Short answer: seconds, not hours. On four questions from our recent posts, the assistant answered in between 1.4 and 12.5 seconds, for at most two cents of model time each. In the runs shown, every number matches the chemistry engine, and every case it computed is there to check in the app.
Both Expert Buffer Designer and Expert Solution Operations Designer now have an assistant panel. You ask in plain language. It sets up the designs in the app, runs them on the same engine you would, and answers with the numbers. Here are four questions we put to it, timed from pressing send to the finished answer.
The scorecard
Each question has a short video of the actual run: click Watch to jump to it.
| Question | Answer in | Video | What it computed | Model cost |
|---|---|---|---|---|
| Compare four buffer families from pH 3 to 9.5 | 7.8 s | ▶ Watch | 108 buffer designs | about 1¢ |
| Make four buffers two ways each, and compare | 3.2 s | ▶ Watch | 8 buffer designs, with grams per liter | under 1¢ |
| Build a viral-inactivation step, then sweep starting pH against four acids | 12.5 s | ▶ Watch | A 6-step workflow, then 24 more runs with two titrations each | about 2¢ |
| Make a histidine buffer at 5 °C, then read it from 4 to 37 °C | 1.4 s, then 2.3 s | ▶ Watch | 1 design, then a 12-point temperature sweep | under 1¢ |
Where the time goes. Almost all of it is the language model deciding what to do and writing the answer. In an instrumented run of the first question:
- the assistant called the model six times, for 6.9 seconds in all;
- all 108 buffer designs, solved in full, took 1.2 seconds.
The viral-inactivation study is the heaviest. Each of its 24 cases is a whole workflow with two titrations, about a fifth of a second each.
How we timed it. These runs are on a local build of the suite: the same engine and the same model (Gemini 3.5 Flash-Lite) as the beta. On the beta, each answer also carries the round trips to our server. Times vary from run to run: the buffer comparison has taken between 4.8 and 10.8 seconds, the 24-case study between 12.5 and 18.3. Each video below shows the actual run, with a live stopwatch.
1. 108 buffers, 7.8 seconds
Asked to compare Tris, citrate, acetate and phosphate at 50 mM, it swept all four across pH 3 to 9.5. It came back with where each one buffers, its peak capacity, its temperature sensitivity and its ionic strength. Tris drops 0.028 pH units per degree, so a Tris buffer made to pH 8.0 at room temperature reads 8.6 in the cold room.
2. Same buffer, different chemicals, 3.2 seconds
Acetic acid plus sodium acetate, or acetic acid plus sodium hydroxide? It made each buffer both ways, with the grams for each, and confirmed they come out identical. It put all eight side by side on the Compare tab.
3. A 24-case process study, 12.5 seconds
From one paragraph describing the pool, the acid and the base, it built the viral-inactivation workflow, ran it, then swept the starting pH against four acids. The answer gives the acid volume, the base volume and the final volume for all 24 cases. Hydrochloric acid needs about a sixteenth of the acetic acid volume.
Done by hand, that is 24 workflows with two titrations each: an afternoon in a spreadsheet. On the app’s Sweep tab, it’s a few minutes of setting up the axes. With the assistant, it’s one message.
4. How to ask
The last video covers four habits that get the best answers:
- Say what, how much, and the conditions: concentration, pH, salt and temperature. “No salt” and “at 5 °C” change the answer.
- It works on your real design. The recipe and the charts update, so check them.
- Ask for a sweep or a comparison when you want many cases. For a buffer made once and then read at other conditions, say “make once, read at each”.
- Open the steps under any answer to see exactly what it ran, and what it cost.
Check its work
The assistant shows its working. Every design it makes sits in the app’s own tabs: the buffer on the left, every case on the Sweep tab, compared designs on the Compare tab. The steps under each answer list every tool it called. We checked every number in this post against the engine directly, and it isn’t always right: on the 24-case study, about one run in three came back with cases missing or volumes wrong. That’s why every case lands in the app where you can see it. Treat its answers the way you would a capable colleague’s: fast, usually right, and worth a look before they go into a batch record.
Try it
Both apps are in invited beta. Ask for an invitation, or read about Expert Buffer Designer and Expert Solution Operations Designer.