Anthropic API Pricing, Models & Rate Limits (2026)

1. 5-minute versus 1-hour cache-TTL decision surface

Formula: prefix write + (reuses − 1) × cache-read + output. No TTL price is substituted from the ordinary input rate; choose a TTL only after the documented minimum prefix and expiry rule are confirmed.

2. Cache + batch + long-context interaction matrix

Long-context thresholds are discrete: 200K and 201K are not interchangeable. A mechanism is applied only when its source supports that input size and traffic mode.

3. Direct Claude API versus Bedrock/Vertex parity audit

Verified 2026-08-14. Data owner: Luna. “Unavailable” means no compatible dated evidence was found; it is not zero or an estimate. Re-verify dated rates, specs, and policy before production use. First-party source · Run this scenario →

1. Tool-loop cache-placement ledger

Formula: system + tool-definition cache write + (steps − 1) × compatible cache reads + changing history + tool results + output. Cache placement, read/write rates, and tool-unit billing are Unavailable; no whole-prefix TTL curve is reused.

2. Thinking and final-output budget surface

Eligibility formula: declared input + thinking tokens + final output must fit context and max-output limits. Thinking tokens and any quality effect are not inferred from the final-output cap.

3. Message Batches completion ledger

Formula: submitted = 100; completed = submitted × (1 − failed/expired share); retried requests add a second compatible request bill. Turnaround, failure probability, and review success are sourced, user-supplied, or Unavailable.

Verified 2026-08-14. Data owner: Luna. “Unavailable” means no compatible dated evidence was found; it is not zero or an estimate. Source / registry · Run this scenario →

1. Context-editing and tool-result-clearing ledger

Formula / rule: net bill=post-edit input+output+replay Source: pricing registry verified 2026-08-26. Rate: Claude Sonnet 5, $2.0000 input/M + $10.0000 output/M.

2. Thinking-block integrity through tool loops

Formula / rule: integrity=order∧signature∧association∧reviewer Source: pricing registry verified 2026-08-26. Rate: Claude Sonnet 5, $2.0000 input/M + $10.0000 output/M.

3. Code-execution sandbox and artifact lifecycle audit

Formula / rule: artifact=exit∧hash∧download∧expiry Source: pricing registry verified 2026-08-26. Rate: Claude Sonnet 5, $2.0000 input/M + $10.0000 output/M.

1. Prompt-cache breakpoint invalidation ledger across a model-version upgrade

Formula / rule: Worst-case re-write cost = (full prefix tokens × input rate + response tokens × output rate)/1M at the Claude Sonnet 5 registry rate, applied if a dated snapshot upgrade forces a full cache miss. Whether a version bump actually invalidates the cache is Unavailable without a sourced Anthropic cache-versioning statement, so the figure below is a labelled worst-case bound, not an observed post-upgrade result. Source: pricing registry verified 2026-08-26.

2. Extended-output (128k) truncation-and-continuation ledger

Formula / rule: Continuation total = Σ(per-call token bill) across the frozen continuation chain at the Claude Sonnet 5 registry rate. Requested-versus-emitted length, stop-reason state, duplicated content, and reviewer acceptance require a matched long-generation run, which is not present in the registry, so only the chained-call cost below is reproducible. Source: pricing registry verified 2026-08-26.

3. Tool-use error-recovery loop cost audit

Formula / rule: Cost per resolved tool call = Σ(initial call + self-correction retries) token bill at the Claude Sonnet 5 registry rate ÷ resolved calls. Retry count and resolved-versus-abandoned outcome require a matched malformed-argument/unavailable-tool run, which is not present in the registry, so only the retry-budget cost below is reproducible. Source: pricing registry verified 2026-08-26.

1. Priority-versus-standard service-tier latency/cost tradeoff ledger

Formula / rule: Standard-tier cost = frozen prompt-set token bill at the Claude Sonnet 5 registry rate. The documented priority-tier price multiplier, measured or documented latency delta, and cost per accepted response at each tier require a sourced tier-rate card and a matched run, neither of which is present in the registry, so only the standard-tier base cost below is reproducible. Source: pricing registry verified 2026-08-26.

2. Image-input token-cost-by-resolution reconciliation

Formula / rule: Reconciliation requires Claude's documented image-tiling or token-estimation rule, a computed token count for each fixed resolution, and a returned image-token usage figure. None of these are present in the pricing registry, so every field below is Unavailable rather than estimated from the text-token rate, and this ledger is distinct from the equivalent OpenAI image-tiling ledger. Source: pricing registry verified 2026-08-26.

3. PDF page-count-to-token estimation accuracy audit

Formula / rule: Billed-page-token bill = frozen-document token bill at the Claude Sonnet 5 registry rate assuming a declared per-page token estimate. The declared per-page estimation method and its measured deviation from actual billed tokens across varying layout density require a matched document-ingestion run, which is not present in the registry, so only the assumed-estimate cost below is reproducible and no deviation figure is reported. Source: pricing registry verified 2026-08-26.

1. Admin API usage-and-cost-report reconciliation ledger

Formula / rule: Billed-sequence total = Σ(per-request token bill) at the Claude Sonnet 5 registry rate for the fixed request sequence. Whether the documented Admin API usage/cost-report endpoint reconciles exactly to this token-bill total, or reports on a different granularity/lag, requires a matched Admin API report pull, which is not present in the registry, so only the billed-sequence total below is reproducible. Source: pricing registry verified 2026-08-26.

2. 1M-context-window tier pricing delta ledger

Formula / rule: Standard-context cost = frozen prompt-set token bill at the Claude Sonnet 5 registry rate under the standard context-window tier. A documented separate per-token rate for the extended 1M-context tier is not present in the pricing registry, so only the standard-tier cost below is reproducible and no extended-tier premium is computed. Source: pricing registry verified 2026-08-26.

3. Remote MCP connector cost-attribution ledger

Formula / rule: Base-conversation cost = frozen prompt-set token bill at the Claude Sonnet 5 registry rate, excluding any remote MCP connector call. A documented per-call or per-token surcharge for a remote Model Context Protocol connector requires a matched connector-invocation run, which is not present in the registry, so only the base-conversation cost below is reproducible. Source: pricing registry verified 2026-08-26.

1. Admin API usage-and-cost-report reconciliation ledger

Formula / rule: Billed-sequence total = Σ(per-request token bill) at the Claude Sonnet 5 registry rate for the fixed request sequence. Whether the documented Admin API usage/cost-report endpoint reconciles exactly to this token-bill total, or reports on a different granularity/lag, requires a matched Admin API report pull, which is not present in the registry, so only the billed-sequence total below is reproducible. Source: pricing registry verified 2026-08-26.

2. 1M-context-window tier pricing delta ledger

Formula / rule: Standard-context cost = frozen prompt-set token bill at the Claude Sonnet 5 registry rate under the standard context-window tier. A documented separate per-token rate for the extended 1M-context tier is not present in the pricing registry, so only the standard-tier cost below is reproducible and no extended-tier premium is computed. Source: pricing registry verified 2026-08-26.

3. Remote MCP connector cost-attribution ledger

Formula / rule: Base-conversation cost = frozen prompt-set token bill at the Claude Sonnet 5 registry rate, excluding any remote MCP connector call. A documented per-call or per-token surcharge for a remote Model Context Protocol connector requires a matched connector-invocation run, which is not present in the registry, so only the base-conversation cost below is reproducible. Source: pricing registry verified 2026-08-26.

1. Token-efficient-tool-use versus standard-tool-loop ledger

Formula / scoring rule: Accepted-result cost = registry token bill plus any sourced tool charges, divided by an accepted equivalent answer. Tool definitions, visible results, cache read/write, repairs, latency, and fidelity must be joined from one matched run. Source: pricing registry verified 2026-08-27.

2. Message Batches malformed-row atomicity and partial-failure invoice audit

Formula / scoring rule: Submitted-volume baseline = registry bill for all declared input/output tokens. Per-row billing, invalid-row atomicity, retry subset, and cancellation treatment require a matched batch invoice and are not inferred. Source: pricing registry verified 2026-08-27.

3. Code-execution cold-container versus warm-container reuse ledger

Formula / scoring rule: Base cost = registry token bill for the fixed job sequence. Reuse is accepted only with startup, retained-state, isolation, timeout, artifact-hash, and tool-unit observations from the same matched run. Source: pricing registry verified 2026-08-27.

1. Web-search `max_uses` and domain-filter marginal-cost ledger

Formula / scoring rule: Cost per accepted answer = (model bill + charged searches + repair calls) ÷ reviewer-accepted answers. Configured `max_uses` is a ceiling, not an observed call count; filtering and errors need matched usage. Source: pricing registry and dated evidence index verified 2026-08-27.

2. Computer-use screenshot-loop cost frontier

Formula / scoring rule: Accepted-task cost = screenshot image-input units + action-loop model tokens + human-takeover/retry cost, divided by successful tasks. Resolution alone does not determine transformed image units. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Files API duplicate-upload, deletion, and post-delete-retention audit

Formula / scoring rule: Final bill = unique processed/storage units + re-upload processing + model token bill. Content-hash deduplication, deletion acknowledgement, citation availability, and retention clock require one dated Files audit. Source: pricing registry and dated evidence index verified 2026-08-27.

1. Completed streaming-versus-non-streaming usage parity

Formula / scoring rule: Parity requires equal terminal usage fields and semantic hash; bill = cached-read/creation + input + output units from the terminal event. Source: pricing registry and dated evidence index verified 2026-08-27.

2. Citation character/page-span integrity suite

Formula / scoring rule: Integrity score = exact quoted spans + valid offsets + source association − duplicate/overlap errors − unsupported claims; accepted cost uses matched returned usage. Source: pricing registry and dated evidence index verified 2026-08-27.

3. SDK automatic-retry invoice canary

Formula / scoring rule: Resolved-result cost = all invoice lines attributable to the terminal accepted result ÷ accepted result; retry status alone does not prove idempotency. Source: pricing registry and dated evidence index verified 2026-08-27.

1. Mixed 5-minute/1-hour prompt-cache breakpoint precedence ledger

Formula / scoring rule: At each reuse, bill = cache creation + cache reads + uncached tokens; the innermost changed block refreshes only its eligible scope after expiry. Source: pricing registry and dated evidence index verified 2026-08-27.

2. Priority-tier capacity-exhaustion and fallback canary

Formula / scoring rule: A result qualifies only when requested tier, returned tier, retry state, cache treatment, and terminal usage are joined; automatic fallback is not assumed. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Multimodal tool-result content-block ledger

Formula / scoring rule: Accepted cost = definitions + calls + result-block tokens/media + thinking/output + repair divided by accepted results; unsupported blocks stay unavailable. Source: pricing registry and dated evidence index verified 2026-08-27.

1. API-version and beta-header compatibility canary

Formula / scoring rule: Change one header; compare feature acceptance, schema, usage, latency, replay bill, and rollback evidence. Source: pricing registry and dated evidence index verified 2026-08-27.

2. Automatic-versus-forced tool choice and parallel-tool-use ledger

Formula / scoring rule: Accepted-result cost = definitions + call blocks + result resends + thinking/final usage + repairs; duplicate calls count as a failure. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Organization audit-event completeness and export-lag matrix

Formula / scoring rule: Completeness = joined event IDs / expected control changes; export lag = export time − event time; unsupported types remain named gaps. Source: pricing registry and dated evidence index verified 2026-08-27.

1. Citation-location-unit fidelity ledger

Formula / scoring rule: Accepted-answer cost = input/cache/output usage plus repair calls ÷ reviewer-accepted answer; every citation must join source-block index and character/page span. Source: pricing registry and dated evidence index verified 2026-08-27.

2. API-key creator, last-used, and orphan-cleanup attribution canary

Formula / scoring rule: Attributable spend window = usage between last successful request and cleanup decision, joined to creator/owner, workspace, audit event, blocked replay, and usage bucket. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Workspace deletion and residual-resource close ledger

Formula / scoring rule: Close = preconditions satisfied + cascade/retention verified + post-delete access blocked + late usage/export attributed + recovery path recorded. Source: pricing registry and dated evidence index verified 2026-08-27.

1. Concurrent prompt-cache creation stampede ledger

Formula / scoring rule: Total bill = creator write + reader hits/misses + retry fan-out; creator/read classification, cache-ready time, race failures, answer equivalence, and returned usage must join. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: Anthropic Claude Sonnet 3.7 / Haiku 3.5 pricing verified 2026-08-27; test suite: Batch 30 Anthropic cache-stampede fixture/test suite (run and result recorded 2026-08-27).

2. Message Batches cancellation, expiry, and result-download lifecycle ledger

Formula / scoring rule: Close = terminal row counts + custom IDs + result/error artifact availability + scoped retry + deletion timestamp + invoice/credit state. Retention is not inferred from a successful submit. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: Anthropic Claude Sonnet 3.7 / Message Batches registry rate verified 2026-08-27; test suite: Batch 30 Anthropic Message Batches lifecycle fixture/test suite (run and result recorded 2026-08-27).

3. Remote MCP authentication-expiry and refresh-recovery ledger

Formula / scoring rule: Accepted cost = connector setup + failed/replayed calls + model/cache/output usage; duplicate side effects and authorization scope must be proven. Connector or identity-provider fees are Unavailable. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: Anthropic Claude Sonnet 3.7 / remote-MCP usage registry rate verified 2026-08-27; test suite: Batch 30 Anthropic remote-MCP recovery fixture/test suite (run and result recorded 2026-08-27).

1. Token-count preflight versus Messages returned-usage reconciliation

Formula / scoring rule: Invoice delta = returned billed usage − preflight count; cache creation/read, system/tool overhead, and context-fit are separate fields. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: Anthropic Messages token-count and pricing registry, verified 2026-08-27.

2. Streaming disconnect-before-usage versus after-usage settlement ledger

Formula / scoring rule: Accepted cost = visible stream + final usage + server completion evidence + scoped retry; an unobserved debit is Unavailable, never zero. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: Anthropic streaming usage and pricing registry, verified 2026-08-27.

3. Message Batches duplicate-custom_id, input-order, and result-order conformance audit

Formula / scoring rule: Close = validation locus + accepted IDs + terminal/result-error association + returned usage + scoped retry + artifact integrity; result order need not equal input order. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: Anthropic Message Batches API and pricing registry, verified 2026-08-27.

1. Adaptive-versus-fixed thinking-effort ledger

Formula / scoring rule: Cost per accepted answer = input + cache + thinking + final output usage at the accepted parameter; an effort label is not a quality result. Anthropic Messages effort/thinking pricing evidence registry. Dated registry and evidence index, verified 2026-08-27.

2. Context-compaction trigger and fidelity ledger

Formula / scoring rule: Compaction decision = retained blocks + tool/citation continuity + context headroom; total bill requires returned compaction accounting. Anthropic context-window and compaction pricing evidence registry. Dated registry and evidence index, verified 2026-08-27.

3. Files API upload, processing, reference, expiry, delete, and residual-usage ledger

Formula / scoring rule: Close = file ID state + first/last reference + deletion propagation + storage/tool units + accepted artifact; residual charges are not presumed zero. Anthropic Files API and pricing evidence registry. Dated registry and evidence index, verified 2026-08-27.

1. Tool-search deferred-definition loading ledger

Formula / scoring rule: Discovery utility = selected-tool recall and avoided irrelevant definitions, conditioned on accepted search control and exact returned usage. First-party pricing/evidence registry: Anthropic tool search and Messages pricing evidence, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: Anthropic tool search documentationAnthropic pricing.

2. Programmatic-tool-calling execution and nested-call accounting canary

Formula / scoring rule: Accepted workflow cost = model/cache usage + visible tool calls + orchestrator execution/retry units; hidden intermediate work is not presumed free. First-party pricing/evidence registry: Anthropic programmatic tool calling and code execution evidence, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: Anthropic programmatic tool use documentationAnthropic pricing.

3. Memory-tool create, view, replace, delete, and cross-session carry-forward ledger

Formula / scoring rule: Memory acceptance = authorized operation + stale/deleted canary + resent-text usage + reviewer result; retention/storage pricing remains Unavailable when undocumented. First-party pricing/evidence registry: Anthropic memory tool, privacy, and Messages usage evidence, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: Anthropic memory tool documentationAnthropic pricing.

1. Fine-grained tool-argument streaming UTF-8/JSON reconstruction ledger

Formula / scoring rule: Accepted call = delta sequence + parser state + exact call ID + valid reconstructed JSON + terminal usage; repair bytes remain billable. Anthropic fine-grained tool streaming evidence; dated registry verified 2026-08-27; unsupported units fail closed as Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: Anthropic fine-grained tool streamingAnthropic pricing.

2. `pause_turn` server-tool continuation and settlement canary

Formula / scoring rule: Workflow bill = every returned model/cache/thinking/output unit across initial and continuation turns; pause is not completion. Anthropic pause_turn and server-tool continuation evidence; dated registry verified 2026-08-27; unsupported units fail closed as Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: Anthropic Messages APIAnthropic pricing.

3. `search_result` content-block citation, cache, and debit ledger

Formula / scoring rule: Claim acceptance = serialized result provenance + citation span integrity + cache placement + returned usage; supplied evidence is not provider search. Anthropic supplied search-result block evidence; dated registry verified 2026-08-27; unsupported units fail closed as Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: Anthropic citations documentationAnthropic pricing.

1. Web-fetch redirect, canonical URL, size, PDF, and max_uses ledger

Formula / scoring rule: Fetch acceptance = requested/final URL + accepted limits + extracted citation spans + fetch/model/cache usage + accepted claims + charge. Anthropic web-fetch boundary matched ledger; dated first-party registry verified 2026-08-27; unsupported units fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic web fetch documentationAnthropic pricing.

2. Structured-output grammar compile, cache reuse, and refusal settlement

Formula / scoring rule: Schema acceptance = schema compile/cache identity + validator result + refusal/overflow state + returned usage + accepted object + bill. Anthropic structured-output grammar matched canary; dated first-party registry verified 2026-08-27; unsupported units fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic structured outputs documentationAnthropic pricing.

3. Code-execution network and process-isolation ledger

Formula / scoring rule: Isolation acceptance = accepted/blocked operation + process/filesystem residue + artifact hash + timeout/restart + reviewer result + model/tool charge. Anthropic code-execution isolation matched ledger; dated first-party registry verified 2026-08-27; unsupported units fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic code execution documentationAnthropic pricing.

1. Output-effort and thinking-budget interaction ledger

Formula / scoring rule: Effort result = accepted/effective effort + thinking mode/budget + cache read/write + thinking/final usage + finish state + checker + latency + bill. Anthropic effort and thinking interaction matched ledger; dated first-party evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic Messages API documentationAnthropic model pricing.

2. Anthropic SDK tool-runner hidden-round and retry reconciler

Formula / scoring rule: Runner total = every logical/request ID + tool args/results + automatic retries/stop + per-round cache/thinking/input/output usage + side-effect checksum + final acceptance + invoice. Anthropic SDK tool-runner matched ledger; dated first-party evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic tool runner documentationAnthropic model pricing.

3. Remote MCP discovery pagination and schema-churn canary

Formula / scoring rule: MCP churn acceptance = catalog/page cursor + tool identity/version + serialized definition + stale-selection state + approval/auth + invocation/result IDs + rediscovery + side-effect checksum + charge. Anthropic remote MCP catalog matched canary; dated first-party evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic MCP connector documentationAnthropic model pricing.

1. Citation location-type and offset-integrity ledger

Formula / scoring rule: Citation integrity = source/block/page/character location + cited-substring hash + claim support + usage + malformed-location repair + reviewer acceptance + bill. Anthropic citation-location matched ledger; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic citations documentationAnthropic model pricing.

2. `cache_control` breakpoint-count and placement-boundary canary

Formula / scoring rule: Breakpoint result = serialized prefix + accepted/rejected breakpoint + cache creation/read input + expiry probe + equivalence + latency + invoice. Anthropic cache-breakpoint matched canary; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic prompt caching documentationAnthropic model pricing.

3. Overload and acceleration-limit recovery ledger

Formula / scoring rule: Recovery = request/status/error + rate-limit/retry headers + admission time + partial usage + backoff/replay + duplicate-effect check + accepted completion + charge. Anthropic overload-recovery matched ledger; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic rate limits documentationAnthropic model pricing.

1. `redacted_thinking` block preservation, replay, and debit ledger

Formula / scoring rule: Thinking continuity = block/message IDs + replay shape + signature/redaction state + cache creation/read + thinking/final usage + answer continuity + repair + acceptance + bill. Anthropic redacted-thinking matched ledger; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic extended thinking documentationAnthropic model pricing.

2. Web-fetch redirect, canonical, content-encoding, and partial-failure settlement canary

Formula / scoring rule: Fetch settlement = requested/final URL + fetch/result/citation spans + exposed bytes + cache/input/thinking/output/search usage + retry subset + accepted claims + charge. Anthropic web-fetch matched canary; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic web fetch documentationAnthropic model pricing.

3. Message Batches `custom_id` uniqueness, result-association, and partial-retry ledger

Formula / scoring rule: Batch association = batch/request/result/error IDs + identifier validation locus + completed/errored/expired/cancelled counts + usage + scoped replay + artifact hash + accepted coverage + invoice. Anthropic Message Batches identifier matched ledger; matched evidence/pricing registry verified 2026-08-27; unsupported fields fail closed as Unavailable. Dated first-party pricing/evidence registry, verified 2026-08-27. Module-local first-party sources: Anthropic Message Batches documentationAnthropic model pricing.