AI & API Cost Reference

Overview

This page is the authoritative reference for what every external paid API call in the codebase costs. It has three layers:

The audit — every paid call site, its exact model id, and how it is invoked.
The pricing — verified per-model rates, pulled from vendor pricing pages.
The math — per-stage, per-pipeline, and corpus-scale cost projections.

Verified: August 14, 2026. Every model id below was read out of the code on that date; every price was read from the vendor’s own pricing page on that date. Where a number is an estimate rather than a measurement, it is labelled (est.).

The older Cost Tracking page covers the same ground as of July 3, 2026 and also documents fixed monthly subscriptions and Trigger.dev credit budgeting. Where the two disagree on a model or a price, this page is correct — see What changed.

Part 1 — The audit

Model registries

Four files decide which model runs. Change a model here and it changes everywhere.

File	Controls
`apps/api/src/config/ai.ts`	`MODELS`, `FEATURE_MODELS`, `MODEL_TIERS`, `GEMINI_IMAGE_MODELS`, `getReasoningOptions()` — everything the Hono API calls
`packages/trigger/src/lib/anthropic/models.ts`	`CLAUDE_PIPELINE_MODELS`, max-token budgets, and the hardcoded Anthropic Batch price table
`packages/trigger/src/lib/gemini/clients.ts`	`EXTRACTION_MODEL`, `IMAGE_MODEL`, `IMAGE_MODEL_FAST`, `IMAGE_MODEL_PRO`, `VIDEO_MODEL_OMNI_FLASH`
`packages/ai-prompts/src/models.ts`	Legacy metadata + its own cost table (advisory only; nothing routes through it)

packages/ai-prompts/src/models.ts carries a stale cost table of its own. Its costPerInputToken / costPerOutputToken fields are per 1K tokens and are not read by any live call path. Do not treat it as a pricing source — it is model metadata that happens to include numbers.

How calls reach a provider

Three distinct mechanisms are in use, and the distinction matters for cost:

Vercel AI Gateway

gateway("openai/gpt-5.6-luna") or getModelForFeature(). Zero markup — you pay provider list price. All Luna, and all Claude routed through the API app.

Direct provider SDK

anthropic.messages.create() and the GoogleGenAI client. Full list price, no gateway involved. Used by the strain pipeline’s Claude passes, all Gemini image generation, and Lyria.

Anthropic Batch API

anthropic.messages.batches.create(). 50% off input and output. Only reachable in the strain pipeline’s batch mode.

Paid call sites

Reasoning effort is cost-critical. GPT-5.6 Luna bills reasoning tokens as output ($1.20/1M), and they are invisible in the response. Luna’s provider default is medium. Every call site below pins an explicit effort — the Effort column is the single biggest per-call cost lever in the codebase.

Strain research pipeline

File:line	Stage	Service	Model	Mechanism	Effort
`lib/exa.ts:363`	01 Search	Exa	`search`, `type: "auto"`, `livecrawl: "fallback"`	Direct SDK	—
`lib/firecrawl.ts`	03 Scrape	Firecrawl	`asyncBatchScrapeUrls`	Direct SDK	—
`lib/gemini/extraction.ts:380`	04 Extract	OpenAI	`gpt-5.6-luna`	Gateway	`low`
`lib/anthropic/processing.ts:210`	05 FactFinder ×3	Anthropic	`claude-haiku-4-5-20251001`	Direct SDK	—
`lib/anthropic/processing.ts:271`	05 Copywriter	Anthropic	`claude-sonnet-5`	Direct SDK	—
`lib/anthropic/processing.ts:341`	05 Image Prompts	Anthropic	`claude-sonnet-5`	Direct SDK	—
`lib/anthropic-batch.ts:537-608`	05 (batch mode)	Anthropic	same three models	Batch API, −50%	—
`lib/gemini/image-generation.ts`	06 Imaging	Google	`gemini-3.1-flash-image` @ 2K	Direct SDK	—

All paths under packages/trigger/src/. Stage 05 runs direct by default: strain-research-pipeline.ts:62 sets useDirectMode to true, so the 50% Batch discount does not apply to a normal self-serve run.

The openai/ prefix on EXTRACTION_MODEL lives in a directory called lib/gemini/. Text extraction moved off Gemini Flash-Lite onto Luna on Aug 13 2026 but the directory was never renamed. lib/gemini/extraction.ts does not call Gemini — only image-generation.ts and clients.ts still do.

Other pipelines

File:line	Pipeline	Service	Model	Mechanism	Effort
`lib/paper-summarizer.ts:361`	Paper summarize	OpenAI	`gpt-5.6-luna`	Gateway	`low`
`lib/paper-summarizer.ts:525`	Paper (2nd pass)	OpenAI	`gpt-5.6-luna`	Gateway	`none`
`lib/paper-quality.ts:654`	Paper quality gate	Anthropic	`claude-haiku-4-5-20251001`	Direct SDK	—
`lib/paper-image-generator.ts:208`	Paper thumbnail	Google	`gemini-3.1-flash-image` @ 2K, 16:9	Direct SDK	—
`tasks/strain-music/config/ai-prompt-builder.ts:197`	Music prompt	OpenAI	`gpt-5.6-luna`	Gateway	`low`
`tasks/strain-music/03-generate.ts:134`	Music generate	Google	`lyria-3-pro-preview`	Direct SDK	—
`tasks/strain-video/02-scene-director.ts:238`	Video scenes	OpenAI	`gpt-5.6-luna`	Gateway	`low`
`tasks/strain-video/providers/gemini-omni-flash.ts:53`	Video clips	Google	`gemini-omni-flash-preview`	Direct SDK	—
`tasks/social-content/02-generate-text.ts:159`	Social text	OpenAI	`gpt-5.6-luna`	Gateway	`low`
`tasks/social-content/03-generate-images.ts`	Social images	Google	`gemini-3.1-flash-image`	Direct SDK	—
`tasks/gmail/gmail-classify-task.ts:381`	Gmail classify	OpenAI	`gpt-5.6-luna`	Gateway	`none`
`tasks/notebook-media/mindmap-first-party.ts:278`	Notebook mindmap	OpenAI	`gpt-5.6-luna`	Gateway	`none` → `low` on retry
`tasks/feature-notebook-pipeline.ts:406`	Feature notebook	OpenAI	`gpt-5.6-luna`	Gateway	`medium`
`tasks/twitter/twitter-generate-tweet.ts:34`	Tweet	Anthropic	`claude-haiku-4-5-20251001`	Direct SDK	—
`lib/side-effects-generator.ts:114`	Side effects	OpenAI	`gpt-5.6-luna`	Gateway	`low`
`lib/shopping-recommendations.ts:132`	Shopping recs	Anthropic	`claude-sonnet-4-6`	Direct SDK	—
`lib/firecrawl-agent.ts:620`	Menu scan	Anthropic	`claude-sonnet-4-6`, `max_tokens: 16384`	Direct SDK	—

Hono API features

File:line	Feature	Model	Mechanism	Effort
`services/ai-sdk/label-scanner.ts:343`	Label scan pass 1	`gpt-5.6-luna` (vision)	Gateway	`medium`
`services/ai-sdk/label-scanner.ts:418`	Label scan pass 2	`gpt-5.6-luna` (vision)	Gateway	`medium`
`services/ai-sdk/receipt-parser.ts:536`	Receipt parse	`mistral-small-latest`	Gateway	`medium` (ignored by Mistral)
`services/ai-sdk/strain-matcher.ts:70`	Strain match	`gpt-5.6-luna`	Gateway	`none`
`services/ai-sdk/section-streaming.ts:98`	Notebook sections	tier-resolved (below)	Gateway	—
`services/professor-high/agent.ts:105`	Professor High agent	`claude-sonnet-4.6`	Gateway	—
`routes/chat/index.ts:544`	Chat (Professor High)	`claude-sonnet-4.6`	Gateway	—
`services/ai-sdk/deep-research.ts:479,702`	Deep research	`claude-sonnet-4.6`	Gateway	—
`services/ai-sdk/menu-notebook.ts:44`	Menu notebook	`claude-sonnet-4.6`, `maxOutputTokens: 4000`	Gateway	—
`services/blog/`, `services/hero-images/`, `services/discover/`, `services/app-landing/`	Image generation	`gemini-3.1-flash-image`	Direct REST, not gateway	—

No embeddings anywhere. A repo-wide sweep for embed(, .embeddings., and vector-model ids returns nothing in apps/api/src, packages/trigger/src, or apps/mobile/convex. Similarity and matching are done with SQL and string heuristics, not vectors — so there is $0 embedding spend to account for.

getReasoningOptions() returns { openai: { reasoningEffort } }. Non-OpenAI providers ignore the openai key, so it is harmless to spread into a Mistral or Claude call — the medium on the receipt parser has no effect and no cost.

Notebook section tiers

A full notebook is 23 sections: 2 computed (free) and 21 AI-billed, resolved through MODEL_TIERS in apps/api/src/config/ai.ts. Section definitions and their estimatedOutputTokens live in packages/ai-prompts/src/config/section-config.ts.

Tier	Model	Sections	Est. output tokens
`creative`	`claude-opus-4.8`	7	4,300
`analysis`	`claude-sonnet-4.6`	7	11,500
`simple`	`gpt-5.6-luna`	7	5,900
`computed`	none	2	0

Things that look paid but are not

gpt-4o-mini-tts — retired, not a live call site

NOTEBOOK_ARTIFACT_MODEL_IDS in packages/report-contracts/src/artifact-plan.ts:88 still contains 'gpt-4o-mini-tts', and its own comment explains why: notebook narration is retired and no executor references the model. The value survives only because the enum is persisted to notebookArtifacts.executor.modelId and a Zod enum cannot be empty. There is no TTS call anywhere in the codebase — no audio/speech, no ElevenLabs, no .speech. call. This line costs $0.

Delisted OpenAI models still in the registry

MODELS in apps/api/src/config/ai.ts still lists gpt-5, gpt-5-mini, o3, and o4-mini. gpt-5 and gpt-5-mini are annotated in the file as delisted by OpenAI in June 2026. No FEATURE_MODELS entry points at any of them, so none can be reached.

Mistral Large / Magistral Medium

Present in MODELS, referenced by no feature. Note also that magistral-medium-2506 no longer appears on Mistral’s pricing page at all — the medium tier is now Mistral Medium 3.5.

SCRAPE_CONFIG.maxUrls is dead config

config/pipeline.config.ts declares maxUrls: 15, but tasks/stages/03-scrape.ts never reads it. Every filtered URL is scraped — which is why a ranked strain measures ~36 scrapes, not 15. Budget against the measured number, not the config.

Part 2 — Verified pricing

All rates in USD per 1M tokens unless stated otherwise. Read from vendor pricing pages on August 14, 2026.

Text and vision models

Model	Input	Cached in	Output	Batch in/out	Notes
`gpt-5.6-luna`	$0.20	$0.02	$1.20	—	Cut 80% from $1/$ 6 on Jul 30 2026
`gpt-5.6-terra`	$2.00	$0.20	$12.00	—	Cut 20% same day; unused
`gpt-5.6-sol`	$5.00	$0.50	$30.00	—	Unchanged; unused
`gpt-5`	$1.25	$0.125	$10.00	—	Delisted Jun 2026
`gpt-5-mini`	$0.25	$0.025	$2.00	—	Delisted Jun 2026
`gpt-5-nano`	$0.05	$0.005	$0.40	—	Unused
`o3`	$2.00	$0.50	$8.00	—	Unused
`o4-mini`	$1.10	$0.275	$4.40	—	Unused
`claude-opus-4.8`	$5.00	$0.50	$25.00	$2.50 /$ 12.50
`claude-sonnet-5`	$2.00	$0.20	$10.00	$1.00 /$ 5.00	See warning below
`claude-sonnet-4.6`	$3.00	$0.30	$15.00	$1.50 /$ 7.50
`claude-haiku-4.5`	$1.00	$0.10	$5.00	$0.50 /$ 2.50
`gemini-3.7-flash`	$0.75	—	$3.75	—	Rises Jan 1 2027
`gemini-3.5-flash-lite`	$0.30	—	$2.50	—
`mistral-small-latest`	$0.15	—	$0.60	—	Mistral Small 4

Claude Sonnet 5’s introductory pricing became permanent. Anthropic’s pricing page now states that the

2/

10 launch rate “is now the standard price” and that the scheduled increase to $3/$ 15 on September 1, 2026 will not occur.Our code has not caught up. packages/trigger/src/lib/anthropic/models.ts hardcodes SONNET_5_STANDARD_PRICING_START = Date.UTC(2026, 8, 1) (September 1 2026) and will begin reporting $1.50/$ 7.50 batch instead of the correct $1.00/$ 5.00 from that date. This is a cost-reporting bug only — Anthropic bills correctly regardless — but every batch-mode cost figure the pipeline logs after Sep 1 will be 50% too high.

Gemini Flash pricing doubles on January 1, 2027. The

0.75/

3.75 rate is explicitly labelled “through December 31, 2026”, rising to $1.50 /$ 7.50 starting January 1, 2027. Context caching doubles on the same date. We route no production traffic to gemini-3.7-flash today, so the exposure is latent — but the label scanner’s documented fallback (FEATURE_MODELS.labelScanning reverting to gemini-3.7-flash) would land on the doubled rate if exercised after that date.

Claude 4.7-and-later models use a new tokenizer that produces ~30% more tokens for the same text. This affects claude-sonnet-5 (our copywriter and image-prompt passes) and claude-opus-4.8 (notebook creative sections). claude-sonnet-4.6 and claude-haiku-4.5 use the previous tokenizer. Sonnet 5’s headline rate is lower than Sonnet 4.6’s (

2/

10 vs

3/

15) but ~30% more tokens claws back roughly a third of that saving. Every Sonnet 5 and Opus 4.8 figure on this page has the 30% uplift applied.

Image, audio, video

Service	Model	Unit	Price
Google	`gemini-3.1-flash-image`	per image, 1K	$0.067
Google	`gemini-3.1-flash-image`	per image, 2K (what we use)	$0.101
Google	`gemini-3.1-flash-image`	per image, 4K	$0.151
Google	`gemini-3.1-flash-image`	text/image input	$0.50 / 1M
Google	`gemini-3.1-flash-lite-image`	per image, 1K only	$0.0336
Google	`gemini-3-pro-image`	per image, 1K–2K	$0.134
Google	`lyria-3-pro-preview`	per full song	$0.08
Google	Lyria 3 Clip	per 30s clip	$0.04
Google	Veo 3.1	per second	$0.05 –$ 0.60
OpenAI	`gpt-4o-mini-tts`	text in / audio out	$0.60 /$ 12.00 per 1M — retired, $0 spend

Search, scraping, compute

Service	Unit	Price
Exa	search request (covers first 10 results)	$7 / 1k
Exa	each result beyond 10	$1 / 1k results
Exa	`/contents` per page, per content type	$1 / 1k pages
Firecrawl	scrape / crawl	1 credit per page
Firecrawl	browser `interact`	2 credits per browser-minute
Firecrawl	Standard plan	$83/mo yearly → 100k credits → $ 0.00083/credit
Vercel AI Gateway	markup	$0.00 — no markup, no platform fee
Trigger.dev	run invocation	$0.000025
Trigger.dev	`small-1x` / `medium-1x` / `medium-2x` / `large-1x`	$0.0000338 /$ 0.000085 / $0.00017 /$ 0.00034 per sec

Checkpointed waits are free. Trigger.dev does not bill compute for any wait.for / wait.until longer than 5 seconds, which is why a 5-minute pipeline run bills for only ~5 minutes of active machine time.

One number we could not pin down: whether Exa bills contents on top of search. lib/exa.ts:363 calls exa.search() with a text: { maxCharacters: 5000 } contents option. Exa’s pricing page documents /contents at “

1 / 1k pages, per content type" but **does not state whether that charge applies to contents requested inline with `/search`**. Every Exa figure below uses the **conservative-low** reading (search + extra-results only). If contents *do* bill, add **

0.025/strain** unranked and $0.040/strain ranked — about +4% on a total run. Confirm against a live Exa invoice before relying on the lower figure.

Part 3 — Per-run cost math

Assumptions

Measured from production (not estimates)

Unranked strain: 25 results searched → 6–16 scraped → 6–16 extractions. 11 used as midpoint.
Ranked strain: 40 searched → ~36 scraped → ~34 extractions.
Extraction truncates article content at 50,000 chars ≈ 12,500 tokens per article (lib/gemini/extraction.ts:30).
Stage timings, typical unranked run (ms): search 3,200 · filter 1,300 · scrape 78,000 · extract 35,000 · process 150,000 · imaging 40,000 · save 1,100.
Ranked refresh with skipImaging+skipMusic (ms): search 4,630 · filter 2,055 · scrape 304,735 · extract 76,414 · process 86,625 · imaging 142 · save 1,897.
One paper summary measured at **1,345 in / 649 out = $0.00105** — which reproduces to the cent at Luna's$ 0.20/$1.20, independently confirming the rate.
Machine presets read from code: orchestrator medium-1x; stages 03/04 medium-2x; stages 05/06 large-1x; stages 01/02/07 default small-1x.

Estimated (token counts we do not measure)

Extraction: ~13,300 input tokens/article (12,500 content + ~800 prompt + schema); ~2,000 output tokens/article including low reasoning. (est.)
FactFinder: ~12,000 input/pass unranked, ~30,000 ranked; ~2,500 output/pass against a 4,096 cap. (est.)
Copywriter: ~15,000 input unranked / ~30,000 ranked; ~6,000 output against an 8,192 cap. (est.)
Image prompts: ~6,000 input; ~2,500 output against a 4,096 cap. (est.)
Notebook: ~4,000 input tokens per section (order context + strain data + prior sections). (est.)
A 30% token uplift is applied to all claude-sonnet-5 and claude-opus-4.8 figures for the new tokenizer.
Orchestrator billed at ~10s of active compute; the rest of its wall-clock is checkpointed and free. (est.)

(a) Typical unranked strain — direct mode

Stage	Service / model	Units	Cost
01 Search	Exa, 25 results	1 search + 15 extra results	$0.0220
02 Filter	local	—	$0.0000
03 Scrape	Firecrawl	11 credits	$0.0091
04 Extract	`gpt-5.6-luna`, effort `low`	146k in + 22k out	$0.0557
05a FactFinder ×3	`claude-haiku-4.5`	36k in + 7.5k out	$0.0735
05b Copywriter	`claude-sonnet-5`	19.5k in + 7.8k out	$0.1170
05c Image Prompts	`claude-sonnet-5`	7.8k in + 3.3k out	$0.0481
06 Imaging	`gemini-3.1-flash-image`	2 × 2K images	$0.2040
07 Save	Supabase RPC	—	$0.0000
Compute	Trigger.dev, 8 runs / ~309s	—	$0.0850
Total			$0.614

(b) Ranked strain — 40 results, direct mode

Stage	Service / model	Units	Cost
01 Search	Exa, 40 results	1 search + 30 extra results	$0.0370
02 Filter	local	—	$0.0000
03 Scrape	Firecrawl	36 credits	$0.0299
04 Extract	`gpt-5.6-luna`, effort `low`	452k in + 68k out	$0.1720
05a FactFinder ×3	`claude-haiku-4.5`	90k in + 7.5k out	$0.1275
05b Copywriter	`claude-sonnet-5`	39k in + 7.8k out	$0.1560
05c Image Prompts	`claude-sonnet-5`	7.8k in + 3.3k out	$0.0481
06 Imaging	`gemini-3.1-flash-image`	2 × 2K images	$0.2040
Compute	Trigger.dev, 8 runs / ~540s	—	$0.1307
Total			$0.905

(c) Ranked refresh — `skipImaging` + `skipMusic`

Stage	Cost
01 Search (Exa, 40)	$0.0370
03 Scrape (36 credits)	$0.0299
04 Extract (Luna)	$0.1720
05a FactFinder ×3 (Haiku 4.5)	$0.1275
05b Copywriter (Sonnet 5)	$0.1560
05c Image Prompts (Sonnet 5)	$0.0481
06 Imaging	skipped — $0.0000
Compute (measured 476s active)	$0.0956
Total	$0.666

Skipping imaging removes

0.204 of model spend and

0.0136 of large-1x compute — a 26% cut on a ranked run. It is by far the cheapest optimization available, and the right default for any re-research pass where the existing hero and mascot images are still good.

Where the money goes

Provider	Unranked	Share	Ranked	Share
Anthropic (Claude)	$0.239	39%	$0.332	37%
Google (Gemini images)	$0.204	33%	$0.204	23%
Trigger.dev compute	$0.085	14%	$0.131	14%
OpenAI (Luna)	$0.056	9%	$0.172	19%
Exa	$0.022	4%	$0.037	4%
Firecrawl	$0.009	1%	$0.030	3%

Luna is not the cost problem. Even on a ranked strain doing 34 extractions over ~450k input tokens, Luna accounts for 19% of the run. Two Gemini images cost more than every Luna token combined on an unranked strain. The optimization targets, in order, are: image resolution, the Sonnet 5 copywriter, and large-1x compute on stage 05.

Music pipeline

Auto-fires after research completes unless skipMusic is set.

Component	Cost
Prompt builder (`gpt-5.6-luna`, effort `low`)	$0.0028
`lyria-3-pro-preview`, full song	$0.0800
Trigger.dev compute (~5 runs, ~80s)	$0.0060
Total	$0.089

Combined per strain

Profile	Research	+ Music	Combined
Unranked, direct	$0.614	$0.089	$0.703
Ranked, direct	$0.905	$0.089	$0.994
Ranked refresh (`skipImaging`+`skipMusic`)	$0.666	—	$0.666
Unranked refresh (`skipImaging`+`skipMusic`)	$0.397	—	$0.397

Extrapolations

All figures below are estimates, built on the token assumptions listed above. They are accurate to roughly ±20% on the model-spend lines and better than that on the fixed per-unit lines (images, songs, Exa, Firecrawl). The blended column assumes 30% of strains carry a popularity rank and therefore take the 40-result path — verify that ratio against strains_v2 before treating a blended total as a budget.

Cost to research N strains (research + music)

Strains	All unranked	All ranked	Blended (30% ranked)
100	$70.32	$99.40	$79.04
500	$351.59	$497.01	$395.22
1,000	$703.19	$994.02	$790.44
2,750 (full corpus)	$1,933.77	$2,733.55	$2,173.71

Cost to RE-research the full corpus

Re-research keeps existing images and audio, so skipImaging + skipMusic apply throughout.

Scope	Per strain	2,750 strains
All unranked-shaped	$0.397	$1,091.20
All ranked-shaped	$0.666	$1,831.75
Blended (30% ranked)	$0.478	$1,313.37

A full corpus refresh at ~ $1,313** costs 40% less than a fresh full research pass at ~$ 2,174 — entirely because imaging and music are skipped. If a refresh does need new images, add $0.204 × 2,750 = **$ 561.

Per-unit costs — other pipelines

Unit	Composition	Cost
Notebook (full, 21 AI sections)	Opus 4.8 ×7 + Sonnet 4.6 ×7 + Luna ×7	$0.591
Professor High reply, 8-step worst case	Sonnet 4.6, agentic tool loop	$0.318
Professor High reply, typical 3 steps	Sonnet 4.6, agentic tool loop	$0.086
Deep research run	Sonnet 4.6, up to 5 tool calls	$0.255
Strain song	Luna prompt + Lyria 3 Pro + compute	$0.089
Menu notebook	Sonnet 4.6, 4k output cap	$0.090
Menu scan	Sonnet 4.6 (16k cap) + Firecrawl credits	$0.182
Research paper	Luna summary + Haiku gate + 1 × 2K thumbnail	$0.109
Shopping recommendations	Sonnet 4.6	$0.105
Social post, 1 image	Luna text + 1 × 2K image	$0.104
Social carousel, 4 images	Luna text + 4 × 2K images	$0.407
Blog article image	1 × 2K Gemini image	$0.101
Label scan	Luna vision ×2 passes, effort `medium`	$0.0077
Tweet	Haiku 4.5	$0.0045
Side-effects backfill	Luna, effort `low`	$0.0022
Receipt parse	Mistral Small 4	$0.0012
Gmail classify, 100-email batch	Luna, effort `none`	$0.0066
Gmail classify, per email	Luna, effort `none`	$0.00007
Strain match	Luna, effort `none`	$0.0007

Professor High is the only per-message cost on this page, and it is agentic. MAX_AGENT_STEPS = 8 (services/professor-high/agent.ts:56), and each step re-sends the full accumulated context to Sonnet 4.6. A single chat reply therefore ranges from $0.086 to$ 0.318 depending on how many tool calls it takes — and a chatty session multiplies that per message, not per session. It is the only feature here whose cost scales with user engagement rather than with content volume. Model spend on chat should be watched separately from pipeline spend.

Notebook breakdown

Tier	Model	Sections	Cost	Share
`creative`	`claude-opus-4.8`	7	$0.322	54%
`analysis`	`claude-sonnet-4.6`	7	$0.257	43%
`simple`	`gpt-5.6-luna`	7	$0.013	2%
`computed`	—	2	$0.000	0%

Seven Luna sections cost 2% of a notebook. Seven Opus sections cost 54%. Moving even two creative sections down to Sonnet 4.6 would cut a notebook by roughly 15% — the single highest-leverage model change on this page.

Could not be priced

Item	Why
Strain video pipeline (`gemini-omni-flash-preview`)	Google publishes no per-second or per-video rate for `gemini-omni-flash-preview`. Veo 3.1 spans $0.05–$ 0.60/sec, a 12× range, so no defensible estimate is possible. Measure a real run before budgeting video.
Exa inline contents	See the warning in Part 2 — vendor docs do not state whether search-inline `text` bills the `/contents` rate. Range given instead.
Firecrawl `interact` per menu scan	`lib/firecrawl-agent.ts:769` comments “7 credits/min”; Firecrawl’s current page says 2 credits/browser-minute. The code comment is stale, and session duration is not instrumented.

What changed since the July page

Item	July 3, 2026 page	Now (Aug 14, 2026)
Extraction model	`gemini-3.1-flash-lite`	`gpt-5.6-luna` via Gateway, effort `low`
Copywriter / image prompts	`claude-sonnet-4-6`	`claude-sonnet-5` (+30% tokenizer)
Stage 05 machine	`medium-2x`	`large-1x`
Orchestrator machine	`small-1x`	`medium-1x`
Exa cost/strain	$0.007	$0.022 unranked /$ 0.037 ranked — the old figure ignored per-result billing beyond 10
Compute/strain	$0.021	$0.085 unranked /$ 0.131 ranked — stage 05 is `large-1x` and runs ~150s
Ranked strains	not modelled	40-result path, ~34 extractions
Per strain (research)	$0.571	$0.614 unranked /$ 0.905 ranked
Per strain (+ music)	$0.658	$0.703 unranked /$ 0.994 ranked

The headline per-strain number moved less than the underlying changes suggest, because two large moves cancelled: extraction got much cheaper (Luna replaced Flash-Lite) while compute and Exa were both materially undercounted before.

How to update this page

Re-check this page whenever any of the following change. Pricing drift is silent — nothing in CI will catch it.

Re-read the model registries

These four files are the only places a model id is decided. If a model id changed, every number downstream is wrong.

File	What to look for
`apps/api/src/config/ai.ts`	`MODELS`, `FEATURE_MODELS`, `MODEL_TIERS`, `GEMINI_IMAGE_MODELS`
`packages/trigger/src/lib/anthropic/models.ts`	`CLAUDE_PIPELINE_MODELS`, `CLAUDE_PIPELINE_MAX_TOKENS`, `ANTHROPIC_BATCH_PRICING`, `getAnthropicBatchPricing()`
`packages/trigger/src/lib/gemini/clients.ts`	`EXTRACTION_MODEL`, `IMAGE_MODEL*`, `VIDEO_MODEL_OMNI_FLASH`
`packages/ai-prompts/src/config/section-config.ts`	Section count and `modelTier` / `estimatedOutputTokens` per section

Re-check every reasoningEffort call site

Reasoning tokens bill as output. A call site that loses its explicit effort silently falls back to Luna’s medium default and can multiply that call’s output cost.

grep -rn "getReasoningOptions\|reasoningEffort" \
  apps/api/src packages/trigger/src apps/mobile/convex \
  | grep -v "\.test\."

Every hit should either be the definition in config/ai.ts or a call site with an explicit effort. A bare generateText against Luna with no providerOptions is a cost bug.

Re-pull vendor pricing

Vendor	Page
OpenAI	developers.openai.com/api/docs/pricing
Anthropic	platform.claude.com/docs/en/about-claude/pricing
Google Gemini + Lyria	ai.google.dev/gemini-api/docs/pricing
Mistral	mistral.ai/pricing/api
Exa	exa.ai/docs/reference/pricing
Firecrawl	firecrawl.dev/pricing
Trigger.dev	trigger.dev/pricing
Vercel AI Gateway	vercel.com/docs/ai-gateway/pricing

Watch these dated cliffs

Jan 1, 2027 — Gemini Flash input/output doubles ( $0.75/$ 3.75 → $1.50/$ 7.50). Context caching doubles too.
Sep 1, 2026 — no longer a real price change (Anthropic cancelled it), but getAnthropicBatchPricing() still thinks it is. Fix that function or every batch-mode cost log overstates by 50%.
Any OpenAI repricing of gpt-5.6-luna, which moved 80% within three weeks of launch and is the most volatile rate on this page.

Re-measure, don't re-estimate

The token counts in Part 3 are estimates. If a run’s real usage is available — Trigger.dev run metadata, an Anthropic usage block, or an AI Gateway spend report — prefer it and mark the row as measured. The paper-summarizer row is the model: a real 1,345/649 measurement that reproduced the vendor rate exactly.

Version history

Date	Change
August 14, 2026	Initial publication. Full call-site audit; all pricing re-verified from vendor pages; Luna/Sonnet 5/`large-1x`/Exa-per-result corrections applied; ranked and refresh profiles added; `gpt-4o-mini-tts` identified as retired.

​Overview

​Part 1 — The audit

​Model registries

​How calls reach a provider

Vercel AI Gateway

Direct provider SDK

Anthropic Batch API

​Paid call sites

​Strain research pipeline

​Other pipelines

​Hono API features

​Notebook section tiers

​Things that look paid but are not

​Part 2 — Verified pricing

​Text and vision models

​Image, audio, video

​Search, scraping, compute

​Part 3 — Per-run cost math

​Assumptions

​(a) Typical unranked strain — direct mode

​(b) Ranked strain — 40 results, direct mode

​(c) Ranked refresh — skipImaging + skipMusic

​Where the money goes

​Music pipeline

​Combined per strain

​Extrapolations

​Cost to research N strains (research + music)

​Cost to RE-research the full corpus

​Per-unit costs — other pipelines

​Notebook breakdown

​Could not be priced

​What changed since the July page

​How to update this page

​Version history

Overview

Part 1 — The audit

Model registries

How calls reach a provider

Paid call sites

Strain research pipeline

Other pipelines

Hono API features

Notebook section tiers

Things that look paid but are not

Part 2 — Verified pricing

Text and vision models

Image, audio, video

Search, scraping, compute

Part 3 — Per-run cost math

Assumptions

(a) Typical unranked strain — direct mode

(b) Ranked strain — 40 results, direct mode

(c) Ranked refresh — `skipImaging` + `skipMusic`

Where the money goes

Music pipeline

Combined per strain

Extrapolations

Cost to research N strains (research + music)

Cost to RE-research the full corpus

Per-unit costs — other pipelines

Notebook breakdown

Could not be priced

What changed since the July page

How to update this page

Version history