Giotto Solo and Alto: Free Sovereign AI From Switzerland, €2.10 API Pricing

Lausanne, Switzerland — Giotto, the Swiss AI lab founded by EPFL mathematics graduates in 2017, launched its Solo and Alto model family on Tuesday, October 27, 2026, and the offer is unusually direct: the flagship Alto is free to try in the cloud, no account required, and the smaller Solo runs on a single GPU you can own. The pitch is not that Giotto beats the frontier labs — Giotto's own benchmark table shows Alto trailing OpenAI's GPT 6 Astra and Anthropic's Claude Fable 5.1 on two of the four tests it publishes. The pitch is that sovereignty is a feature you can buy: European hosting, no US cloud dependency, and a model you can take home. For a lab that placed second out of roughly 1,500 teams at ARC-AGI 2025, that is a coherent bet. But the price list shows it is not a cheap one.

What Giotto actually launched

Giotto Solo and Giotto Alto are two reasoning models wrapped in the same system. Solo is the portable one: Giotto calls it the smartest model that runs on a single GPU, built for a workstation or a small server. Alto is the flagship for demanding reasoning and enterprise workloads, and it runs on eight NVIDIA B300 GPUs. Around both sits Giotto's operating system layer — an assistant, document analysis, code tools, integrations for email and calendars, task-specific agents, and monitoring that shows where workloads run and how infrastructure performs.

There are three ways to get it. Shared cloud, hosted in Europe, is free to use, with a free anonymous chat that needs no account and iOS and Android apps. Dedicated cloud instances, starting at one GPU, are rented monthly and include the Giotto software licence. On-premises deployments are licensed annually or over multiple years, with pricing based on the number and type of GPUs where Giotto is deployed. Giotto says its cloud services are GDPR-aligned, do not depend on US cloud infrastructure, and hold ISO/IEC 27001:2022 certification.

This is not Giotto's first launch. The lab introduced its first portable general-purpose model and operating system in May 2026, opened access to institutional partners and public authorities in June, and expanded to individual users in July. The October 27 announcement turns that single model into a family — Solo for the edge, Alto for the heavy lifting — and puts a free tier in front of everyone at once.

The benchmarks: Giotto's own table, read honestly

Every benchmark number in this section comes from Giotto's own website, and that matters. No independent evaluation of Solo or Alto exists yet, and the comparison set — which rivals appear in each row — was chosen by Giotto. What makes the tables worth reading is that Giotto publishes its losses alongside its wins.

Alto's table puts the flagship against frontier models. On AIME 2026, Alto scores 100.0%, tied with GPT 5.5 at extra-high reasoning effort and Claude Opus 4.8 at maximum effort, ahead of Gemini 3.1 Pro in preview at 98.3% and Kimi K3 at 96.7%. A three-way tie at 100% is also a sign the test is saturated at the top: AIME 2026 no longer separates the best models from each other. On Math-500, Alto scores 99.80%, ahead of GPT 5 at high effort (99.40%) and Claude Sonnet 4 (99.10%).

On the two harder tests, Alto trails. On GPQA-Diamond, Alto scores 91.9% against 96.1% for GPT 6 Astra at maximum effort, 93.7% for Claude Fable 5.1 at maximum effort, and 95.3% for Gemini 3.8 Flash at high effort. On HLE, the expert-reasoning test, Alto scores 52.0% against 59.1% for Claude Fable 5.1 and 54.7% for GPT 6 Astra. So the honest summary of Alto's own table is: best-in-class on maths, behind the frontier on science and expert reasoning, with a 7-point gap on HLE that no marketing copy should be allowed to round away.

Solo's table is the more interesting one, because the single-GPU class is where Giotto's efficiency argument lives or dies. Solo scores 100.0% on AIME 2025, ahead of NVIDIA's Nemotron 3 Nano 30B-A3B at 99.2% and GPT-OSS-120B at 97.9%. On AIME 2026 it scores 96.6% against 89.2% for Google's Gemma 4 31B. On GPQA-Diamond it scores 88.3% against 84.3% for Gemma 4 31B and 80.1% for GPT-OSS-120B. On HLE it scores 37.5% against 26.5% for Gemma 4 31B and 19.0% for GPT-OSS-120B. On Math-500 it scores 99.80%. Within the comparison set Giotto chose, Solo leads its class on every test — and the HLE gap, 11 points over the nearest rival, is the number that makes the "smartest model on a single GPU" claim more than a slogan. It is still Giotto's claim, on Giotto's tests, against Giotto's rivals. Independent testing is the missing piece, and until it exists, treat every figure here as a company-reported score, not a verdict.

What it costs

The free tier is real and simple: shared-cloud chat costs nothing and asks for no account. The paid story starts with the Alto API, and Giotto publishes the rates on its API page in euros. Input costs €2.10 per million tokens, output costs €10.50 per million tokens, a cache hit costs €0.21 per million tokens, and a cache write costs €2.10 per million tokens. Giotto's own currency converter turns those into $2.47 for input, $12.35 for output, $0.25 for a cache hit and $2.47 for a cache write, and CHF 1.89, CHF 9.45, CHF 0.19 and CHF 1.89 respectively.

Two pricing details deserve attention. First, the press release says API access on the shared cloud uses a flat fee for all-you-can-use requests, limited to one concurrent request at a time, with token packages available for multi-concurrency. That flat fee is a genuinely different billing model from per-token pricing — and Giotto publishes no figure for it. Until the number appears, heavy users cannot work out whether the flat fee or the token rate is the better deal. Second, dedicated cloud and on-premises pricing are quote-only: monthly rental for dedicated cloud, annual or multi-year licensing for on-premises, both priced by GPU count and type. There is no published price for either tier.

Against per-token rivals, Alto is not the cheap option. The table below uses rates from our own verified coverage, and Giotto's euro prices converted at the rate on Giotto's own site.

Alto's output price, $12.35, sits above every rival in that table, and its input price, $2.47, sits above the $2.00 standard of Sol, Kimi K3.1 Pro and Sonnet 5.5. The only way Alto wins on price is the unpublished flat fee: if a single concurrent request at a fixed monthly price covers a heavy workload, the per-token comparison stops mattering. Giotto has not published the number that would let anyone test that. For now, the honest way to describe Alto's pricing is: free to try, mid-pack to expensive per token, and potentially cheap for heavy single-stream use — with the last clause unverifiable until Giotto names the fee.

The sovereignty pitch — and where it holds up

Sovereignty is the whole product. Giotto's chief executive, Aldo Podestà, put it bluntly in the launch release: "Europe must offer more than a different flag on the same dependency. Sovereignty has to belong to the people and organisations using AI, and not just to the companies building it." The June access announcement made the same point from the other direction, noting that recent restrictions on entry to advanced AI systems had shown European users that access to frontier AI "can no longer be taken for granted."

That pitch lands differently now than it would have a year ago. Europe's sovereign-AI push is no longer a single announcement: Polarise launched its sovereign AI cloud in the same month, and Giotto itself has been signing the partnerships that make sovereignty concrete rather than rhetorical. In September, Groupe Mutuel — Switzerland's global insurer, serving over 1.3 million clients — began deploying Giotto on its own premises, keeping client data inside the organisation. Giotto has also signed a cooperation agreement with RUAG, the technology partner to the Swiss Armed Forces, to integrate its system into RUAG's LLARA AI platform, and it collaborates with SAP on reasoning capabilities in SAP Joule agents.

But sovereignty has limits, and Giotto's own materials state them plainly if you read them. First, the hardware: Alto runs on eight NVIDIA B300 GPUs, and Solo on a single GPU — NVIDIA is a US company, and the world's most advanced AI accelerators are not European. Hosting in Europe removes the US cloud from the stack; it does not remove the US chip. Second, the software: Giotto publishes no parameter counts, no context-window figure, and no open weights or public licence. This is proprietary software, licensed for deployment — portable in the sense that you can run it on your hardware, not in the sense that you can download the weights and walk away. That is a real distinction from Mistral Large 4, whose open weights were due October 27. Third, the evidence: all benchmark figures are company-reported, and the comparison tables are Giotto's own selection. Sovereignty is a claim about control, and control is verifiable — but capability still needs independent testing.

Pros and cons

Pros: genuinely free to try, with anonymous chat and no account; single-GPU deployment at this capability level is unusual; flat-fee API option could make heavy use predictable; European hosting, GDPR alignment and ISO 27001 certification; an honest benchmark table that publishes losses; ARC-AGI pedigree and real enterprise deployments already running on-premises.

Cons: per-token output price above Sol, Kimi K3.1 Pro and Sonnet 5.5; flat-fee and enterprise prices unpublished; no parameter counts, context window, latency figures or open weights; all benchmarks company-reported against a Giotto-chosen comparison set; Alto still trails the frontier on GPQA-Diamond and HLE; B300 hardware keeps a US dependency in the stack.

FAQ

Is Giotto really free?
Yes, for the shared-cloud chat: free access with no account, plus free iOS and Android apps. The API, dedicated cloud and on-premises deployments are paid. The shared-cloud API uses a flat fee for all-you-can-use requests at one concurrent request, but Giotto has not published the fee.

Can I run Giotto Solo on my own GPU?
Yes — that is the point of Solo. You can download Giotto to run on your own GPUs, buy a certified Giotto appliance with the software preinstalled, or rent a dedicated cloud instance starting at one GPU. On-premises deployments are licensed annually or over multiple years, priced by the number and type of GPUs. No dollar or euro figure is published for any of those tiers.

Is Alto better than GPT 6 Astra or Claude Fable 5.1?
Not on Giotto's own benchmark table. Alto ties both at 100% on AIME 2026 — a saturated test at the top — and leads on Math-500 at 99.80%. But it trails GPT 6 Astra on GPQA-Diamond (91.9% vs 96.1%) and trails Claude Fable 5.1 on HLE (52.0% vs 59.1%). Alto's case is efficiency and control, not raw benchmark leadership.

What does the Alto API cost in dollars?
Giotto prices the API in euros: €2.10 per million input tokens and €10.50 per million output tokens, with cache hits at €0.21 and cache writes at €2.10. Giotto's own converter shows $2.47 for input, $12.35 for output, $0.25 for a cache hit and $2.47 for a cache write. Those dollar figures come from the site's own conversion rate, not from a separate dollar price list.

Is Giotto open source?
No. Giotto publishes no model weights, no parameter counts and no public licence. It is proprietary software that you license to run on hardware you choose — portable deployment, not open weights. If downloadable weights matter to you, Mistral Large 4 and Kimi K3.1 Pro are the closer comparisons.

Key takeaways

• Giotto launched Solo and Alto on October 27, 2026: free anonymous cloud chat with no account, plus iOS and Android apps.
• Solo runs on a single GPU; Alto runs on eight NVIDIA B300 GPUs; both sit inside Giotto's operating system of assistant, documents, code, agents and monitoring.
• The Alto API costs €2.10 per million input tokens and €10.50 per million output tokens (about $2.47 and $12.35), with a flat-fee option whose price Giotto has not published.
• On Giotto's own benchmarks, Alto hits 100% on AIME 2026 but trails GPT 6 Astra and Claude Fable 5.1 on GPQA-Diamond and HLE; Solo leads its single-GPU class on every test Giotto publishes.
• Sovereignty is real but partial: European hosting, GDPR alignment and ISO 27001, with no US cloud dependency — but NVIDIA hardware, proprietary licensing and no open weights.

October 27, 2026 is the day sovereign AI stopped being a press release and started being a free chat window. Giotto's bet is that enough people will pay for control — even at $12.35 per million output tokens — to build a European alternative that does not need to win every benchmark. The free tier makes that bet costless to test. The missing numbers — the flat fee, the enterprise prices, the independent benchmarks — are what will decide whether it is a movement or a niche.