What actually launched
Strip the sovereignty language out of the release and the offer is concrete. Solo is a reasoning model sized for one GPU, aimed at a workstation. Alto is the flagship for heavier work, and it needs eight B300s, which is not a workstation at all. The same family covers both ends by changing the hardware bill, not by changing the sales pitch.
Giotto offers free access to Alto through its European-hosted shared cloud, including anonymous chat without an account. API access follows a different commercial model: a flat fee for unlimited usage at one concurrent request, with additional token packages for greater concurrency. Dedicated European cloud instances are rented monthly, while on-premises installations are licensed according to GPU count and type. The release does not publish currency-denominated prices for those paid arrangements.
There is also a free anonymous chat that needs no account, plus new iOS and Android apps. Free, anonymous access gives prospective users a way to test Alto against their own prompts without accepting the company's benchmark claims. It also makes the product easier to evaluate before committing to paid infrastructure.
The scores are Giotto's own
Giotto's homepage carries the comparison table the release leaves out. Alto is listed at 100.0% on AIME 2026, 91.9% on GPQA-Diamond, 52.0% on HLE and 99.80% on Math-500. Solo, on one GPU, is listed at 88.3% on GPQA-Diamond and 37.5% on HLE. In Giotto's own comparison table, Solo leads the listed single-GPU competitors on GPQA-Diamond and HLE.
Giotto's table does not show Alto winning every comparison. It trails the highest-listed competitor on GPQA-Diamond and HLE. That is useful context, but it does not independently validate the scores. Giotto has not published the evaluation harness, prompt settings, sampling methodology or third-party results needed to reproduce the table. The numbers describe Giotto's reported performance, not an independently established ranking.
What is missing is the model
For a launch built on the word control, the omissions are specific. Neither the launch release nor the product pages reviewed disclose parameter counts, context windows or model cards. There are no open weights. Downloading Giotto means a licensed private deployment on your hardware, which is a real form of control over where data sits, and a much weaker form of control over what the model is. You can choose the building. You cannot inspect the tenant.
That distinction is the practical test for buyers. A European company that needs data to stay in Europe, on infrastructure it can point to, gets something real here today, with free shared-cloud access available for initial testing, while dedicated and on-premises deployments carry separate commercial terms. A team that needs to audit the model itself, reproduce the benchmark table, or plan around a known context window gets none of that from this launch. Giotto says its ARC-AGI entries placed 11th in 2024 and second in 2025 among roughly 1,500 teams, which is a genuine research signal, and it is still the company's own account of itself.
The offer will be easy to check. Alto is free and anonymous, so its everyday quality is now a public experiment anyone can run. What cannot be checked from outside is whether Solo really is, as Giotto puts it, the smartest model running on a single GPU. That claim rests on a table one company wrote, comparing models it selected, at sizes it has not disclosed for its own entry.
Sources
- [1] Giotto.ai SA media release, "Giotto launches sovereign AI family to make AI public infrastructure," RNS Reach / EQS News ID 2421676, October 27, 2026, 12:00 CETRead source
- [2] Giotto official site, benchmark tables for Alto and Solo and deployment options, accessed October 27, 2026Read source