Changelog¶

New features, improvements, and fixes in each Silico release. Hosted workspaces receive releases automatically; there is nothing to install.

September 10, 2026¶

Publish research code to GitHub¶

You can now send a thread's codebase to a connected GitHub repository as a draft pull request. Ask Silico to publish a new codebase, and it can create a private repository in your account. Later changes update the thread's draft pull request.

Before Silico organizes experiment code into the shared codebase, you can review and edit its proposed folder layout. Approving the layout starts the work.

Dictate your messages¶

Use the microphone in a new thread or chat to speak your message. Stop recording to review and edit the text, or use Send to submit it after transcription.

Catch up on research progress¶

When you return to a thread, Catch me up summarizes the work completed while you were away and links to the relevant experiments. You can also type catch me up in the conversation. The recap stays in the thread for later reference.

A research brief now develops as findings come in, with figures from the original experiment reports. Open See full summary under Where things stand to read it.

Guide experiments through a pause¶

When an experiment needs your input, a decision card explains the blocker and shows the question to answer. In the worker chat, you can change the model or return to an earlier message and try a different instruction. The worker is the agent that runs an experiment; rewinding its conversation preserves files and submitted jobs.

GPT-6 Astra by default¶

GPT-6 Astra is now the default model for planning and experiment work. You can choose another model from the model picker.

More reliable work and clearer reports¶

Long conversations recover from more failures while condensing older history. Compute setup can recover after an agent timeout or Lab restart, and interrupted job submissions retain the information needed to find the job.

Message drafts survive a page refresh. Running experiments remain visible in the sidebar when another task needs attention, while report fixes improve chart error bars and the display of large figures.

September 4, 2026¶

GPT-6 Astra and GLM 5.3¶

You can now choose GPT-6 Astra or GLM 5.3 from a thread's model picker.

More reliable long-running research¶

Auto-research can resume when a turn ends without a next step. Updates from the worker, the agent that runs your experiment, reach the main conversation more reliably. Long threads also recover from more failures while condensing older conversation history to make room for new work.

Clearer reports and more flexible experiments¶

Updated report and figure guidance helps Silico explain results and make charts easier to read. For exploratory experiments, a plan can name measurements the worker may adjust as it learns, with a set budget and stopping rule. The worker is instructed to record each design it tries and check the selected design on fresh data.

Easier to review outputs and organize work¶

Task outputs use a consistent layout. Error and warning banners share the same format across the Lab, and you can reorder pinned reports by dragging them in the sidebar.

Better support for large-data transfers¶

Command-line transfers can split large folders into archive parts, with a listing file to help reconstruct the folder. You can set the number of parallel uploads and supply the transfer code from a protected file for unattended runs. The new large-data guide explains where to store datasets and how to move them to the compute that needs them.

September 1, 2026¶

Claude Fable 5.1 for everyone, sessions that stay signed in, faster long-running threads, and HTML apps you can take anywhere.

Claude Fable 5.1 is now available to everyone¶

Anthropic's most capable model is generally available in Silico today. Pick it from the model picker on any thread. Fable 5.1 is billed at Anthropic's published prices, including its lower rate for cached context, which is where most of an agentic run's tokens go.

Stay signed in¶

You no longer have to sign back in every day. Web and desktop sessions refresh themselves quietly in the background and stay active through long stretches of regular use, so you land straight in your work every time.

Faster, smoother long-running threads¶

Long research threads are noticeably snappier. Experiments launch instantly even on very large workspaces, chat stays fluid on threads with thousands of events, and turns that used to get stuck after a very large tool call now recover on their own. A dedicated report-builder agent writes up results while the worker keeps running the experiment, so you get polished reports without slowing the science.

The Codebase panel has a new design and shares one file tree with the experiment's Code tab, so what you see in the thread is exactly what the worker is editing. In the sidebar, drag to reorder pinned threads, drag a thread to pin it, and hover any thread for a card with its activity and estimated time to finish.

Download HTML apps as a single file¶

Any HTML app your agent builds can now be downloaded as one .zip with every asset included, ready to open, host, or hand to a colleague. Choose the exact version you are viewing or a snapshot of the current draft.

Also in this release¶

Follow the worker's interactive cluster sessions live in the Compute tab, connect your GitHub repositories during signup, a new LLM-as-judge skill for dataset evaluations, and a broad set of stability improvements across the Lab, cluster jobs, and serverless compute.