Voice Input, a Refreshed Model Lineup & AI Guardrails

Voice Input

You can now talk to Conception. Hit the microphone in the chat input and start speaking - a live waveform shows you that we're listening, and when you stop, your words are transcribed straight into the input where you can edit them before sending.

It's built for the way ideas actually arrive. Dictate a half-formed thought, keep typing to sharpen it, then send. Transcripts append to whatever is already in the box rather than replacing it, so you can layer voice and typing together. Changed your mind mid-sentence? Discard the recording and nothing is sent.

A Refreshed Model Lineup

We rebuilt the model picker from the ground up this cycle. The headline: several new frontier models, two new providers, and a long list of older models retired.

New this cycle:

• xAI Grok 4.5 - xAI's most intelligent model, built for coding and agentic work.
• OpenAI GPT-5.6, in three variants tuned for different jobs:
  • Sol - for demanding reasoning and long-horizon work.
  • Terra - frontier intelligence balanced against cost.
  • Luna - the fastest of the three, for cost-sensitive everyday work.
• Anthropic Claude Fable 5 - Anthropic's flagship model for the hardest problems.

On Fable 5 - in our last update we had to remove it shortly after launch, when a U.S. export control directive required Anthropic to suspend access for foreign nationals. We said we'd restore it the moment it became available again. It's back, and it's available to Pro and Premium users today.

We also welcomed two new model providers to Conception: Z.ai, with GLM-5.2, and Moonshot AI, with Kimi K2.6 - both strong at long-horizon reasoning and coding, and both a meaningful widening of the range of choice available to you. They join DeepSeek V4 Pro, Claude Sonnet 5, Claude Opus 4.8, Gemini 3.1 Pro, and the rest of the current lineup.

Just as importantly, we retired a long tail of older models - the GPT-4 series, Claude 3 and Claude 4 generations, Grok 3, Gemini 2.5, and others. Keeping them around padded the list without helping anyone choose. Every model in the picker now is one we'd actually recommend for something.

The picker itself is clearer, too. Each model carries its provider's icon, newly added models are badged, and anything outside your plan is marked up front rather than failing after you send. If you'd rather not think about any of this, Best stays at the top of the list and picks a model for each query on your behalf.

Attachments That Follow the Model

Images and PDFs now work properly in chat and search. Attach a screenshot, a scanned page, or a research paper, and models that can see them will read them directly as part of your question.

The file picker also adapts to the model you've selected. Not every model can read images, and not every model can read PDFs - so instead of letting you attach a file that would quietly fail, Conception now only offers the formats your chosen model can actually handle. We've also broadened the set of supported image formats along the way.

Guardrails: How We Test the AI

Adding a model to Conception used to be a judgement call. It isn't anymore.

Every model now has to pass an automated evaluation suite before it reaches you, and again whenever its provider updates it. We measure four things:

• Quality - each model is graded against a fixed set of real tasks, so a provider's silent update can't quietly make your results worse without us noticing.
• Safety - we run adversarial tests to confirm that a malicious webpage, a pasted document, or a crafted note can't hijack the assistant into ignoring its instructions or leaking your data.
• Speed - we benchmark how quickly each model starts responding, because a brilliant answer that takes thirty seconds is often the wrong trade.
• Cost - so we can keep capable models on the free tier and keep them there sustainably.

The results decide which models we ship, which model powers which part of the product, and which ones we drop. Several of the retirements above came directly out of this process - models that no longer earned their place once we had the numbers in front of us.

This is the same instinct behind last cycle's security work: the parts of Conception you can't see should be the parts you can most rely on. We'd rather tell you how we test the AI than ask you to take our word for it.