Cheapest LLM API

Run a JavaScript async function in an isolated sandbox, composing catalog_search, model_detail, offer_compare, benchmark_results, provider_list, benchmark_list, coding_agent_list, coding_agent_detail, coding_agent_releases, coding_agent_upgrade, and price_watch as catalog.* methods. Return a small JSON value. Example: async () => { const page = await catalog.catalog_search({query: 'claude', page_size: 5}); return page.data.models.map(m => ({slug: m.slug, prices: m.bestOffer?.prices})); }. No network, secrets, or database binding is exposed. Shared worldwide execution budgets: 100 per UTC minute and 5000 per UTC day. Limits: 30 executions per client per minute (IPv6 /64), 20 catalog calls and read units (offer_compare and price_watch use 5), at most two concurrent calls, 5 seconds wall time, 50ms sandbox CPU, 32KiB code, 64KiB JSON output. JavaScript only; no imports or TypeScript syntax. Use /mcp?mode=code for only search + execute.

Discover exact benchmark IDs, editions, score units, configurations, provenance, and reviewed result counts before calling benchmark_results. Empty benchmarks are retained with resultCount 0.

Get reviewed scores for an exact benchmark_id from benchmark_list. Join result.modelId to catalog_search model.id. Compare scores only within the same benchmark edition and configuration. Includes source provenance and reported coding-agent usage when published. Follow pagination.hasMore.

Find canonical model slugs, context limits, and provider prices. Optional provider_slug filters availability, not the provider in bestOffer. bestOffer is selected for 10M input + 2M output tokens; use offer_compare for your workload. Follow pagination.hasMore.

Get a coding-agent profile by agent_slug from coding_agent_list, including official source, latest stable release, freshness, and history coverage. An unavailable archive is explicit in state. This does not assert product benchmark rankings or subscription prices.

List coding-agent profiles, interfaces, model access, official sources, latest stable release, archive counts, freshness, coverage limits, and sourced model/provider compatibility, BYOK, billing and setup facts. Includes agents whose release archive is unavailable; inspect state. Use coding_agent_releases to page through source-linked changelogs.

Search archived coding-agent releases by agent_slug, query (title, version, excerpt), topic, and channel. Stable only by default; channel all includes prereleases. Newest first; advance offset by limit while pagination.hasMore. Includes source URLs, abbreviated excerpts, and per-agent freshness/coverage. Topics are keyword mentions, not assessments. Partial coverage is explicit in meta.completeness; all unavailable returns an error. Archives reflect configured source history, not every product announcement.

Find archived changes published after an installed version or exact release ID. Get agent_slug and source_id from coding_agent_detail; default source_id is primary. Only the same publisher source and recognizable product stream are compared. Stable only by default; use channel all for prereleases. The ready response is publication order, not semantic-version precedence or an upgrade safety assessment; backports may appear. An unknown or ambiguous baseline is explicit in data.status, never treated as no changes. Follow data.pagination.hasMore with offset and limit.

Get capabilities, context/output limits, and current sourced provider offers by model_slug from catalog_search. Prices are base-context rates; long-context tiers, cache writes, batch discounts, and free-tier quotas can change actual cost.

Rank current chat offers for up to five canonical model slugs with the website calculator: context-tier prices, cache reads/writes, retries, supported batch rates, capability/benchmark requirements, provider/route preferences, and evidence filters. Input totals include cache shares; cache_hit_percent plus cache_write_percent must be <=100. Average request sizes determine tiers and capacity. Batch is Anthropic direct only. Unknown required evidence excludes offers; missing cache-read rates use standard input rates. Inspect costs, qualification, warnings, assumptions, excludedModels, and pagination.hasMore. Free quotas, non-token fees, and media token accounting are excluded.

Read current offers meeting a monthly baseline_usd minus minimum_saving_usd. Uses the same workload, qualifications and evidence filters as offer_compare. Omit model_slugs to watch all qualifying models. Returns up to 50 cheapest matches; no account or delivery subscription is created. The same query is available as RSS at /api/v1/price-watch/feed.xml; your RSS reader handles polling and notifications. Inspect offer warnings before switching.

Discover accepted provider slugs, roles, and current comparable model/offer counts. Provider minimum input and output prices are independent minima and may belong to different models; use offer_compare for an actual quote.