE and GPQA while fitting on one GPU, it would be a credible private-deployment option for a hospital, insurer or ministry. If the gap shrinks or reverses on your own tasks, the sovereignty benefit has to carry the decision on its own.
What builders should do this week
- Do not migrate anything yet. One launch day and a vendor table is not enough to switch a production workload.
- Run a bake-off. Pick 30 to 50 real tasks from your own logs, remove anything sensitive, and score Alto, your current model and one open model blind.
- Test the boring parts. Latency, rate limits, context handling, tool calling and uptime decide whether a model is usable, and none appear in the tables.
- Ask for the paperwork. Request the data processing terms, sub-processor list, hosting region and retention policy in writing, plus the appliance and self-host licence if you plan to run Solo on your own hardware.
- Re-check in two weeks. If independent leaderboards or community tests appear, compare them with Giotto's own rows. Large gaps are informative either way.
How to check the claims yourself
- Install the free app and run a held-out set of your own prompts rather than public benchmark questions.
- Compare against an open model you can run locally on the same prompts, with matched effort settings.
- For Solo, ask Giotto for the GPU spec, quantization and context length, then measure tokens per second.
- Watch independent leaderboards: our coverage of the Artificial Analysis Intelligence Index v4.2 shows how third parties typically follow launches.
- For data residency, ask where inference runs, whether prompts are logged and which sub-processors are involved.
Where it fits among European models
Giotto joins a crowded European field. Aleph Alpha's Kolibri is an open-weight sovereign MoE and appears, as "Kolibri-1 (en)," to be one of Giotto's comparison rows. Switzerland's own Apertus 1.5 technical report describes fully open models, and Mistral Large 4 is an open-weights preview. Giotto's difference is the packaged operating system and the single-GPU pitch, not openness. For what "sovereign" does and does not guarantee, read what is sovereign AI.
The buyer trade-off is simple. Open weights let you verify and self-host but you assemble the stack. A closed sovereign product offers integration and support, but you rely on the vendor's claims until audits and independent benchmarks arrive.
Bottom line
The launch is real; the performance claims are unproven. A lab publishing its own tables is normal, and some rows, such as Solo's HLE versus Gemma 4 31B, would be notable if reproduced. Until independent evaluators or users confirm them, treat Alto as a candidate to test and Solo as a claim to measure. We will update this post when third-party results appear.
Related reading
- What is sovereign AI? Layers, models and trade-offs
- Aleph Alpha Kolibri: sovereign open-weight MoE
- Apertus 1.5 technical report: Swiss open models
- Mistral Large 4 open-weights preview
- Muse Wonder vs GPT-6.1 Sol vs Opus 5.5 comparison
- Artificial Analysis Intelligence Index v4.2
Details are accurate as of October 27, 2026, based on giotto.ai and Giotto's June 16, 2026 press release; benchmark figures are the company's own.