Talos-1 went live for business customers on September 15, Ablatic AI announced — the official launch this blog flagged as the thing to watch when the JKU Linz spin-off first went public in August with a beta running across roughly 100 companies. That beta period is now dated precisely: 2.5 months, ending with market availability. The pitch itself hasn't changed — a large language model built, trained, and operated entirely on European infrastructure, positioned against the US CLOUD Act's reach into EU-hosted data — but what's new this week is the part the August announcement didn't have yet: a named model with published numbers, real pricing, and a specific technical claim about why a small university spin-off could afford to train it at all.
The efficiency claim is the actual news here
Ablatic says Talos-1 uses an in-house model architecture that breaks the standard quadratic-cost relationship between context length and compute — in a conventional transformer, doubling document length roughly quadruples processing cost. Talos-1's architecture instead uses up to 14 times less computational power and moves up to 79 times less data through memory than a conventional model of the same size when generating a response, with the advantage growing with context: eightfold at 256,000 tokens (roughly 500 pages), fourteenfold at one million. These are architecture-derived figures from Ablatic itself, not an independent benchmark, but the practical framing is concrete and worth taking at face value as a real engineering constraint solved rather than a marketing number: CTO Julian Mauerkirchner's own quote is that the team had to cut training and operating costs before it could be competitive at all, and the company states plainly that training on long sequences — which for large providers consumes data-center capacity at a scale of billions — was only feasible within Ablatic's actual EuroHPC compute allocation because of this architecture. For a team this size, riding on public EuroHPC compute rather than a hyperscaler's own cluster, a sub-quadratic architecture isn't an optimization, it's the precondition for the whole project working at all.
At launch, Talos-1 processes 200,000 tokens per request; Ablatic plans to raise that to 512,000 later this year and eventually to 1 million, which is what the architecture is designed for — the company says the remaining constraint is server capacity, not the model itself.
What the trial period actually tested
The press release names specific pilot use cases rather than staying abstract: marking homework on a learning platform, searching a telecommunications company's internal documentation, and posting receipts in an accounting department. That's a useful concreteness check against the January-scale framing this blog flagged in August, that Talos is meant for AI that "carries out work" rather than answers questions — reading contracts, filling in forms, completing a process across multiple systems. Three real pilot tasks across three different industries, run over 2.5 months with 100+ companies, is a genuinely more substantial trial than most model launches disclose before going to market.
The benchmark numbers, with the standard caveat
Ablatic's own comparisons put Talos-1 at 74.4% on Toolathlon (a test of autonomous agent tasks in real software — email, calendars, spreadsheets), against 80.6% for Claude Opus 5 and 74.9% for GPT-5.6 Sol; at 82.2% on IFBench (precise instruction-following), ahead of GPT-5.6 Sol, DeepSeek V4 Pro, and GLM-5.3; and at 90.8% on OmniDocBench (parsing real documents with tables and formulas), two points behind the top score. All of this is Ablatic's own comparison table, published at ablatic.ai/models — the same caveat that applies to every vendor benchmark chart this blog covers applies here: real, specific, and checkable claims against named competitors, but not yet an independent aggregator's confirmation. What's worth noting on its own terms is the shape of the result: a small European spin-off's first shipped model landing within single digits of Claude Opus 5 and ahead of GPT-5.6 Sol on at least one agentic benchmark, rather than trailing by the wide margins usually associated with a first-generation regional model.
Pricing and the legal framing, now with numbers attached
Talos-1 is priced at €1 per million input tokens, €4 per million output tokens, and €0.10 per million cached input tokens, served through an OpenAI-compatible API so existing integrations migrate without rework. Ablatic states customer data isn't used for training and that it operates as a data processor under GDPR Article 28 — the same jurisdictional argument this blog covered in August, that EU hosting alone doesn't block foreign legal compulsion the way EU incorporation and EU-only operation can, now attached to an actual live, priced product rather than a beta commitment. Access itself is still gated rather than self-serve: companies reach Ablatic directly and start with a pilot built around their own use case, with open self-registration planned once server capacity expands — a sequencing that matches a team still scaling infrastructure to meet the context-window roadmap it just published.
What the market launch actually settles is narrower than "Europe has its Claude or GPT competitor." It settles that the promise this blog covered as beta-stage in August shipped roughly on schedule, with more disclosure than most model launches carry — a specific architectural mechanism for the cost claim, named pilot use cases, and published pricing — while leaving the harder question, whether those benchmark numbers hold up once someone other than Ablatic runs them, exactly where it was two weeks ago.