Giotto launches Alto and Solo, putting deployment control at the center

Swiss AI developer Giotto announced Alto and Solo on October 27, pairing new reasoning models with a pitch for more control over where AI runs. The launch offers European cloud access and commercial private-deployment options, rather than asking customers to treat a remote chatbot as their only route to the technology.

The practical question is not just how the models score. It is whether an organization can move useful AI workloads to infrastructure it controls without taking on an impractical hardware or support burden. Giotto’s launch makes that an explicit part of the product proposition, but its published performance figures and deployment claims still require customer testing.

Two models, different infrastructure targets

In its October 27 announcement, the Lausanne company says Solo runs on one GPU while flagship Alto runs on eight NVIDIA B300 GPUs. It describes free shared-cloud chat, dedicated cloud environments in Switzerland or elsewhere in Europe, and licensed installations on compatible customer hardware.

Those are different access arrangements, not one universally free product. The release describes monthly rental for dedicated infrastructure and annual or multiyear licensing for on-premises deployments. The public website also provides an Alto API entry point. PromptsZone has verified these published access paths, but has not run the models or validated their throughput.

Giotto’s current offering page separates managed European cloud, deployment on private GPUs, and ready-to-use workstations or servers with support. That distinction matters: operating a model locally means owning more of the capacity planning, updates and reliability work. A single-GPU claim is not a promise that any consumer graphics card meets the requirements.

A scorecard, not an independent league table

The company’s benchmark tables report Alto at 100.0% on AIME 2026, 91.9% on GPQA-Diamond, 52.0% on Humanity’s Last Exam and 99.80% on Math-500. Solo is listed at 96.6%, 88.3%, 37.5% and 99.80% on those respective tests.

These are developer-reported results. The tables do not provide a complete evaluation recipe alongside the numbers, including all settings needed to establish a controlled comparison. Our accompanying scorecard reproduces the published figures without converting them into an overall intelligence score or a claim of equal-compute superiority. A percentage on one test should not be averaged casually with a percentage on another.

For a deployment decision, the missing bridge is task-level evidence. An organization should compare exact model configurations on representative work, measure failure rates as well as successful answers, and include latency and operating costs. A strong benchmark score can justify an evaluation; it cannot replace one.

Check the commercial terms behind the access claim

The API page, checked October 29, lists Alto at €2.10 per million input tokens and €10.50 per million output tokens, with separate cache rates. That differs in presentation from the launch release’s description of flat-fee shared-cloud API access with concurrency limits. The pages do not establish that every API customer receives unlimited use, so buyers should confirm the plan and billing terms that apply to their account.

Likewise, Giotto’s stated goal of user-controlled AI should be distinguished from the contractual rights a customer actually receives. Local deployment, access to model weights, permission to modify them and freedom to redistribute them are separate questions. The materials reviewed here do not establish an unrestricted open-source release.

The useful development is another deployment choice from a European model developer. Its value will depend on reproducible performance, clear licensing and the ability to keep the promised control when a real workload moves from a trial into production.

Graphic: PromptsZone, rendered from Giotto’s published benchmark data, checked October 29, 2026. Original scorecard; developer-reported results, not independently replicated by PromptsZone.