Omni

26 new models shipped in September, from 16 providers. This week: Gemini 4, Claude Sonnet 5.5, GPT-6.1 Sol. Last week: Opus 5.5, GPT-6 Sol and Luna, Grok 4.7. The open models keep pace. GLM-5.3, Kimi K3, and now Ember-1, a post-trained K3 from Fireworks that reasons in about 40% fewer tokens. Then there are entirely new categories! Jev from TypeSafe AI and OpenAI's Decisions API don't generate text at all. You hand them a question and a list of allowed answers and get back a typed decision with calibrated confidence. TypeSafe claims Jev is up to 200x faster than an LLM on classification. The pace of development is increasing and you want to hit the right edge of the Pareto efficiency curve for all of the use cases you're targeting. This means: - Using the right model for the right use case via dynamic routing. Data modeling and application building on Astra or Fable. Querying on Sonnet. Validation on Jev. Data summarization on Luna. And the state of the art changes every week. - Elaborate evals, so you know where a new model helps and what needs work before it reaches a user. - Efficient harness architecture: tool and prompt caching, progressive discovery of context, determinism, so a cheaper model can do the same job. I met with a company this week whose 3-person data team is building a custom data harness. It's fun! But it's increasingly complex (expensive) with no end in sight. That's the work we do at Omni so our customers don't have to. The semantic layer and the evals ensure your agentic analytics stays accurate and efficient. The models underneath are constantly improving.

Ask an LLM how confident it is and it will give you an answer, but it's basically grading its own homework. 😅 Enter everyone's favourite topic: Jev. It's a small model from TypeSafe built for quick judgment calls, and it does about 1,000 labels in a second. The more useful bit is that its confidence actually means something. We tested classifying ~3,200 rows and found that when Jev was at least 90% sure, it was right 86% of the time. Under 50% sure, it was right 1/4 times. In Omni's AI Hub, we use LLMs to label our AI conversations: what the person asked about, how it went, how they felt, and whether we could help. In that flow, a more efficient model takes a first pass and rates its outputs as "high" or "low" confidence. The low-confidence ones are then sent to a smarter model for second pass. Jev isn't in that pipeline today but I am looking forward to making it a part of this workflow. I've had many conversations with teams over the past week or so about where these small classifiers fit. Our session labeling example is a great one because: 1. we can be more token efficient and more accurately escalate to smarter models when we need to and 2. at the end of the day, if we miss flagging a session that didn't go as well as it could have, the consequences are not super dire. Would love to hear use cases from people already using Jev today or where you might be thinking about using it 👀

When a number looks off, the first question is "can I trust this?" Usually that turns into a Slack thread with the data team. Your analytics engineers already did the work that answers it. In dbt, they documented the fields, wrote the tests, and tracked where each model comes from. Omni brings that work to the people using the data. Descriptions from your dbt project show up next to the fields business users pick. Test results and freshness show up in a data health tile next to the numbers they cover. To learn more about how to use dbt Labs + Omni, join today's webinar at 9 AM PT/ 5 PM BST with Victoria Perez Mola and Chris Merrick: https://luma.com/wrexoa5j

We're back in London for Snowflake World Tour ❄️ At 16:15, head to Breakout 1 to hear Daniel Jordan (Wayflyer) dive into their Tableau → Omni migration. He'll cover how they grew weekly active users 23x by enabling business teams to build things the data team never scoped or imagined: https://lnkd.in/gwfr3eHg You can also find our team all day in pink at the Omni booth, see you soon!

We're officially recommending an OpenAI model as the default model family for Omni customers. Before making the call, we ran GPT-6 Sol against Claude Sonnet 5.5, Sonnet 5, Opus 5.5, and Opus 5 on our hardest analytics eval. Sol got 92% right at $0.23 a question. Sonnet 5.5, out this week, got 72% at $0.19. Sonnet 5, our default until now, got 60% at $0.32. Opus 5.5 came closest at 87%, but it cost $0.54 a question. (Sol was faster too, 29 seconds median vs 40.) It's interesting where the savings came from. Sol made almost twice as many model calls per question as Opus, about nine to five. But Opus spent about half its bill writing to the prompt cache, and Sol wrote a lot less. We ran 16 model and effort combinations across OpenAI and Anthropic. Sol on medium effort came out on top. Customers have been asking us for new models, and since last week I've been pointing them at Sol. So for now, new US customers start on GPT-6 Sol. If Claude wins the next round, we'll switch back.

Showbie migrated from Looker to Omni. With the migration behind them, the team is now focused on expanding Omni adoption, exploring more advanced features, and integrating additional data sources. Throughout the journey, the Showbie team brought us complex use cases, custom dashboard requirements, and challenging problems that required more than a one-size-fits-all approach. That’s where having the right partner made a difference. When we asked Erin to rate her experience with Shearwater, her answer was: "Five starts, hands down." Thank you, Erin Orris and the Showbie team, for your trust. And congrats to Gabriella El Khoury Ghanem for representing Shearwater so well. Planning a move to Omni? Let's talk. #Omni #BusinessIntelligence #DataAnalytics #LookerMigration #CustomerSuccess

Guten Morgen, Berlin ❄️ If you’re at Snowflake World Tour, swing by our booth to catch a live demo and meet our team. At 2:15, head to the Station Berlin Expo Stage to hear Victoria Perez Mola (dbt Labs) share how she uses Omni: https://lnkd.in/gAukNiD7 See you there!