Model selection
Each pipeline stage reads from its own model slot, and picking the right model for each slot is where most of your cost-versus-quality trade-offs live. This page walks through the slots, the current registry, and a few configurations that tend to work well in practice.
The model slots
Aparture exposes six independent model slots plus one for the optional podcast add-on. Every slot ships with a default so you can run without touching them, and every slot can be swapped per run without restarting the server.
| Slot | What it drives | Default |
|---|---|---|
filterModel | Stage 2 quick filter (YES / MAYBE / NO triage) | gemini-3.1-flash-lite |
scoringModel | Stage 3 abstract scoring (0–10 + justification) | gemini-3-flash |
postProcessingModel | Stage 3.5 comparative re-scoring (optional) | gemini-3-flash |
pdfModel | Stage 4 deep PDF analysis | gemini-3.1-pro |
briefingModel | Stage 5 synthesis, the hallucination audit, and the refinement flow | gemini-3.1-pro |
quickSummaryModel | Per-paper quick-summary compression (text-only), runs just before synthesis | gemini-3.1-flash-lite |
A separate notebookLMModel slot drives the podcast-bundle generator in the podcast add-on — no default, you pick it per generation. The related quickSummaryConcurrency knob (default 5, clamped 1–20) controls how many quick-summary calls fire in parallel before synthesis.
The quickSummaryModel default is a small Flash-Lite model because the work is text compression (the input is already the text of each deep-analysis report, not the original PDF). If you're running single-provider and don't want a Google key alongside your main one, swap this slot to a small model in your chosen provider. Quick-summary failures are non-fatal in practice — a missing key quietly drops the inline-expansion feature but doesn't break the briefing.
The scoring rubric itself lives in the pipeline documentation — see Stage 3 scoring for what the scoringModel is actually asked to do.
Current lineup
The registry in utils/models.js is the authoritative source. As of April 2026 it holds 17 models across three providers. Prices below are per million tokens, current at time of writing.
Anthropic (Claude)
Current generation:
| User-facing ID | API ID | Context | Input / Output | Adaptive thinking |
|---|---|---|---|---|
claude-opus-4.7 | claude-opus-4-7 | 1M | $5 / $25 | Yes |
claude-opus-4.6 | claude-opus-4-6 | 1M | $5 / $25 | No |
claude-sonnet-4.6 | claude-sonnet-4-6 | 1M | $3 / $15 | No |
claude-haiku-4.5 | claude-haiku-4-5 | 200k | $1 / $5 | No |
Legacy (still registered, mostly kept for comparison):
| User-facing ID | API ID | Context | Input / Output |
|---|---|---|---|
claude-opus-4.5 | claude-opus-4-5 | 200k | ~$5 / ~$25 |
claude-opus-4.1 | claude-opus-4-1 | 200k | $15 / $75 (pricier) |
claude-sonnet-4.5 | claude-sonnet-4-5 | 200k | ~$3 / ~$15 |
claude-haiku-3.5 | claude-3-5-haiku-20241022 | 200k | ~$1 / ~$5 |
Opus 4.1 is worth flagging specifically because its pricing is higher than the newer Opus generations — the registry keeps it for backwards compatibility, but there's usually no reason to pick it over Opus 4.6 or 4.7.
OpenAI (GPT-5.4 family)
| User-facing ID | API ID | Context | Input / Output | Cached input |
|---|---|---|---|---|
gpt-5.4 | gpt-5.4 | 1M | $2.50 / $15 | $0.25 |
gpt-5.4-mini | gpt-5.4-mini | 400k | $0.75 / $4.50 | $0.075 |
gpt-5.4-nano | gpt-5.4-nano | 400k | $0.20 / $1.25 | $0.02 |
OpenAI caches automatically when prompt prefixes repeat, so the cached-input column tends to apply on later runs of the day that share a profile.
Google (Gemini)
Preview tier (Gemini 3.x). These are the defaults Aparture ships with — the user-facing IDs will stay stable when Google graduates them to general availability, even though the underlying API IDs will change.
| User-facing ID | API ID | Input / Output (≤200k) |
|---|---|---|
gemini-3.1-pro | gemini-3.1-pro-preview | $2 / $12 |
gemini-3-flash | gemini-3-flash-preview | $0.50 / $3 |
gemini-3.1-flash-lite | gemini-3.1-flash-lite-preview | $0.25 / $1.50 |
Stable tier (Gemini 2.5):
| User-facing ID | API ID | Input / Output (≤200k) |
|---|---|---|
gemini-2.5-pro | gemini-2.5-pro | $1.25 / $10 |
gemini-2.5-flash | gemini-2.5-flash | $0.30 / $2.50 |
gemini-2.5-flash-lite | gemini-2.5-flash-lite | $0.10 / $0.40 |
Google also offers a free tier with daily request caps on most of these models. For someone running a couple of briefings per day on a modest profile, the free tier often covers the whole workflow.
All 17 models support PDF content blocks, so any of them can technically drive the pdfModel slot. Speed, quality, and price are the real differentiators.
Per-stage recommendations
Each stage puts different pressure on the model. Matching the model to the pressure tends to be the cleanest way to keep quality up and costs down.
filterModel — fast, cheap, directional. Stage 2 produces three-bucket triage with a short summary. The prompt is 50–150 tokens per paper and you'll call it on everything fetched, so throughput and cost matter more than nuance. Good picks: gemini-3.1-flash-lite, gemini-2.5-flash-lite, claude-haiku-4.5, gpt-5.4-nano. Anything pricier than Haiku is usually overkill at this stage.
scoringModel — careful, calibrated, still batch-friendly. Stage 3 assigns a 0–10 score with a justification from the abstract, and downstream stages use those scores to pick what gets deep-analysed. Calibration matters more than at the filter stage, but it's still a per-abstract task that benefits from throughput. Good picks: gemini-3-flash, claude-sonnet-4.6, gpt-5.4-mini. Haiku 4.5 or a Flash-Lite model can work if your profile is very clear-cut.
postProcessingModel — same as or stronger than scoringModel. Stage 3.5 compares top papers head-to-head for calibration, so calibration matters more than throughput. Most configurations set it to the same model as scoringModel.
pdfModel — quality-sensitive, the most expensive slot. Stage 4 reads full PDFs, including figures, equations, and tables, and writes the structured deep analysis that feeds the briefing. Spending more here usually pays off in what lands in front of you. Good picks: gemini-3.1-pro, claude-opus-4.7, claude-sonnet-4.6, gpt-5.4. Cheaper fallback: gemini-3-flash or gemini-2.5-pro.
briefingModel — editorial judgment and long context. Stage 5 writes the executive summary, themes, and paper cards. It holds your profile, the day's quick summaries and deep analyses, and recent briefing history in one context, so quality here correlates strongly with what you actually read. Good picks: gemini-3.1-pro, claude-opus-4.7, claude-sonnet-4.6, gpt-5.4. Cheaper fallback: gemini-3-flash works if you're willing to accept slightly less polished prose.
Example configurations
Cost estimates below are rough — token counts vary with profile length, abstract length, and how many papers pass the filter. The "100-paper run" numbers assume 100 papers fetched, ~50 passing the filter to scoring, 20 reaching deep PDF analysis, and one briefing produced. The same scenario is used for the worked calculations on the API keys pages.
Three configurations worth spelling out: the Aparture default, a free-tier-only setup, and one mix across all three providers. Any model from the registry works in any slot — these are just common starting points.
Default (Aparture ships with this)
All-Google, current-generation Gemini 3.x. What DEFAULT_CONFIG sets.
| Slot | Model |
|---|---|
filterModel | gemini-3.1-flash-lite |
scoringModel | gemini-3-flash |
postProcessingModel | gemini-3-flash |
pdfModel | gemini-3.1-pro |
briefingModel | gemini-3.1-pro |
quickSummaryModel | gemini-3.1-flash-lite |
Ballpark cost for a 100-paper run: roughly $1.00–2.50. Every model here except gemini-3.1-pro is free-tier-eligible; the Pro slots require a paid Google tier.
Free tier (Gemini 2.5 throughout)
Every slot set to a stable Gemini 2.5 model. A fresh Google AI Studio account can run the full pipeline at no cost, subject to daily request caps. Slightly less polished than the 3.x previews, but fully capable of end-to-end briefings.
| Slot | Model |
|---|---|
filterModel | gemini-2.5-flash-lite |
scoringModel | gemini-2.5-flash |
postProcessingModel | gemini-2.5-flash |
pdfModel | gemini-2.5-pro |
briefingModel | gemini-2.5-pro |
quickSummaryModel | gemini-2.5-flash-lite |
Ballpark cost for a 100-paper run: effectively $0 within Google's free-tier caps, or roughly $0.50–1.20 on paid Tier 1. See Google API key setup for caps and signup.
Mixed providers
One way to spread work across all three providers: Gemini Flash-Lite on the high-volume filter, Claude Sonnet for careful abstract scoring, GPT-5.4 on deep PDF analysis, Gemini 3.1 Pro for briefing synthesis, and Claude Haiku compressing quick summaries. Requires keys for all three.
| Slot | Model |
|---|---|
filterModel | gemini-3.1-flash-lite |
scoringModel | claude-sonnet-4.6 |
postProcessingModel | claude-sonnet-4.6 |
pdfModel | gpt-5.4 |
briefingModel | gemini-3.1-pro |
quickSummaryModel | claude-haiku-4.5 |
Ballpark cost for a 100-paper run: roughly $1.75–3.00. Anthropic prompt caching trims scoring and quick-summary cost on repeat runs with the same profile.
This is one mix among many. Providers are interchangeable per slot, and the pipeline only checks that each slot's selected model has a valid key in .env.local — a missing key breaks only the stage that needs it, and the rest of the pipeline keeps running.
Adaptive thinking on Anthropic
Claude Opus 4.7 supports adaptive thinking: the model allocates more internal reasoning tokens for harder tasks and fewer for straightforward ones. Aparture passes thinking: {type: "adaptive"} on every Anthropic call, so picking Opus 4.7 turns it on automatically.
In practice, this matters most for briefingModel and pdfModel, where the model is asked for editorial judgment or structured deep analysis. Filter and scoring decisions are mechanical enough that thinking tokens don't buy much. Thinking tokens count against the output-token bill, so Opus 4.7 tends to cost more than Opus 4.6 on the same inputs — usually worth it for the briefing, often not worth it for the filter.
The default maxTokens for Anthropic calls is raised to 16000 to give thinking room to breathe.
Anthropic prompt caching
Aparture enables Anthropic's prompt caching automatically on every Anthropic call. On runs that share a profile, the repeated-prefix cost drops substantially — the second run of the day is usually noticeably cheaper than the first. The terminal log prints [anthropic cache] read=N create=N after each call if you want to confirm cache hits.
When the previews change
Gemini 3.x models are marked -preview in their API IDs. The user-facing IDs (gemini-3.1-pro, gemini-3-flash, gemini-3.1-flash-lite) will stay stable when Google graduates them to general availability; the API IDs will change. utils/models.js handles the mapping, so upgrades are a one-file change.
Snapshot taken 2026-04-19. Model lineup from utils/models.js. Pricing verified against the provider pricing pages below. Check these for current prices before making high-volume decisions:
- Anthropic: platform.claude.com/docs/en/docs/about-claude/models
- OpenAI: developers.openai.com/api/docs/models
- Google: ai.google.dev/gemini-api/docs/models
Next
The pipeline → — where each model slot slots in, and what each stage is actually asking for.
Also worth reading:
- API keys — set up the keys for the providers you picked.
- Tuning the pipeline — thresholds, batch sizes, and review gates to adjust alongside model choice.
- Reading a briefing — what the
briefingModelactually produces, so you can judge whether a slot change is helping.