Changelog
What’s new in Superbrain
Release notes for the beta. New builds reach testers automatically through in-app updates — no reinstall needed.
August 2026
- v0.4.10
- Fixed long-running tasks stopping partway through and asking you to resend your message — Superbrain now renews your session in the background and the task keeps going.
- Changed session renewal to cover every tier, so Auto and your own API key behave the same on long tasks across every supported provider.
- Added automatic recovery from a dropped connection — Superbrain waits and retries for up to 5 minutes, then picks the task up exactly where it left off.
- Added a reconnecting notice while the connection is down, so a pause is never mistaken for a freeze.
- Changed brief provider outages, overloads, and rate limits to be waited out rather than ending the task.
- Changed a replaced API key to take effect on the next step of a running task, instead of after a restart.
- Changed Stop to take effect immediately while Superbrain is waiting to reconnect.
- Changed sign-in and billing messages to state what actually needs doing rather than asking you to try again.
- Changed cost reporting to price each provider's prompt cache at that provider's own rates, so the figure you see matches the bill you get.
- Fixed reasoning blocks and images counting as zero tokens in cost and context figures.
- Fixed cached context being dropped early on providers other than Anthropic, which quietly made follow-up questions cost more than they should.
- Fixed delegated work re-sending its shared setup on every sub-agent run.
- Changed
@fileattachments to be budgeted per file, and to say plainly when part of one was left out. - Changed citations to resolve to exactly one file or none — an ambiguous reference is now reported instead of guessed.
- Added a closing note listing any statement in an answer that isn't backed by a file Superbrain actually read.
- Fixed a citation pointing at a line Superbrain never read being presented as verified.
- Changed read tracking to carry across turns, so a file read earlier in a session still counts as evidence for a later answer.
- Fixed answers being treated as edit requests because the wording happened to contain a word like "update" or "fix".
- Fixed a symbol's name influencing what kind of request Superbrain thought you had made.
- Fixed paths differing only in letter case being treated as two files on case-insensitive filesystems.
- Added
superbrainandsbterminal commands — open any folder in Superbrain straight from your shell. - Added automatic setup of the terminal command on macOS at first run, without a password prompt where possible.
- Added the terminal commands to the Windows and Linux packages, so they are on your PATH right after install.
- Added shell completions for the terminal commands.
- v0.4.9
- Fixed broad, whole-codebase questions being worked one file at a time — the agent now delegates wide exploration to fast background sub-agents and answers from their summaries, keeping the chat quick and lean.
- Fixed the agent being limited to a single background exploration per session — it now runs several in parallel and revisits new areas as the task moves.
- Fixed background exploration not saying what it left uncovered, so the agent no longer over-claims or repeats searches it already ran.
- Fixed the debugging helper not reasoning through a failure before proposing a fix.
- Fixed code review routing the whole diff through the main chat — the review sub-agent now fetches it directly, keeping the conversation lean.
- Fixed a multi-second freeze when renaming a symbol or repeating an edit across a very large file.
- Fixed a hang when an edit's search text was empty, and confusing failures when two edits to the same file overlapped.
- v0.4.8
- Fixed edits to files with Windows line endings occasionally corrupting the edited line — replacements now keep the surrounding characters exactly intact.
- Fixed the agent stopping after five files when asked to edit more than five in one request; it now edits every file you name.
- Fixed whitespace-only edits, like adding a trailing blank line, being wrongly rejected as "no change."
- Fixed the agent sometimes editing the wrong part of a file on the first try, so more edits land correctly without a retry.
- Fixed the agent occasionally refusing to re-open a file it had already read during a long editing session.
- Fixed slow edits on very large files.
- v0.4.7
- Added free Auto on every surface — the chat panel, inline chat, and any extension that asks Superbrain for a model now run on your beta credits, with no API key required.
- Changed Auto to stay free even after you add your own API key — Auto runs on your beta credits and never spends your key.
- Added a usage meter to the account menu showing how much of your free window you've used and when it renews.
- Changed Superbrain to tell you exactly when your free window renews, and offer to continue on your own key, instead of failing quietly when it runs out.
- Changed the context panel to read "Auto Mode" and "Sub-agent" while on Auto, instead of showing the underlying model name.
- Fixed answers arriving all at once — replies now stream in as the model writes them, so the first words appear right away.
- Fixed Stop not taking effect while a request was waiting to retry, which could leave the button unresponsive for up to 45 seconds.
- Fixed reasoning appearing twice in some chat surfaces.
- Fixed sub-agents continuing to run after you dismissed them.
- Fixed TokenFold missing generator functions — the agent can now find and reason about
function*andasync function*code. - Changed TokenFold to surface better matches on large repositories, so the right code turns up more often.
- Fixed slow code search on large repositories — results now come back noticeably faster.
- Changed TokenFold to make a freshly opened project ready to search sooner.
- Changed TokenFold to use less memory while it works.
- Fixed the first question of a session being slow to start — sign-in, keys, and connection setup now happen while the window loads.
- Fixed Superbrain signing you out when several windows started at once.
- Fixed macOS builds never updating themselves — every Mac had been stuck on the version it was first installed with, and updates normally from this release onward.
- Changed TokenFold to use your Mac's performance cores, making its first pass over a large project several times faster.
- v0.4.6
- Added the model's reasoning to chat — it streams live while the model thinks, then folds away when the answer starts.
- Added a
superbrain.thinking.displaysetting to keep reasoning open after a turn or hide it entirely. - Changed chat to follow the work as it happens, while leaving you in control the moment you scroll up to read.
- Added a pinned header naming the question the current answer belongs to; clicking it takes you back there.
- Changed the chat layout to a flatter typographic hierarchy, replacing the boxes and rails that framed each part of a turn.
- Changed "Early Access" to "Beta" throughout the app.
- Fixed the macOS download being named VSCode instead of Superbrain.
- v0.4.5
- Added a macOS build for Apple Silicon.
- v0.4.4
- Fixed the agent editing a file it guessed at from a vague request — it now front-loads a file for editing only when you name it (a path, a symbol, or a backticked identifier) and otherwise finds the target itself.
- Fixed the agent working from a stale copy of a file that changed mid-task — after a formatter, a codegen or build step, a
git checkout, or your own save, it re-reads the current contents before editing. - Fixed the agent skipping its read-before-edit checks when a request scoped out one action ("don't fix the tests, just rename this function").
- Fixed the agent stalling or ending a task too early on long, multi-step work — it keeps going while it is making verified progress.
- Fixed internal diagnostic and rate-limit text leaking into replies — the agent now answers from the evidence it gathered.
- Fixed the agent cutting work short on non-default models and on long background runs.
- Fixed wasted cost from the agent re-reading files it already had open — it reuses what it has read and re-reads only when a file changed.
- Fixed the agent treating files it hadn't touched as already edited.
- Fixed codebase audits leaning on assumptions about which bug matters most — findings are now ranked by impact demonstrable from the code.
July 2026
- v0.4.3
- Fixed test runners and build tools that expect a POSIX shell failing on Windows — the agent now runs commands in Git Bash when it's installed (the same shell environment as macOS and Linux), falling back to the standard Windows shell otherwise.
- Added a
SUPERBRAIN_WINDOWS_SHELLsetting to force the standard Windows shell or pin a specific shell path. - Fixed a build or an empty test selection being counted as a passing test run; a pass is now credited only when a run's output shows tests actually ran and passed.
- Fixed a project's own custom or compiled test command not being recognized as verification when its output shows a real pass.
- Fixed the agent giving up after one failed attempt to run your tests — it now tries alternate invocations until one works in your environment.
- Fixed the agent abandoning an approach after a single command failure — it now diagnoses the error and tries a focused fix before switching, raising a blocker only after investigating.
- Fixed the agent running the whole test suite for a change that doesn't need tests — a comment, doc, or rename is now confirmed with the lightest sufficient check, such as a type-check.
- Fixed the agent reporting a change as verified when no test run confirmed it; it now states plainly what it could not verify.
- Fixed the agent misjudging whether a request is an edit or a question from specific wording — detection is now keyed to what a request does, for consistency across languages and multi-package repositories.
- v0.4.2
- Added line numbers to the code the agent reads, so the line references in its answers are exact.
- Changed test verification to report a pass only when tests actually ran: an empty selection, a build with nothing to compile, or a test binary older than the last edit no longer counts as passing.
- Added recognition of Rust (
cargo), Go, and Gradle test output, so a passing run in those toolchains isn't reported as inconclusive. - Changed the agent to use a project's own build and test commands from
AGENTS.mdorCLAUDE.md, and to detect the toolchain per directory in multi-language and multi-package repositories. - Removed limits that stopped the agent mid-task from running project builds, reading documentation while editing, and searching a codebase to completion.
- Changed build-tool, test-command, and edit detection to cover Python, Go, Rust, C++, and Java rather than assuming a JavaScript project.
- Changed context compaction to scale with the selected model's window, up to 1M tokens, instead of a fixed fraction.
- Added confinement of file writes, edits, and deletes to the workspace folders you have open, including across a multi-root workspace.
- Fixed multi-file edits resolving a relative path to the drive root instead of the workspace.
- Fixed a delay finalizing an answer that contained a very long unbroken token such as a hash list or an encoded blob.
- v0.4.1
- Fixed a background update that could fail partway and leave an installation that would not start, requiring a manual reinstall. The updater now verifies it can replace every file before it deletes anything; if something still holds a file, the update is skipped and retried later instead of applied halfway.
- Fixed the update window showing the wrong product name in its title bar.
- v0.4.0
- Added 18 models from 7 providers. The picker now carries 25 models across Anthropic, OpenAI, Google, DeepSeek, Kimi, xAI, Mistral and Sarvam; before this release every selectable model was Anthropic.
- OpenAI — GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GPT-5.5, GPT-5.4 Mini, GPT-5.3 Codex
- Google — Gemini 3.1 Pro, Gemini 3.5 Flash, Gemini 2.5 Pro, Gemini 2.5 Flash
- Kimi — K3, K2.7 Code, K2.7 Code HighSpeed
- DeepSeek — V4 Pro, V4 Flash
- xAI — Grok 4.5
- Mistral — Large 3
- Sarvam — 105B
- Added per-provider API keys, held in the OS keychain and settable from the account menu or
Superbrain: Set API Key. Each provider also reads its conventional environment variable (OPENAI_API_KEY,DEEPSEEK_API_KEY,MOONSHOT_API_KEY, and so on). - Context windows now follow the selected model, from 128K on Sarvam 105B to 1M on Fable 5, Sonnet 5, GPT-5.6, Gemini, DeepSeek V4 and Kimi K3. Opus 4.6–4.8 and Sonnet 4.6 run at 500K, GPT-5.3 Codex and GPT-5.4 Mini at 400K, Kimi K2.7 at 256K, Haiku 4.5 at 200K.
- Added prompt caching on every provider that supports it. Repeated context bills at the provider's cached-input rate — 10% of the fresh rate on Anthropic and Kimi, under 1% on DeepSeek.
- Changed the reasoning-depth picker to list only the levels the selected model accepts, so each option maps to a distinct setting. Models with no depth control no longer show the picker.
- Changed sub-agent exploration to run on the cheapest model in the selected provider's lineup, and on the parent model where nothing cheaper exists.
- The context panel now reports tokens and cost for the model in use, counting cached reads separately from fresh input.
- Fixed the account menu showing beta credits after an Anthropic key was removed while keys for other providers were still set.
- v0.3.3
- Changed the agent to ground answers in code it has actually read. An empty search result is no longer treated as proof that something does not exist, and a file, function, or feature is confirmed missing before it is reported as missing.
- Added explicit reporting of what was not checked. When part of a project has not been explored, the agent names the gap instead of inferring across it.
- Improved answers to questions spanning several parts of a codebase, which now cite the specific files and lines they rest on.
- v0.3.2
- Fixed being signed out after an update or on the next launch. Sessions are restored and refreshed in the background.
- Fixed a brief network failure at startup — waking from sleep, or Wi-Fi or VPN reconnecting — ending the session instead of being retried.
- v0.3.1
- Changed TokenFold to skip non-source directories: dependency downloads, package and build caches, and build output. The first TokenFold checkpoint of a large repository completes sooner and results cover only your own code.
- Changed TokenFold to pace its work against available memory and recover on its own after a stall, so it does not compete with the editor on large or unusual projects.
- v0.3.0
- Improved how fast TokenFold checkpoints and refreshes a codebase. Projects become searchable sooner and stay responsive while you edit.
- Changed TokenFold to size its work to the host machine and run in the background.
- Fixed stale search results after file changes. Additions, edits, and deletions are reflected reliably, with a background pass to catch anything missed.
- TokenFold checkpoints and searches your codebase entirely on-device. Source code is not sent anywhere for Superbrain to understand it.
- v0.2.9
- Added Claude Sonnet 5 and made it the default model, available to all beta users with no setup.
- v0.2.8
- Changed the model picker to group models by provider.
- Added premium models, selectable once you add your own API key.
- Fixed the newest Opus models failing to respond when selected.
- v0.2.7
- Added encryption for the data Superbrain keeps on your device, which also covers copies made by backup or file sync. On by default, with no setup and no passwords.
- v0.2.6
- Added a Feedback window, reachable from the account menu, the Help menu, or the command palette. Bug reports can carry reproduction steps and screenshots.
- v0.2.5
- Fixed diagnostic logs reaching the team incomplete, so faults reported during the beta can be diagnosed from a full record. Logs are encrypted on disk.
- v0.2.4
- Fixed accounts not yet on the invite list seeing a repeating "session refreshed" message instead of an explanation and how to request access.
- v0.2.3
- Internal telemetry accuracy fixes. No user-facing changes.
- v0.2.2
- Changed the Windows installer to carry Superbrain branding throughout.
- v0.2.1
- Internal telemetry work. No user-facing changes.
- v0.2.0
- Internal reliability and telemetry work. No user-facing changes.
Loading earlier releases…
Showing 2 of 3 months of releases.
Superbrain is in beta · 31 releases since June 2026