OpenAI API Pricing, Models & Rate Limits (2026)

1. API invoice-surface reconciler · fixed text + tool workload

Formula: token amount = (input $/M × input tokens + output $/M × output tokens) ÷ 1,000,000. The workload bill is therefore a sum of compatible token units; tool, search, storage, and media units remain separate.

2. Pinned-model versus alias drift ledger

Decision boundary: a price or lifecycle change never promotes an unpinned alias alone. The exact canary must match model ID, prompt, tool state, output cap, and replay cost.

3. Portfolio capacity envelope · sequential worker results

Worker rows are sequential fixed-token calculations. Published parallel quota/SLA coverage is Unavailable; no concurrency capacity is inferred from serial speed.

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. Stateful versus stateless request ledger

Formula: token bill = (input $/M × resent input + output $/M × output) ÷ 1,000,000. Retained server state, storage, and tool units are not counted as zero; their accounting is Unavailable.

2. Service-class eligibility-and-bill matrix

3. Organization/project/key chargeback canary

Chargeback formula: allocated invoice = sum of compatible line items grouped by model ID, endpoint, service class, cached tokens, tools, failed requests, budget alert, and quota. Budget-alert and quota evidence remain Unavailable; replay the fixed 20-request portfolio before promotion.

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. Remote MCP-and-connector permission canary

Formula / rule: pass=discovery∧scope∧host∧approval∧no duplicate side effect Source: pricing registry verified 2026-08-26. Rate: GPT-5.4, $2.5000 input/M + $15.0000 output/M.

2. Versioned prompt-object rollout ledger

Formula / rule: bill=(input×input$/M+output×output$/M)/1M; shadow is a control Source: pricing registry verified 2026-08-26. Rate: GPT-5.4, $2.5000 input/M + $15.0000 output/M.

3. Server-side context-compaction fidelity gate

Formula / rule: fidelity=constraints∧tools∧citations∧open task∧reviewer acceptance Source: pricing registry verified 2026-08-26. Rate: GPT-5.4, $2.5000 input/M + $15.0000 output/M.

1. Fine-tuning job lifecycle and per-epoch cost ledger

Formula / rule: Accepted-job cost = Σ(training-token consumption × per-epoch training rate) + checkpoint/deployment fees; gpt-5.4 has no dated per-epoch fine-tuning training rate in the pricing registry, so training, checkpoint, and cancelled/failed-job partial-consumption cost are all Unavailable rather than assumed zero. Deployed-model inference after a completed job bills at the base gpt-5.4 registry rate shown in row three of module three. Source: pricing registry verified 2026-08-26.

2. Audio-endpoint (transcription/speech) duration-to-unit billing reconciliation

Formula / rule: Bill = returned duration or character count × declared per-minute/per-character rate; OpenAI audio-endpoint per-minute and per-character rates are not present in the pricing registry, so there is no rate card to reconcile a returned duration against, and rounding behavior for failed/retried segments is equally unsourced. Source: pricing registry verified 2026-08-26.

3. Parallel tool-choice determinism canary

Formula / rule: Cost per accepted tool-call set = (frozen prompt+schema tokens × input rate + response tokens × output rate)/1M at the gpt-5.4 registry rate. Selected-tool-set order, duplicate/omitted calls, and schema-validity failures across repeated identical requests require a matched run, which is not present in the registry, so only the base-request cost below is reproducible. Source: pricing registry verified 2026-08-26.

1. Reasoning-effort (low/medium/high) cost and token-accounting ledger

Formula / rule: Final-answer-only floor cost = (frozen prompt tokens × input rate + visible-completion tokens × output rate)/1M at the gpt-5.4 registry rate, excluding the separately-billed reasoning-token count. Per-effort reasoning-token consumption and latency require a matched run at each effort level, which is not present in the registry, so total cost including reasoning tokens is Unavailable at every tier. Source: pricing registry verified 2026-08-26.

2. Embeddings-endpoint cost-per-dimension and truncation ledger

Formula / rule: The pricing registry carries no dated OpenAI embeddings-endpoint per-token rate or documented dimension-reduction billing rule for gpt-5.4-class embedding models, so cost-per-dimension and the default-versus-reduced-dimension token-to-vector comparison are both Unavailable rather than derived from the completion-token rate. Source: pricing registry verified 2026-08-26.

3. Spend-limit and usage-tier-escalation interaction ledger

Formula / rule: Base-request cost = frozen-request-sequence token bill at the gpt-5.4 registry rate for the stated volume. Whether crossing a documented cumulative-spend milestone changes the rate-limit ceiling or unlocks batch/fine-tuning eligibility, and the documented behavior of in-flight requests at a hard spend cap, require a sourced escalation rule and a matched account-tier run, neither of which is present in the registry, so only the base-request cost below is reproducible. Source: pricing registry verified 2026-08-26.

1. Predicted Outputs cost-and-latency delta ledger

Formula / rule: Baseline cost = (frozen prompt tokens × input rate + completion tokens × output rate)/1M at the gpt-5.4 registry rate, no prediction supplied. The rejected-prediction-token surcharge rate, the accepted-prediction latency delta, and the accepted-versus-rejected token split for a supplied prediction require a matched Predicted Outputs run, which is not present in the registry, so only the no-prediction baseline below is reproducible. Source: pricing registry verified 2026-08-26.

2. Multi-project rate-limit pooling ledger

Formula / rule: Base-request cost = frozen fixed-project-set token bill at the gpt-5.4 registry rate for one billed organization. Whether documented per-project rate limits draw from one shared organization-level pool or are enforced independently per project requires a sourced rate-limit-pooling policy document, which is not present in the registry, so pooling behavior below is Unavailable and named rather than assumed. Source: pricing registry verified 2026-08-26.

3. Vector-store storage-and-query billing ledger

Formula / rule: The pricing registry carries no dated OpenAI vector-store per-GB storage rate or per-query retrieval rate, so both storage cost and query cost are Unavailable rather than derived from the completion-token rate, and this ledger is distinct from the standard file-search/retrieval-tool token cost. Source: pricing registry verified 2026-08-26.

1. Predicted Outputs cost-and-latency delta ledger

Formula / rule: Baseline cost = (frozen prompt tokens × input rate + completion tokens × output rate)/1M at the gpt-5.4 registry rate, no prediction supplied. The rejected-prediction-token surcharge rate, the accepted-prediction latency delta, and the accepted-versus-rejected token split for a supplied prediction require a matched Predicted Outputs run, which is not present in the registry, so only the no-prediction baseline below is reproducible. Source: pricing registry verified 2026-08-26.

2. Multi-project rate-limit pooling ledger

Formula / rule: Base-request cost = frozen fixed-project-set token bill at the gpt-5.4 registry rate for one billed organization. Whether documented per-project rate limits draw from one shared organization-level pool or are enforced independently per project requires a sourced rate-limit-pooling policy document, which is not present in the registry, so pooling behavior below is Unavailable and named rather than assumed. Source: pricing registry verified 2026-08-26.

3. Vector-store storage-and-query billing ledger

Formula / rule: The pricing registry carries no dated OpenAI vector-store per-GB storage rate or per-query retrieval rate, so both storage cost and query cost are Unavailable rather than derived from the completion-token rate, and this ledger is distinct from the standard file-search/retrieval-tool token cost. Source: pricing registry verified 2026-08-26.

1. Responses API versus Chat Completions text-request invoice-parity ledger

Formula / scoring rule: Exact bill = (effective input tokens × input rate + effective output tokens × output rate)/1M at the gpt-5.4 registry rate. Endpoint parity is accepted only when usage, hidden endpoint tokens, finish state, latency, and replay scope match; SDK compatibility is not billing evidence. Source: pricing registry verified 2026-08-27.

2. Web-search `search_context_size` marginal-cost and citation-yield canary

Formula / scoring rule: Baseline bill uses the model registry rate; search units, retrieved-source tokens, repair searches, and accepted citations must come from the same matched run and are never inferred from tool availability. Source: pricing registry verified 2026-08-27.

3. Image-generation tool size/quality/format billing reconciliation

Formula / scoring rule: Token baseline is reproducible at the gpt-5.4 registry rate, but total image-tool bill requires a dated output-image rate and delivered-artifact invoice. Requested pixels never substitute for delivered pixels. Source: pricing registry verified 2026-08-27.

1. Realtime `truncation=auto` versus retention-ratio cache-economics ledger

Formula / scoring rule: Total bill = retained text/audio units × applicable input rates + uncached units + replayed tool-state units + output units; retained-prefix ratio = retained units ÷ prior-turn units. Continuity is accepted only from the matched session. Source: pricing registry and dated evidence index verified 2026-08-27.

2. Web-search empty-result, tool-error, and retry billing canary

Formula / scoring rule: Accepted-answer cost = (model token bill + charged search-call units + retry bill) ÷ accepted answers. Empty-result, timeout, provider-error, citation yield, and retry scope must come from the same canary; availability is not a call count. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Code Interpreter container session-duration and reuse ledger

Formula / scoring rule: Session bill = container/session units + model token bill + retained-artifact/retry charges. Reuse is reported only when the dated container rate, documented session window, expiry, cold start, and artifact hash are all observed. Source: pricing registry and dated evidence index verified 2026-08-27.

1. HTTP failure-class token-debit and invoice ledger

Formula / scoring rule: Total bill = returned input/output usage × dated model rates + separately sourced tool units; a missing body is never treated as zero usage. Source: pricing registry and dated evidence index verified 2026-08-27.

2. Evals candidate-versus-grader cost ledger

Formula / scoring rule: Job bill = upload/storage units + Σ(candidate tokens + grader tokens + tool units + retry units) across completed cases; score coverage = accepted scored cases ÷ submitted cases. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Moderation batching and decision-cost canary

Formula / scoring rule: Cost per accepted decision = attributable endpoint bill ÷ human-accepted decisions; qualify only when endpoint eligibility, item limits, and matched false-allow/false-block review are present. Source: pricing registry and dated evidence index verified 2026-08-27.

1. Organization Usage API late-arrival and pagination reconciliation ledger

Formula / scoring rule: Variance = usage-export total − joined invoice lines; close only when request IDs, UTC buckets, cursors, duplicate rows, and late adjustments reconcile. Source: pricing registry and dated evidence index verified 2026-08-27.

2. Multipart Uploads API abort-and-orphan accounting canary

Formula / scoring rule: Upload cost = visible accepted bytes/parts + sourced File/storage units; orphan bytes are never inherited from vector-store rates. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Responses include-field expansion ledger

Formula / scoring rule: Compare identical prompts by returned input/output/reasoning usage and payload bytes; extra response bytes are not presumed billed tokens. Source: pricing registry and dated evidence index verified 2026-08-27.

1. Organization/project least-privilege matrix

Formula / scoring rule: Gate = documented grant/deny × canary result × resource scope; undocumented permission is not treated as denied. Source: pricing registry and dated evidence index verified 2026-08-27.

2. Credential rotation-and-revocation propagation ledger

Formula / scoring rule: Duplicate-work ceiling = accepted replay after revocation + in-flight work; close only when last-accepted, rejection, and recovery timestamps join. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Credit-grant, prepaid-balance, and invoice-liability reconciler

Formula / scoring rule: Variance = grant + paid usage + adjustments − expiry − invoice lines; close only when promotional and purchased value remain separate. Source: pricing registry and dated evidence index verified 2026-08-27.

1. Scale Tier committed-token-unit utilization ledger

Formula / scoring rule: Effective accepted-token cost = PAYG-equivalent accepted input/output charge ÷ accepted tokens; compare contracted units, unused capacity, overflow/rejection, eligibility, and service window. Undisclosed contract or overage rates stay Unavailable. Source: pricing registry and dated evidence index verified 2026-08-27.

2. API data-sharing opt-in incentive and liability reconciler

Formula / scoring rule: Net invoice = paid usage + excluded usage − eligible complimentary tokens − credited adjustments; consent scope, effective date, cap, and revocation must be joined. Source: pricing registry and dated evidence index verified 2026-08-27.

3. Project archive/delete final-charge canary

Formula / scoring rule: Close only when blockers, resource disposition, last accepted request, retained usage/export, post-close units, credit/refund state, and final invoice variance all join. Source: pricing registry and dated evidence index verified 2026-08-27.

1. Batch duplicate-ID, cancellation, expiry, and partial-output settlement ledger

Formula / scoring rule: Close = validated IDs + completed/failed/cancelled/expired counts + returned usage + retained output/error files + scoped retry + credits + final invoice variance. Duplicate IDs never become accepted work. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: OpenAI GPT-4o / Batch API registry rate verified 2026-08-27; test suite: Batch 30 OpenAI Batch settlement fixture/test suite (run and result recorded 2026-08-27).

2. Approximate-user-location and locale marginal-yield ledger

Formula / scoring rule: Marginal yield = newly supported geographic claims ÷ additional search calls; parameter acceptance, geographic fit, cited sources, repair searches, latency, and accepted-answer cost must all join. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: OpenAI GPT-4o / web-search usage registry rate verified 2026-08-27; test suite: Batch 30 OpenAI location-yield fixture/test suite (run and result recorded 2026-08-27).

3. Computer-use screenshot/action/confirmation loop ledger

Formula / scoring rule: Accepted-task cost = screenshot input + reasoning/output + tool calls + retries; confirmation-gated actions, duplicate or unsafe actions, latency, and reviewer acceptance are required. Undocumented tool fees remain Unavailable. Source: pricing registry and dated evidence index verified 2026-08-27; provider registry: OpenAI GPT-4o / computer-use usage registry rate verified 2026-08-27; test suite: Batch 30 OpenAI computer-use loop fixture/test suite (run and result recorded 2026-08-27).

1. Realtime server-VAD barge-in, truncation, and abandoned-audio billing ledger

Formula / scoring rule: Accepted-turn cost = committed audio + model input/output usage + replayed work; discarded audio is never assumed free and a missing debit is Unavailable. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: OpenAI Realtime pricing and event registry, verified 2026-08-27.

2. Responses previous_response_id fork-versus-linear continuation accounting ledger

Formula / scoring rule: Continuation cost = inherited instructions/tool state + stored or replayed input + cached/uncached usage + output; stored state is not treated as free. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: OpenAI Responses usage and stored-response registry, verified 2026-08-27.

3. Vector-store file replacement, duplicate-content, and delayed-deletion settlement ledger

Formula / scoring rule: Close = ingestion/version + chunk/search state + deletion propagation + storage day boundary + query units + invoice; undocumented deduplication credits remain Unavailable. First-party registry: allaiask.com pricing and evidence registry, verified 2026-08-27. Provider/model source: OpenAI vector-store file/search registry, verified 2026-08-27.

1. Responses background-mode poll/cancel/expiry settlement ledger

Formula / scoring rule: Final bill = returned input + cache + reasoning + output + tool usage; polling/cancellation credits are Unavailable unless separately returned. OpenAI Responses background-mode pricing/evidence registry. Dated registry and evidence index, verified 2026-08-27.

2. Code-interpreter container creation, reuse, idle-expiry, file-artifact, and concurrent-session cost ledger

Formula / scoring rule: Accepted artifact cost = container/session units + model/tool usage + uploaded/generated file units + scoped restart/replay. OpenAI Code Interpreter containers and pricing evidence registry. Dated registry and evidence index, verified 2026-08-27.

3. Signed-webhook duplicate, delay, reordering, and redelivery attribution canary

Formula / scoring rule: Reconciled charge = terminal API usage joined to event/object ID; webhook transport is not billed without a dated charge record. OpenAI webhook signatures, usage exports, and pricing evidence registry. Dated registry and evidence index, verified 2026-08-27.

1. Stored-completion-to-distillation dataset lineage and job-cost ledger

Formula / scoring rule: Total bill = retained input + output + tool usage + training-job units; undocumented distillation or storage units stay Unavailable. First-party pricing/evidence registry: OpenAI distillation, stored responses, fine-tuning, and usage evidence, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: OpenAI distillation guideOpenAI API pricing.

2. Transcription diarization, timestamps, and overlapping-chunk settlement matrix

Formula / scoring rule: Accepted transcription cost = returned audio/input/output units + retry stitching; overlap is not silently deducted. First-party pricing/evidence registry: OpenAI transcription endpoint and audio pricing evidence, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: OpenAI speech-to-text guideOpenAI API pricing.

3. Realtime WebRTC, WebSocket, and SIP transport-parity ledger

Formula / scoring rule: Transport comparison joins session IDs, committed/abandoned audio, usage, accepted turns, and final charge; recovery is not a transport equivalence claim. First-party pricing/evidence registry: OpenAI Realtime transport, audio, and usage evidence, verified 2026-08-27; unsupported units or credits remain Unavailable.. Dated registry and evidence index, verified 2026-08-27. First-party sources: OpenAI Realtime API guideOpenAI API pricing.

1. File Search ranker, score-threshold, and metadata-filter marginal-yield ledger

Formula / scoring rule: Marginal yield = accepted relevant chunks ÷ returned chunks after ranker, threshold, metadata, stale, and ACL gates; bill uses returned usage. OpenAI File Search retrieval-control 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: OpenAI File Search guideOpenAI API pricing.

2. Realtime SIP DTMF, call-transfer, and hang-up atomicity canary

Formula / scoring rule: Accepted call = control acknowledgement + unique event IDs + committed audio/tool continuity + accepted outcome; duplicate actions fail the gate. OpenAI Realtime SIP control 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: OpenAI Realtime API guideOpenAI API pricing.

3. Streamed image-generation partial-preview settlement ledger

Formula / scoring rule: Charge = returned generation input/output usage for completed work; preview, retry, disconnect, and abandoned units are never silently free. OpenAI streamed image preview and image pricing 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: OpenAI image generation guideOpenAI API pricing.

1. Responses max_tool_calls aggregate-cap ledger

Formula / scoring rule: Aggregate-cap acceptance = accepted max_tool_calls + attempted/completed calls by type + stop state + side-effect checksum + returned usage + invoice. OpenAI Responses aggregate tool-cap 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: OpenAI Responses API documentationOpenAI API pricing.

2. Web-search domain and approximate-location interaction canary

Formula / scoring rule: Search acceptance = submitted/effective filters + requested/resolved location + final URLs + unique supported claims + reviewer result + marginal search bill. OpenAI web-search domain/location interaction 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: OpenAI web search documentationOpenAI API pricing.

3. Realtime out-of-band response isolation and settlement ledger

Formula / scoring rule: Out-of-band acceptance = call/session/response/conversation IDs + committed input + state isolation + side-effect uniqueness + modality usage + invoice. OpenAI Realtime out-of-band response 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: OpenAI Realtime API documentationOpenAI API pricing.

1. Responses `include` selector payload-and-debit ledger

Formula / scoring rule: Include settlement = submitted/effective selectors + exposed source/result/state fields + response bytes + cache/input/output/tool usage + semantic hash + exact invoice; metadata bytes are excluded from tokens. OpenAI Responses selective-expansion 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: OpenAI Responses API documentationOpenAI API pricing.

2. `prompt_cache_key` and cache-retention routing canary

Formula / scoring rule: Routing evidence = worker/key/rotation/retention controls + project/model identity + cached/uncached usage + observable concentration + expiry probe + latency + answer equivalence + bill. OpenAI prompt-cache-key and retention 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: OpenAI prompt caching documentationOpenAI API pricing.

3. Realtime `session.update` mutation-atomicity and marginal-settlement ledger

Formula / scoring rule: Mutation acceptance = event/session/response IDs + accepted/effective boundary + buffered state + duplicate effects + usage + recovery + reviewer result + charge. OpenAI Realtime session mutation 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: OpenAI Realtime API documentationOpenAI API pricing.

1. Responses `safety_identifier` acceptance, normalization, and attribution ledger

Formula / scoring rule: Attribution = submitted/effective identifier + project/key/request/response IDs + safety state + returned cache/input/output/tool usage + export visibility + exact invoice. OpenAI safety-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: OpenAI Responses API documentationOpenAI API pricing.

2. Requested reasoning-summary level and returned-summary settlement canary

Formula / scoring rule: Summary settlement = model/control eligibility + reasoning-item/summary presence + summary bytes + hidden/disclosed units + cache/input/output usage + checker + continuation + bill. OpenAI reasoning-summary 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: OpenAI reasoning documentationOpenAI API pricing.

3. Evals model-grader and custom-grader nested-cost reconciler

Formula / scoring rule: Nested bill = candidate + grader + cache + tool usage across eval/run/sample/grader IDs and retries, divided by human-adjudicated accepted score. OpenAI Evals nested-grader 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: OpenAI Evals documentationOpenAI API pricing.

1. Realtime semantic-VAD eagerness and turn-boundary settlement ledger

Formula / scoring rule: Turn settlement = submitted/effective eagerness + audio-buffer/response IDs + commit/timeout boundary + transcript/tool continuity + audio/text/model usage + repair + acceptance + latency + exact bill. OpenAI Realtime semantic-VAD 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: OpenAI Realtime API documentationOpenAI API pricing.

2. Realtime input-audio noise-reduction and transcription-conditioning canary

Formula / scoring rule: Audio conditioning = accepted configuration + submitted/decoded duration + word/error alignment + committed input + abandoned/duplicate responses + audio/text usage + recovery + reviewer acceptance + charge. OpenAI Realtime audio-conditioning 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: OpenAI Realtime audio documentationOpenAI API pricing.

3. Computer-use viewport, device-pixel-ratio, zoom, scroll, and coordinate-frame atomicity ledger

Formula / scoring rule: Coordinate atomicity = environment/screenshot/action IDs + submitted/resolved coordinates + target hash + side-effect checksum + model/tool/image usage + confirmation/retry scope + acceptance + latency + invoice. OpenAI computer-use coordinate 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: OpenAI computer use documentationOpenAI API pricing.