Version History

2.9 is about seeing what the assistant is doing while it uses its tools.

TOOL STEPS

• Tool steps sat on "Queued" even after they'd finished. Now each one shows a spinner while it works and says Completed or Failed when it's done. That goes for web search, file search, Code Interpreter, image generation, MCP tools and the browser steps of Computer Use.

• Steps have their real names, like Code Interpreter instead of "code_interpreter_call". Tap the Code Interpreter step to see the Python it ran.

• When a step fails, it says Failed. That includes an MCP tool that returns an error, which the app used to miss.

• The answer after a tool call starts its own paragraph instead of running into the sentence before it, and the note the app shows when a browser step fails goes away once the answer starts.

CHARTS AND FILES

• A chart from Code Interpreter shows once. It could show up two or three times before.

• Each file Code Interpreter makes is listed once, and text files still have their contents when you reopen the chat.

ALSO

• Reasoning summaries show their formatting instead of raw asterisks.

• The Computer Use warning says what it actually drives: a browser inside the app, nothing else.

Still no account, still your own key, and your chat history stays on your phone.

2.8 is a big one. OpenAI released GPT-6.1 Sol while I was finishing it, so that's in, and new models should show up on their own from now on.

NEW MODELS

• GPT-6.1 Sol is in every model menu, with the reasoning levels OpenAI lists for it, Low through Max. It doesn't take None, so a preset set to None sends Low instead of failing.

• When OpenAI releases a model your key can use, it shows up at the top of the model menu the next time the app starts. The app reads that model's page on OpenAI's docs once to learn its reasoning levels, so you're not waiting on an update from me.

• Pro reasoning mode works on every GPT-5.6 and GPT-6 model now, not just Astra and 5.6 Sol.

• If OpenAI says a model shuts down within 30 days, it comes off the menus and any preset using it moves to its replacement.

NEW

• Local Python. Turn it on in Settings → Model and the assistant can write Python and run it on your phone. You see the code first and nothing runs until you tap Run. Standard library only, no internet, no files, 30 seconds a run.

• Search your conversations by what they're about, not just the exact words. It uses Apple's on-device language model, so the search never leaves your phone.

• monday.com, Airtable, Intercom and Vercel wouldn't accept the app's own sign-in address, so signing in to them now goes through a page on my site that hands you straight back to the app.

VOICE

• GPT-Live 1 stops when you talk over it. With headphones or Bluetooth and Voice Barge-In on, about 0.2 seconds of you talking pauses it, and if you only said "mm-hmm" it picks up where it stopped. On the phone's speaker it works like before, because the mic hears it there.

• Settings → Model → Voice opened and then closed itself within seconds. It stays open now.

EASIER TO READ AND TAP

• Text on the chat, Settings and Conversations screens has at least 4.5:1 contrast, the chat toolbar buttons have names VoiceOver reads out, and they're all at least 44 points to tap.

• At the largest text sizes the message bar pushed the chat past the edge of the screen. It fits now.

Still no account, still your own key, and your chat history stays on your phone.

OpenResponses 2.7 adds OpenAI's newest models and prepares the app for the releases that follow.

NEW MODELS

• GPT-6 Sol, the new default for new presets, alongside GPT-6 Astra and GPT-6 Luna, with every reasoning level each model supports, from None to Max.
• GPT Image 2.5 Flare and Sunburst, with the new X-High and Max quality settings and partial previews. GPT Image 2 remains available.
• Later general-purpose GPT releases that appear in your account's model list are recognized automatically, with current reasoning and tool controls.

STAYING CURRENT AS OLDER MODELS RETIRE

• Voice sessions saved with a Realtime model OpenAI is retiring switch to GPT Realtime 2.1.
• The offline model list leaves out models with announced shutdowns. Your account's live model list and the model ID field reach every current model your API project can use.
• Image quality adjusts automatically when you switch to an image model that does not accept the chosen setting.

CURRENT API COVERAGE

• GPT-Live 1 voice sessions through OpenAI's new Live API, selectable in Settings → Model → Voice. GPT Realtime 2.1 remains the default.
• Response moderation, live-web-access control for web search, hybrid file-search ranking, Code Interpreter memory limits, image input fidelity and compression, and the Fast and Ultrafast service tiers.
• The API Workbench now reaches every current endpoint the app works with: responses, conversations, models, moderation, embeddings, files, vector stores, containers, batches and Realtime client secrets. Deletions ask first.
• Batch jobs default to the Responses endpoint.
• Voice notes recorded with the microphone button are transcribed with GPT Transcribe into the message field, so you can edit them before sending.

REMOVED BECAUSE OPENAI SHUT THEM DOWN OR DEPRECATED THEM

• Assistants API tools and the Assistant import lab, fine-tuning jobs, published prompts, preview web search and preview computer-use tools, and retired models such as o1, o3, o4-mini and GPT-5. Saved chats keep working, and presets that used a retired model switch to OpenAI's recommended replacement, such as o3 to GPT-5.6 Sol.

RELIABILITY

• MCP tool discovery uses GPT-6 Luna, the lowest-cost current model.
• Concurrency and memory-handling fixes in the API Workbench, conversation saving and voice recording.
• The Computer Use notice now explains that approval is requested when OpenAI flags a safety check.

OpenResponses 2.6 expands the playground with current models, a new API Workbench, continuing voice conversations, and a more dependable tool and storage engine.

CURRENT MODELS AND CONTROLS

• Shared model catalog for GPT-6 Astra and GPT-5.6 Sol, Terra and Luna, with model-aware reasoning and request settings.
• GPT Image 2 controls and partial image previews; current Realtime and transcription models.
• Optional automatic compaction, hosted shell, tool search, async/programmatic calling and multi-agent responses for supported configurations. Model and tool availability depends on your API account.

NEW API WORKBENCH

• Edit literal JSON without smart quotes; recover request drafts after relaunch and confirm replacement before discarding edits.
• Explore HTTP, streaming and WebSocket responses, with invalid schemas rejected before sending.
• Inspect raw events, export complete responses, count input tokens and compact context.
• Retrieve or cancel background responses, inspect known conversations, continue client tool calls with real results, and try steering and current tool templates.

VOICE THAT KEEPS GOING

• More reliable listening after the assistant answers, interruption handling, mute and audio-route recovery.
• Corrected microphone format conversion and playback completion so the voice bars do not remain stuck in a speaking state.

MCP, WEB SEARCH AND BROWSER TOOLS

• New Connections page with provider browser sign-in, several accounts per chat, account naming, reconnect and permissions.
• 44 catalog entries plus live MCP Registry search. Providers requiring app registration or callback setup are clearly labeled.
• Secure OAuth refresh and hosted tool discovery with schema inspection, cancellation and bounded retries.
• More reliable on-device browser navigation, precise page-element actions, back/forward controls and correctly sized screenshots.
• Shared browser action limits, deadlines and cancellation; safer recovery from stale callbacks and duplicate tool calls.
• Corrected hosted web-search settings and clearer source/citation handling.

CONVERSATIONS AND ACTIVITY

• Richer response artifacts, tool results, image previews and multi-agent activity.
• Complete compacted context and better hydration of known remote conversation items.
• Safer chat switching and deletion during streaming, background save checkpoints and clearer storage errors. Existing conversations and presets remain compatible.

FILES AND DEVELOPER LABS

• File upload progress now distinguishes uploading, indexing, failure and search readiness. Check indexing again without uploading duplicates.
• Save complete Batch output and error files through the system share sheet, including available partial results.
• Import and validate reviewed text-chat JSONL datasets for eligible fine-tuning accounts. Chat export is clearly a draft, and restricted service availability is explained.
• Complete cursor pagination for supported file/job lists and improved Notion search pagination.

Retained Assistant JSON can still be imported into Responses presets; retired live Assistants operations are disabled. Demo Mode, Keychain credentials, first-send consent and existing Apple/Notion integrations remain available.

Version 2.5 brings a major under-the-hood overhaul to make the playground experience faster, safer, and more reliable.

Performance & Speed

• Greatly improved rendering speeds for the File Manager, Vector Store uploads, and chat history
• Scrolling through files and conversations is now significantly smoother and more responsive

Reliability Fixes

• Fixed an issue where the chat interface could occasionally lock up if a network error occurred while the AI was generating text
• Enhanced the URL detection engine to better recognize, extract, and render web pages directly within conversations

Privacy & Security

• Strengthened privacy safeguards to ensure authorization tokens and private keys are aggressively redacted and never leaked during app diagnostics

New Models

• Added full support for the new GPT-5.5 family of models, including GPT-5.5, Pro, Mini, and Nano

OpenResponses 2.0 is a major update since the original App Store launch.

• Added GPT-5.4, GPT-5.3, GPT-5.2, GPT-5.1, GPT-5-mini, refreshed model aliases, and better model compatibility controls
• Added Apple Calendar, Contacts, and Reminders tools with safer permission flows and improved local-time handling
• Added Explore Demo, camera capture, better file and image attachments, message quick actions, and cleaner chat and settings UI
• Added Request Inspector, activity feed, reasoning traces, token usage, and more reliable streaming and tool-call follow-through
• Added an explicit first-send OpenAI data sharing notice with Allow & Send consent before live requests leave the device
• Strengthened App Store disclosures, privacy copy, and approval flows for computer use and other sensitive actions