Document search finds the right passage more often
Search in your documents now matches exact words as well as meaning, merges the two and reranks the best passages by the sentences that fit the question. On our test handbook of 30 questions, the right passage came first for 28 (93%), up from 19 (63%) with meaning alone, and was in the first three every time, up from 83%.
Each cited sentence is compared with the passage it cites. When the passage doesn’t share the sentence’s words or numbers, the citation is marked and the source list says what to check.
Workspace
Long conversations and long documents
When a conversation outgrows the model’s context, its earlier messages are summarised instead of resent, and turns that match your question come back word for word; a marker shows where the summary begins, and each conversation can send everything in full instead. Asking for a summary of a very long document summarises it in parts and combines the notes, keeping each citation to its original passage.
Workspace
Steadier answers
An answer that starts repeating itself is stopped, and one cut off by a dropped connection is kept; both can be continued. A tool call with arguments that don’t fit is sent back to the model to fix, and a failed round is retried once. Images, recordings and videos are checked against one set of limits, with the same preview for each.
Models
Auto learns from your blind tests
After a few blind-test votes on the same kind of message, such as code or images, Auto prefers the model you chose, on this device, and says when your votes made the choice.
Models
How we test models
A suite of checks for research, tool calls, citations, memory and long-document summaries now runs on every change, against a scripted stand-in model. The Models page explains the method; results for each model appear there once runs against the real models are on record.
The workspace and shared links follow your device’s light or dark setting, or the one you pick in Settings › Appearance. Diagrams, sheet music, slides and dashboards keep a light card, so they read the same either way.
Dark modeThe same answer in the dark theme.
Recorded in the workspace; answers scripted for the recording, some clips sped up. The Verified checks ran in the browser against a recorded attestation.
16 changes
A much fuller workspace
+10
Read the notesHide the notes
Workspace
Python on your files
Attach a spreadsheet or data file and the model can write Python that runs in your browser, in a worker with no network access, with numpy, pandas and matplotlib. Tables come back sortable, files saved to /out/ come back as downloads, and variables last for the conversation. JavaScript blocks get a Run button in a sandbox of their own.
Python on your filesAttach a spreadsheet and the model writes Python that runs in your browser tab, with tables and charts.
Recorded in the workspace; answers scripted for the recording, some clips sped up. The Verified checks ran in the browser against a recorded attestation.
Projects group conversations, documents and standing instructions. Sync encrypts them in your browser with a key that exists only as your recovery code, and a passkey can unlock it on a new device. A shared link carries its key after the #, so we store only an encrypted copy; links last 1, 7 or 30 days, or open once.
Privacy
Shared projects, end to end encrypted
Invite people to a project as editors or viewers. Its key reaches each member sealed to their own key pair, once the owner’s browser confirms them, and changes when someone is removed. Our server sees membership, roles and sizes, not what’s in the project.
Shared spacesWork in an end-to-end encrypted project with your team: see who’s here, who’s typing, and talk about answers.
Recorded in the workspace; answers scripted for the recording, some clips sped up. The Verified checks ran in the browser against a recorded attestation.
Models can search your documents, run Python and search the web on their own, in capped rounds you can follow. Answers cite numbered passages, such as [1], that open where they came from. Deep research plans searches, reads sources and writes a cited report within a cost ceiling, and a web page can be added as a document from its address.
Auto picks a model for each message, in your browser. A blind test hides which model wrote which answer until you’ve voted.
Workspace
Voice mode, transcripts, video and scans
Voice mode is a hands-free conversation that listens, answers aloud and can be interrupted. Recordings and videos get timed transcripts to download as subtitles, and a video can be asked about from frames taken in your browser. Images can be cropped, marked up and redacted before sending, and a photo of a page can be scanned into a document.
Answers can draw diagrams, SVG pictures and sheet music, show slides as a deck, and build dashboards from attached spreadsheets. Any answer or conversation downloads as Word, PowerPoint or PDF, made in your browser.
Workspace
Code changes as diffs
Changes a model proposes to an attached codebase show as diffs, and the codebase downloads as a .zip with them applied.
Edit a sent message and step between versions of the answer. Search and pin conversations, quote part of an answer in your next message, and continue an answer that stopped at its length limit. Images and files can be pasted into the message box.
Time travelScrub back through every message and edited version of a conversation, and pick one up again.
Recorded in the workspace; answers scripted for the recording, some clips sped up. The Verified checks ran in the browser against a recorded attestation.
Workspace
Memory, saved prompts and conversation settings
Memory notes go with each message as instructions; a model can suggest one, and it’s saved only if you choose. Save prompts and pick them by typing /, and give a conversation settings of its own.
Privacy
A receipt for each answer
An answer from an enclave keeps a receipt of what your browser verified before encrypting: the enclave, the code release it matched and the key the request was encrypted to. Download it to check again with Tinfoil’s verifier.
Workspace
Import from ChatGPT and Claude
Bring in conversations from a ChatGPT or Claude export. The file is read in your browser by a worker with no network access, and you choose which conversations come in.
Developers
A usage page and key limits
Usage shows 30 days of estimated spending, worked out in your browser, with a daily spending alert. An API key can carry its own daily cap and list of models.
Developers
A fuller developer API, and a playground
The API takes images, function calling and structured outputs, and adds embeddings, transcription and speech. An encrypted mode seals requests to the enclave so not even our gateway reads them. The docs have a playground that keeps your key in the page’s memory only.
Kimi K3, Gemma 4 31B and DeepSeek V4.1 Flash can see attached images. They’re shrunk in your browser and sent encrypted, and a model that can’t see images tells you which ones can.
Attach a folder or a .zip of code, or drop it on the message box. It’s read in your browser, leaving out dependencies, build output, binaries, .env files and keys.
Workspace
Recordings up to 30 minutes
Audio files up to 30 minutes are split into pieces in your browser and transcribed inside an enclave, with a summary suggested.
Workspace
Read aloud
Listen to an answer, voiced inside an enclave and played piece by piece.
Workspace
Search inside long documents
Documents over 120,000 characters are indexed inside an enclave, and each message sends only the passages that match it. The index stays in your tab.
Site and plans
Tidier on desktop and phone
Larger tap targets, a message box that grows with your text, and steadier spacing across the site and workspace.
6 changes
Encrypted to the enclave
Read the notesHide the notes
Privacy
Encrypted from your browser to the enclave
The workspace verifies Tinfoil’s enclave in your browser, then encrypts each message, document and recording to it. Undergrove’s relay passes the ciphertext on without being able to read it.
Verified, and encryptedEach answer shows its path, sealed in your tab, and what actually left the device.
Recorded in the workspace; answers scripted for the recording, some clips sped up. The Verified checks ran in the browser against a recorded attestation.
See what your browser checked before sending: the enclave’s hardware attestation and the open-source code release it runs.
Workspace
Fast or Thorough, web search and files
Choose Fast or Thorough for each message, search the web with the sources listed, dictate with voice input, and attach PDF, Office and image files, converted inside an enclave.
Models
Kimi K3 and DeepSeek V4.1 Flash
Two more open models join the register: Moonshot AI’s Kimi K3, and DeepSeek V4.1 Flash with a context of a million tokens. Compare puts two answers to the same question side by side.
Use is measured against a daily allowance at the listed prices. Self-serve API keys and Stripe plans are built into the gateway too; payments aren’t switched on.
Site and plans
A site that leads with the models
The homepage opens with the live models, their context and prices, and how encryption to the enclave works.
4 changes
Live models
Read the notesHide the notes
Models
Three models, live on confidential inference
GLM-5.3, Gemma 4 31B and GPT-OSS 120B answer live through Tinfoil, whose SDK verifies the enclave’s hardware attestation and encrypts each request to it.
Workspace
A streaming workspace
Answers stream as they’re written, with stop and regenerate, visible reasoning and Markdown. Saving conversations on your device is your choice.
Developers
The developer API
An OpenAI-compatible API for key holders at /v1/chat/completions, with streaming, and /v1/models. Requests are read in memory and not stored.