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 any slot can be swapped per run without restarting the server. Five have drop-downs in the Settings panel; postProcessingModel is the exception — it lives only in the saved config, with no Settings control.
| Slot | What it drives | Default |
|---|---|---|
filterModel | Stage 2 quick filter (YES / MAYBE / NO triage) | gemini-3.5-flash-lite |
scoringModel | Stage 3 abstract scoring (0–10 + justification) | gemini-3.6-flash |
postProcessingModel | Stage 3.5 comparative re-scoring (optional) | gemini-3.6-flash |
pdfModel | Stage 4 deep PDF analysis | gemini-3.6-flash |
briefingModel | Stage 5 synthesis, the hallucination audit, and the refinement flow | gemini-3.6-flash |
quickSummaryModel | Per-paper quick-summary compression (text-only), runs just before synthesis | gemini-3.5-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 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 (it now carries per-model pricing fields too). As of July 2026 it holds 23 models across three providers. Prices below are per million tokens, list price, snapshot 2026-07.
Anthropic (Claude)
| User-facing ID | API ID | Context | Input / Output | Adaptive thinking |
|---|---|---|---|---|
claude-fable-5 | claude-fable-5 | 1M | $10 / $50 | Always on |
claude-opus-5 | claude-opus-5 | 1M | $5 / $25 | Yes |
claude-opus-4.8 | claude-opus-4-8 | 1M | $5 / $25 | Yes |
claude-opus-4.7 | claude-opus-4-7 | 1M | $5 / $25 | Yes |
claude-opus-4.6 | claude-opus-4-6 | 1M | $5 / $25 | Yes |
claude-sonnet-5 | claude-sonnet-5 | 1M | $3 / $15 | Yes |
claude-sonnet-4.6 | claude-sonnet-4-6 | 1M | $3 / $15 | Yes |
claude-haiku-4.5 | claude-haiku-4-5 | 200k | $1 / $5 | No |
Opus 5 is the current Opus flagship (registered July 2026), at the same $5 / $25 as the 4.x Opus line — a free upgrade from Opus 4.8 for any slot. Fable 5 sits above it as Anthropic's most capable widely released model, at double the price; see the caveats below before reaching for it. Sonnet 5 launched with introductory pricing ($2 / $10 through 2026-08-31, then $3 / $15 — the table shows the post-intro list price).
Two things to know about claude-fable-5 specifically:
- Its thinking can't be turned off. Aparture sends
thinking: {type: "adaptive"}on every Anthropic call, which Fable 5 accepts — but unlike the Opus and Sonnet models, there's no way to trade thinking tokens away for a cheaper run. At $50 / MTok output, thinking tokens are the expensive kind. - It requires 30-day data retention on your Anthropic organization. Under zero data retention, every Fable 5 request fails with a 400 that reads like a malformed payload rather than a policy rejection — so if Fable 5 is the one model that 400s for you and the others work, check your org's retention setting before debugging the request.
Anthropic also ships Claude Mythos 5, which shares Fable 5's specs and pricing, but it's invitation-only through Project Glasswing with no self-serve signup, so it isn't in the registry. If you have access, adding it is a two-line change in utils/models.js (copy the claude-fable-5 entry, change the ID).
The pre-4.6 legacy entries (Opus 4.5/4.1, Sonnet 4.5, Haiku 3.5) were removed from the registry in July 2026: Haiku 3.5's API ID was retired upstream (selecting it returned a live 404), Opus 4.1 retires 2026-08-05, and the others predate adaptive thinking. Saved configs pointing at any of them are migrated automatically to the closest current-generation equivalent.
OpenAI (GPT)
GPT-5.6 family (GA 2026-07-09, current generation):
| User-facing ID | API ID | Input / Output |
|---|---|---|
gpt-5.6-sol | gpt-5.6-sol | $5 / $30 |
gpt-5.6-terra | gpt-5.6-terra | $2.50 / $15 |
gpt-5.6-luna | gpt-5.6-luna | $1 / $6 |
GPT-5.4 family (previous generation, still fully served with no deprecation notice):
| 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)
Both models added in the July 2026 refresh became defaults: Gemini 3.6 Flash for the scoring, post-processing, PDF, and briefing slots, and Gemini 3.5 Flash-Lite for the filter and quick-summary slots.
| User-facing ID | API ID | Input / Output |
|---|---|---|
gemini-3.6-flash | gemini-3.6-flash | $1.50 / $7.50 |
gemini-3.5-flash | gemini-3.5-flash | $1.50 / $9.00 |
gemini-3.5-flash-lite | gemini-3.5-flash-lite | $0.30 / $2.50 |
gemini-3.6-flash matches 3.5 Flash on input price and undercuts it on output ($7.50 vs $9.00), so it's a straight cost reduction anywhere output volume matters — PDF analysis and briefing synthesis most of all. That's why it's now the default for those slots. Existing installs are not migrated: a saved config keeps whatever you already selected, since gemini-3.5-flash still works fine and silently rewriting a deliberate choice would be worse than leaving a few cents on the table. Switch it in Settings if you want the cheaper output price.
gemini-3.5-flash-lite is a more capable Lite model, and it is not the cheapest one — at $0.30 / $2.50 it lists above gemini-3.1-flash-lite ($0.25 / $1.50). It is nonetheless the default for the filter and quick-summary slots, because the percentage gap is misleading at this scale: both slots are input-dominated and tiny in absolute terms, so on the reference 100-paper run the upgrade costs about 1.7¢ — roughly 1% of total run cost. Paying that for a better model on the stage that decides what the rest of the pipeline even looks at is a good trade.
gemini-3.1-flash-lite is still registered and still the cheapest 3.x option. Switch the Lite slots back to it if you're optimising the last cent, or running high enough volume that the filter stage stops being rounding error.
Gemini 3.x tier. Pro and Flash are still preview-only upstream (Google has already shut down sibling previews, so treat these as churn-prone); Flash-Lite reached GA in May 2026. The user-facing IDs stay stable when Google graduates the remaining previews — only the API IDs 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 | $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 23 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-3.5-flash-lite (the default), gemini-3.1-flash-lite (cheapest 3.x), gemini-2.5-flash-lite, claude-haiku-4.5, gpt-5.6-luna, 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-3.6-flash, gemini-3.5-flash, claude-sonnet-5, 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-3.6-flash (cheapest output of the frontier Flash models), gemini-3.5-flash, claude-opus-5, claude-sonnet-5, gpt-5.6-sol. Cheaper fallback: gemini-2.5-pro. claude-fable-5 will work here, but this slot runs once per paper, so its 2× premium multiplies across the whole deep-analysis set — it's the least cost-effective place to spend it.
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-3.6-flash, gemini-3.5-flash, claude-opus-5, claude-sonnet-5, gpt-5.6-sol. Cheaper fallback: gemini-2.5-flash works if you're willing to accept slightly less polished prose. This is the one slot where claude-fable-5 is arguably worth its price: it runs once per briefing rather than once per paper, so the premium is bounded, and editorial judgment is exactly what it's strongest at.
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. The app computes its own estimates from the registry's pricing fields: the score-review and pre-briefing gates show the projected spend of the next stage, and the end-of-run summary shows a per-stage cost estimated from the token counts the run actually used.
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, GA models only. What DEFAULT_CONFIG sets.
| Slot | Model |
|---|---|
filterModel | gemini-3.5-flash-lite |
scoringModel | gemini-3.6-flash |
postProcessingModel | gemini-3.6-flash |
pdfModel | gemini-3.6-flash |
briefingModel | gemini-3.6-flash |
quickSummaryModel | gemini-3.5-flash-lite |
Ballpark cost for a 100-paper run: roughly $1.00–1.90 at Gemini 3.6 Flash's $1.50 / $7.50 list pricing — in the same band as the previous all-3.x-preview default (roughly $1.00–2.50), since 3.5 Flash undercuts the Gemini 3.1 Pro that used to hold the PDF and briefing slots.
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-5.6 Terra on deep PDF analysis, Gemini 3.6 Flash for briefing synthesis, and Claude Haiku compressing quick summaries. Requires keys for all three.
| Slot | Model |
|---|---|
filterModel | gemini-3.5-flash-lite |
scoringModel | claude-sonnet-5 |
postProcessingModel | claude-sonnet-5 |
pdfModel | gpt-5.6-terra |
briefingModel | gemini-3.6-flash |
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 and Sonnet models from 4.6 onward — plus Opus 5 and Fable 5 — support adaptive thinking: the model allocates more internal reasoning tokens for harder tasks and fewer for straightforward ones. Aparture passes thinking: {type: "adaptive"} automatically on those models, and omits it on Haiku (which rejects it).
Which models get the parameter is driven by a supportsAdaptiveThinking flag on each Anthropic entry in utils/models.js, not by parsing the model name. That matters for Fable 5: the adapter's name-based fallback only recognises claude-opus-* and claude-sonnet-*, so an unregistered claude-fable-5 would be treated as a non-thinking model, and the adapter would force tool_choice — which Fable 5 rejects outright, because its thinking is always on and forced tool choice is invalid whenever thinking is enabled. Any future Anthropic model outside the Opus/Sonnet naming convention needs a registry entry for the same reason.
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 — 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 Pro and Flash are still marked -preview in their API IDs; Flash-Lite reached GA in May 2026 and its API ID is now gemini-3.1-flash-lite (the old -preview alias has since been shut down upstream). Preview IDs carry real shutdown risk — Google has already retired sibling previews — which is why Aparture's shipped defaults moved to GA models in July 2026. The user-facing IDs stay stable across these transitions — only the API IDs change. utils/models.js handles the mapping, so upgrades are a one-file change.
Snapshot taken 2026-07-28. Model lineup and pricing from utils/models.js (pricing fields inputPerMTok/outputPerMTok, snapshot 2026-07). 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.