Model pricing and budget accounting

Reviewed 2026-10-04. USD per million tokens, standard text inference. Source of truth in code: packages/core/src/llm/pricing.ts. No paid built-in tools, batch, priority processing, explicit prompt caching or service-tier upgrades are requested.

Provider/model	Input	Output	Cached input	Cache write (5m / 1h where applicable)
Anthropic Sonnet 4.6	3	15	0.30	3.75 / 6
Anthropic Sonnet 5.5	2	10	0.20	2.50 / 4
Anthropic Opus 5.5	4	20	0.20	5 / 8
Anthropic Fable 5.1	10	50	0.25	12.50 / 20
Anthropic Haiku 4.5 (20251001)	1	5	0.10	1.25 / 2
OpenAI GPT-6.1 Sol	2	10	0.10	2.50
OpenAI GPT-6 Astra	10	50	1	12.50
OpenAI GPT-6 Luna	0.10	0.50	0.01	0.125
OpenAI GPT-5 Pro (legacy)	15	120	15	—
Gemini 3.8 Flash / 3.7 Flash	0.75	3.75	0.075	—
Gemini 3.1 Pro Preview	2	12	0.20	—

Sources: Anthropic pricing, OpenAI pricing, GPT-5 Pro, and Gemini pricing. Cache accounting follows OpenAI prompt caching; Gemini output limits include thoughts per the thinking guide.

For OpenAI GPT-6 requests above 272,000 input tokens, input, cache-read and cache-write rates double, and output increases 1.5×. Gemini 3.1 Pro Preview above 200,000 input tokens costs 4 input / 18 output / 0.40 cached input. Gemini 3.8/3.7 Flash promotional rates end on 2026-12-31; from 2027-01-01 the table automatically uses 1.50 input / 7.50 output / 0.15 cached input.

Reservation and actual spend

The default budget is $1 for every package. Every call, including a retry with error feedback, must fit before any HTTP request is made. The input estimate counts one token per UTF-8 byte of the entire serialized request (system prompt, schema, scoped context and retry feedback included), plus 8,192 tokens for provider framing. This is intentionally conservative; it is not the usual characters-divided-by-four heuristic. The full source file retained locally for patch validation is excluded from requests.

Each request caps output at 8,192 tokens. The reservation uses the highest applicable input/cache-write rate times estimated input, plus the capped output times its rate. Long-context tiers apply to the estimate. Integer nanodollar accounting rounds up. There are no hidden SDK retries. Missing usage after any HTTP failure, timeout or malformed response retains the complete reservation; the report separates that amount from spend confirmed by usage. API error bodies and authentication data are never printed.

Actual cost comes from each response’s usage: Anthropic regular input, cache reads, 5-minute/1-hour cache writes and output; OpenAI input, cached input, any cache-write usage and total output (already includes reasoning); Gemini prompt/cache tokens and candidate plus thinking tokens. Invalid or missing submit_patch is a paid failed attempt and becomes retry feedback. No response text is treated as a patch.

Unknown model IDs use the maximum of every listed rate for that provider, including long-context/cache-write rates, and print a warning. This still enforces the budget against the checked-in table. The table must be maintained when prices change; custom endpoints may charge different prices. The CLI cannot enforce an external service’s invoice or API contract. If returned usage exceeds the reserved token bounds, it reports actual usage and stops all further calls instead of hiding the discrepancy.

Live evaluation

The storefront results include verified zod/stripe runs for all three providers, the Sonnet 4.6 versus 5.5 effort comparison, exact API-echoed model IDs, and spend versus reservations. Sonnet 5.5 with medium effort is the Anthropic default. Its $2 input / $10 output per MTok rates were checked against the Sonnet 5.5 overview on 2026-10-04. The small fixture comparison does not establish a general quality ranking.

Offline CI replays wire fixtures, including a captured no-tool response. Live calls are opt-in, with environment keys only. Build core first, then use the evaluation helper against a clean, installed storefront checkout outside this repository:

pnpm --filter @uptide/core build
pnpm exec tsx scripts/eval-providers.ts --smoke --provider=anthropic
pnpm exec tsx scripts/eval-providers.ts --repo=/scratch/storefront --only=zod --provider=openai
pnpm exec tsx scripts/eval-providers.ts --repo=/scratch/storefront --only=stripe --provider=gemini
pnpm exec tsx scripts/eval-providers.ts --repo=/scratch/storefront --only=zod --provider=anthropic --model=claude-sonnet-5-5 --effort=high

Repeat both packages for each model/effort. --effort=high explicitly reproduces the Sonnet 5.5 API default; omission uses Uptide’s selected medium effort. Each run gets an isolated clone. Only aggregate metrics are emitted by this helper; keep any full local fix reports outside Git.