DeepSeek's composer teardown
Updated June 16, 2026
DeepSeek splits speed and depth across an Instant/Expert toggle above the bar and DeepThink/Search chips inside the card. The strategy is honest tiering: casual users see a calm default with web search already on, while power users opt into slower reasoning without leaving the same surface.
Calm default
What works
- Centered card on a blank canvas. The job is obvious: type a message.
- “Start chatting with Instant” ties the header to the active tier before the first keystroke.
- Search ships on by default in Instant. Research intent does not require hunting a toggle.
- DeepThink and Search live inside the card as peer chips, not buried in a + menu.
What we would push on
- Instant and Expert sit above the card while DeepThink sits inside it. Model tier and reasoning mode split across two surfaces.
- Search on by default may surprise users who expected a private draft answer.
Business strategy
DeepSeek competes on price and speed. A calm Instant default with search already armed signals “answers with sources” as the baseline product, not an upsell.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Search on by default in Instant | Research-ready first send without mode hunting | Users who want offline-style chat must notice and disable Search |
Takeaway
Steal the tiered header plus in-card capability chips. Make the default tier label and the armed chips scannable together before send.
Pattern: Tool Switching in Composer
Pattern: Model Selection UI
DeepThink
What works
- DeepThink reads as a mode, not a model name. Tooltip explains the job in plain language.
- Chip highlight inside the card shows armed state before send.
- DeepThink can stack with Search in Instant. Users can ask for reasoning plus sources in one prompt.
What we would push on
- Tooltip only appears on hover. Mobile users may never see “solve reasoning problems.”
- No latency or cost cue on the chip itself when DeepThink is active.
Business strategy
DeepThink lets DeepSeek compete on hard questions without forcing everyone onto Expert. Reasoning becomes an opt-in tax users understand before they spend time waiting.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| DeepThink as an in-card chip with hover tooltip | Clear reasoning contract without a second composer | Cost and latency invisible until after send |
Takeaway
Label reasoning modes by outcome in tooltips and keep them as chips users can arm and disarm in one glance.
Pattern: Input Mode ToggleReasoning is a pre-send chip with outcome copy in the tooltip, not a separate “thinking” product.
Pattern: Tool Switching in Composer
Instant vs Expert
What works
- Segmented Instant/Expert control uses speed vs quality metaphors (lightning vs diamond), not model IDs.
- Header copy tracks the toggle. Users always know which tier they are about to use.
- Expert mode simplifies the card: DeepThink remains, Search disappears. Tier changes the capability set.
What we would push on
- No subtext on Expert about latency, cost, or when to switch back to Instant.
- Search vanishing in Expert may confuse users who toggled tiers mid-session.
Business strategy
Instant keeps the free-tier feeling fast and searchable. Expert is the quality lane for users willing to trade speed, which supports upgrade storytelling without a separate paywall screen.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Instant/Expert toggle above the composer | Honest speed vs depth choice before typing | Capability set changes between tiers without an explicit changelog |
Takeaway
Pair tier toggles with header copy that updates in sync. When tiers drop features, say why in the toggle or card.
Pattern: Model Selection UI
Pattern: Tool Switching in Composer
Attach limits in Instant
What works
- Attach limits appear on demand via tooltip, not a blocking modal.
- Copy is specific: text extraction only in Instant, max 50 files, 100MB each.
- DeepThink and Search stay visible while the tooltip is open. Context is not lost.
What we would push on
- Limits surface on hover over paperclip. Discoverability depends on curiosity.
- “Text only” in Instant may frustrate users who uploaded images expecting vision.
Business strategy
Capping multimodal attach in Instant protects compute on the default tier while nudging heavy file users toward Expert or paid plans without a hard error on first click.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Tier-specific attach limits in a tooltip | Keeps Instant fast and cheap to serve | Vision expectations may break on first upload |
Takeaway
When tiers limit modalities, surface limits before upload completes, not only on icon hover.
Pattern: Cost Transparency
File in the card
What works
- File chip lives inside the card with filename and parsing state. Users see work in progress before send.
- Modes stay armed during upload. Attach does not reset DeepThink or Search.
- Send stays available once parsing completes. One surface for file plus prompt.
What we would push on
- Parsing spinner with no time estimate. Large files may feel stuck.
- No preview thumbnail for images during parse. Hard to verify the right file.
Business strategy
In-card parsing keeps users in the compose loop instead of routing to a separate upload modal, which reduces drop-off for document Q&A flows.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Inline file chip with parsing state | File plus modes visible in one pre-send view | Card height grows and parsing feedback is minimal |
Takeaway
Show parse progress and a removable chip inside the composer. Pair with tier limits so users know what the file will be used for.
Pattern: Context Chip Management
Expert armed
What works
- Header, tier toggle, and DeepThink chip align before send. No hidden Expert state.
- Typed prompt stays editable with mode chips still one tap away.
- Send affordance stays prominent on the right. The card does not morph into a form wizard.
What we would push on
- No combined cue for “Expert + DeepThink” cost or expected wait time.
- Attach icon absent in this Expert capture. Tier may still change attach rules without a reminder.
Business strategy
Expert plus DeepThink is the quality stack for hard prompts. Keeping it in the same card rewards power users who outgrow Instant without teaching a new UI.
Tradeoff
| Decision | Benefit | Cost |
|---|---|---|
| Stack Expert tier with DeepThink chip | Depth without a separate reasoning product | Wait and cost opaque until the answer streams |
Takeaway
Let users stack tier and reasoning chips, but add pre-send cues when the combination is slow or expensive.
Pattern: Model Selection UI
Pattern: Input Mode Toggle
Steal this
- Instant/Expert toggle with header copy that tracks the active tier
- In-card DeepThink and Search chips with outcome tooltips
- Search on by default for a research-ready Instant tier
- Tier-specific attach limits explained before upload finishes
- Inline file chip with parsing state inside the composer
Skip this
- Splitting tier toggle above the card and reasoning chips inside without linking them
- Hiding attach limits behind paperclip hover only
- Dropping Search in Expert with no explanation when users switch tiers
- Reasoning and tier changes with no latency or cost preview
How others design the composer
Same job, different product bets, and what each tradeoff reveals.
ChatGPTChatGPT hides modes behind + and starter pills. DeepSeek puts Instant/Expert and DeepThink/Search in the open on a single card.
Read teardown
ClaudeClaude defaults web search on with model on the right. DeepSeek pairs a tier toggle above the card with reasoning chips inside it.
Read teardown
GrokGrok keeps a minimal bar with Fast on the right. DeepSeek centers a card with tier and reasoning controls before the first message.
Read teardownOriginal gallery pages: Tool Switching in Composer