Three models in September: The real-world test for structured records

Following the summer of 2026, the conversation around language models has shifted markedly. It is rarely about who writes the longest prose anymore. In enterprise applications, a different criterion counts: are the numbers exact, do table rows match, and does extraction break down on complex layouts?

Three models shape September 2026 in our daily evaluation environment:

• OpenAI o-series: Models with upfront reasoning steps that validate arithmetic internally before producing output.
• Claude 3.7 Sonnet (Anthropic): A hybrid architecture combining visual document parsing with continuously configurable thinking time.
• Mistral Large 2 / Llama 3.3: High-performance open-weight models for autonomous on-premise operation without US cloud data leakage.

Technical documentation on current reasoning techniques can be found directly with the developers: OpenAI: Learning to Reason with LLMs.

The Swiss stress test: Cents, rappen, and value-added tax

Standard language models tend toward plausible-sounding estimates when faced with columns of figures. On a Swiss contractor invoice with 38 line items, early-payment discounts, and mixed VAT rates (8.1% standard, 2.6% reduced), an estimate is worthless. If the final sum is off by five rappen, automatic reconciliation in the ERP fails.

Here, the new reasoning models of September 2026 demonstrate a genuine leap: because they calculate intermediate steps before emitting the final JSON structure, the error rate in mathematical consistency checks drops from nearly 12% in older models to under 0.8%.

A language model does not calculate — a reasoning model checks its intermediate steps before replying.

Multilingualism without context switching

A hallmark of the Swiss market remains multilingualism on a single document: a delivery note from Ticino with an Italian header, German item descriptions, and French payment terms routinely overwhelms rigid OCR rulebooks.

All three models tested in September detect language switches within the same table without prompting and map fields semantically correctly. The time spent on manual post-corrections in mail processing drops drastically as a result.

The limit: A model is not an accounting system

Despite these advances, one fundamental rule stands: no language model may book unchecked directly into a general ledger or ERP. In professional architectures, the model serves semantic extraction — subsequent validation (checksums, IBAN check digits, VAT register lookup) must be enforced by deterministic code and hard guards.

For Swiss enterprises, this means: the models of September 2026 are fully mature for demanding document pipelines — provided they are embedded in an architecture that maintains audit trails and escalates ambiguities to human operators.