← Frontier Inference Margins · all research reports

Research artifact — each page states its own provenance class in its header: verbatim originals carry SHA-256 stamps; adopted-findings summaries and reconstructions say so explicitly. Conclusions are synthesized (and where needed corrected) in the main report; each page carries its own run date.

Preset grounding pack — dated re-audit (GPT-5.6 Pro re-emission, 2026-07-11)

Provenance class: author-model re-emission, not an independently verified original. The original pack's full row set lived in conversation-sandbox files that expired. On 2026-07-11 the same GPT-5.6 Pro project was asked to re-emit the pack inline; it reports having located its own retained "22:29 UTC final export" and re-emitting it exactly (run 21m0s, Pro-verified, completion_path: api, conversation). That exactness claim is the model's own and cannot be independently byte-verified — which is why the site's authoritative parameter record is the adopted grounding ledger, generated from the deployed preset registry. Where this page and the ledger differ, the ledger wins. SHA-256 (response body below the delta section): c4a8bc8219d952895f6c73004e6547c3872f46e6930cbdd37e994c2ec5e6d197

Delta vs the adopted values (checked 2026-07-11)

Tariff and architecture rows were checked against the deployed presets: all mappable tariff rows (prices, cached-input ratios, context/output limits) and all disclosed architecture rows match the adopted values. The differences are where the site's council-gated ship list deliberately departed from the pack's recommendations, plus two modeling-mechanism differences:

area this re-audit says site adopted why the site differs
DeepSeek V4-Flash fleet blend Ascend .50 / H800 .35 / H20 .15 H800 50 / H20 20 / Ascend 30 Council ship list kept the V4-family blend consistent with the V4 Pro dive scenario; both are labeled assumptions
GLM-4.7 fleet blend Ascend .70 / H800 .20 / H20 .10 Ascend 40 / H800 40 / H20 20 Adopted blend mirrors GLM-5.2's (same serving estate); both speculative
GPT-5.6 Terra/Luna blends include a B200 class H100/H200/GB200/GB300 The calculator's hardware table does not model B200 separately; GB200 row carries the Blackwell class
Gemini 3.5 Flash fleet TPUv7 .60 / v6e .30 / v5e .10 TPU v7 100% Older TPU generations are not modeled; the single-row simplification is stated in the preset note
xAI valuation mechanism GB300-only replica, chip-hour override ($0.60 / $5.27), MFU override 0.02, util = 1 rentMult 0.156 / 1.37 on the Colossus blend (weighted rent $3.85/hr → $0.60 / $5.27), util 48 Same $/GPU-hour by construction; the site expresses it through the blend-relative multiplier so the lens composes with the model's fleet
Ant H20 SLO mechanism effective-MFU override 0.16875 (Pro tier) H20 effDec default 0.17 + stack 1.0 Same operating point within rounding (the site's neutral default was fit to this anchor)
China on-demand multiplier 6.152857142857143 6.15 Display rounding; the ledger records 6.15 as adopted

Rows describing candidate presets the council rejected (Anthropic-strategic merge, reported-margin-implied, Kimi K2.6, Grok 4.3) are absent from both the re-audit's final export and the site — consistent.


I’m reconstructing the annex from project artifacts and live presets, verifying volatile values against authoritative sources, and explicitly marking any irrecoverable rows rather than inferring them.

The archived consultation confirms 192 rows and preserves adopted tariff scenarios and China procurement multipliers; I’m extracting deployed static definitions to avoid fresh approximations.

I located the exact July 10 final JSON and Markdown exports, plus earlier versions for delta checking, enabling exact recovery of the row set.

Inference Margins preset grounding pack — durable inline export

According to a document from 2026-07-10, this is the recovered 22:29 UTC final export. It contains 192 preset parameter rows: 78 PERSPECTIVE rows and 114 MODEL rows. OMITTED-UNRECONSTRUCTED: none. fileciteturn33file8

The recovered evidence labels map to the requested source classes as follows: DISCLOSEDfirst-party; CREDIBLY REPORTEDcredible-secondary; COMMUNITY ESTIMATEcommunity; SPECULATIONassumption.

Schema and implementation semantics — not counted in the 192 preset rows

parameter value unit source (URL) source class as-of date note
meta.presetRowCount 192 rows https://inference-margins.pages.dev/ community 2026-07-10 78 PERSPECTIVE rows plus 114 MODEL rows.
meta.outputCostUSDPerToken (chipHourUSD / 3600 / outputTokPerSecPerChip) / utilization formula https://inference-margins.pages.dev/ community 2026-07-10 Calculator output-token cost equation.
meta.outputTokPerSecPerChip dense8bitFLOPS × precisionMult × effectiveMFU × interactivityMult / (2 × activeParams) formula https://inference-margins.pages.dev/ community 2026-07-10 Calculator throughput equation.
meta.tokenPriceUnit USD per 1,000,000 tokens convention https://inference-margins.pages.dev/ community 2026-07-10 Applies to input, cached input, cache write and output tariff fields.
meta.chipHourUnit USD per accelerator-hour convention https://inference-margins.pages.dev/ community 2026-07-10 “Card-hour” and “GPU-hour” rows are normalized to the named accelerator.
meta.throughputUnit aggregate output tokens/s/accelerator convention https://inference-margins.pages.dev/ community 2026-07-10 Unless a row explicitly states another denominator.
meta.parameterCountUnit billions of parameters convention https://inference-margins.pages.dev/ community 2026-07-10 Applies to total and active parameter fields.
meta.nullSemantics unknown / leave unset, never zero rule https://inference-margins.pages.dev/ assumption 2026-07-10 A published zero tariff remains zero; an undisclosed field is null.
meta.effectiveMFUOverrideRule mutually exclusive with fallback interactivity multiplier rule https://inference-margins.pages.dev/ assumption 2026-07-10 Never apply both representations to one calculation.
meta.measuredThroughputUtilizationRule utilization = 1 for measured saturated per-occupied-chip throughput rule https://inference-margins.pages.dev/ assumption 2026-07-10 Fleet occupancy is a separate allocation question.
meta.batchShareSemantics customer-token share receiving a published asynchronous batch tariff rule https://inference-margins.pages.dev/ assumption 2026-07-10 Internal continuous batching is not customer-visible Batch API usage.
meta.chinaPublicCloudFallback rentMult.globalFallback = 6.152857142857143 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Prefer supported per-chip multipliers; use this only when one scalar is required.
meta.tariffScenarioSemantics price and endpoint metadata grounded; topology, parameters, precision, workload and fleet remain scenarios rule https://inference-margins.pages.dev/ assumption 2026-07-10 Applies to visibly marked TARIFF SCENARIO presets.

Provider tariffs and endpoint limits — 55 rows

parameter value unit source (URL) source class as-of date note
DeepSeek V4-Flash.contextTokens 1000000 tokens https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 Official standard context.
DeepSeek V4-Flash.maxOutputTokens 384000 tokens https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 Official maximum output.
DeepSeek V4-Flash.price.inputMiss 0.14 USD/million input tokens https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 Official rate card checked 2026-07-10.
DeepSeek V4-Flash.price.inputHit 0.0028 USD/million cached-input tokens https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 Official rate card checked 2026-07-10.
DeepSeek V4-Flash.price.output 0.28 USD/million output tokens https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 Official rate card checked 2026-07-10.
DeepSeek V4-Flash.price.cacheWrite null (leave unset) USD/million cache-write tokens https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 No separate cache-write tariff shown.
DeepSeek V4-Flash.price.cacheStoragePerMtokHour null (leave unset) USD/million cached tokens-hour https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 No cache-storage tariff shown.
DeepSeek V4-Flash.batchDiscount null (leave unset) fraction https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 No V4-Flash batch tariff shown.
DeepSeek V4-Flash.concurrency 2500 concurrent requests https://api-docs.deepseek.com/quick_start/pricing first-party 2026-07-10 Official V4-Flash concurrency tier; not sequences per accelerator.
GLM-4.7.contextTokens 200000 tokens https://docs.z.ai/guides/llm/glm-4.7 first-party 2026-07-10 Official model guide.
GLM-4.7.maxOutputTokens 128000 tokens https://docs.z.ai/guides/llm/glm-4.7 first-party 2026-07-10 Official model guide.
GLM-4.7.price.inputMiss 0.6 USD/million input tokens https://docs.z.ai/guides/overview/pricing first-party 2026-07-10 Official rate card checked 2026-07-10.
GLM-4.7.price.inputHit 0.11 USD/million cached-input tokens https://docs.z.ai/guides/overview/pricing first-party 2026-07-10 Official rate card checked 2026-07-10.
GLM-4.7.price.output 2.2 USD/million output tokens https://docs.z.ai/guides/overview/pricing first-party 2026-07-10 Official rate card checked 2026-07-10.
GLM-4.7.price.cacheWrite null (leave unset) USD/million cache-write tokens https://docs.z.ai/guides/overview/pricing first-party 2026-07-10 No separate cache-write price shown.
GLM-4.7.price.cacheStoragePerMtokHour 0 USD/million cached tokens-hour https://docs.z.ai/guides/overview/pricing first-party 2026-07-10 Rate card states “Limited-time Free”; zero is a published temporary tariff, not unknown.
GLM-4.7.batchDiscount null (leave unset) fraction https://docs.z.ai/guides/overview/pricing first-party 2026-07-10 No GLM-4.7 batch tariff shown.
GLM-4.7.concurrency null (leave unset) concurrent requests https://docs.z.ai/guides/llm/glm-4.7
https://docs.z.ai/guides/overview/pricing
first-party 2026-07-10 No model-specific public concurrency tier identified.
GPT-5.6 Terra.contextTokens 1050000 tokens https://developers.openai.com/api/docs/models/gpt-5.6-terra first-party 2026-07-10 Official endpoint page.
GPT-5.6 Terra.maxOutputTokens 128000 tokens https://developers.openai.com/api/docs/models/gpt-5.6-terra first-party 2026-07-10 Official endpoint page.
GPT-5.6 Terra.longContextThresholdInputTokens 272000 input tokens https://developers.openai.com/api/docs/models/gpt-5.6-terra first-party 2026-07-10 Above 272K input, the entire request reprices at 2× input and 1.5× output.
GPT-5.6 Terra.price.inputMiss.short 2.5 USD/million input tokens https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Official standard short-context rate.
GPT-5.6 Terra.price.inputHit.short 0.25 USD/million cached-input tokens https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Official cached-input short-context rate.
GPT-5.6 Terra.price.cacheWrite.short 3.125 USD/million cache-write tokens https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Official cache-write short-context rate.
GPT-5.6 Terra.price.output.short 15 USD/million output tokens https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Official standard short-context output rate.
GPT-5.6 Terra.price.inputMiss.long 5 USD/million input tokens https://developers.openai.com/api/docs/models/gpt-5.6-terra
https://developers.openai.com/api/docs/pricing
first-party 2026-07-10 Official >272K whole-request tariff.
GPT-5.6 Terra.price.inputHit.long 0.5 USD/million cached-input tokens https://developers.openai.com/api/docs/models/gpt-5.6-terra
https://developers.openai.com/api/docs/pricing
first-party 2026-07-10 Official >272K whole-request tariff.
GPT-5.6 Terra.price.cacheWrite.long 6.25 USD/million cache-write tokens https://developers.openai.com/api/docs/models/gpt-5.6-terra
https://developers.openai.com/api/docs/pricing
first-party 2026-07-10 Official >272K whole-request tariff.
GPT-5.6 Terra.price.output.long 22.5 USD/million output tokens https://developers.openai.com/api/docs/models/gpt-5.6-terra
https://developers.openai.com/api/docs/pricing
first-party 2026-07-10 Official >272K whole-request tariff.
GPT-5.6 Terra.price.cacheStoragePerMtokHour null (leave unset) USD/million cached tokens-hour https://developers.openai.com/api/docs/pricing first-party 2026-07-10 No persistent cache-storage tariff shown.
GPT-5.6 Terra.batchDiscount 0.5 fraction https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Batch and Flex are exactly half Standard rates.
GPT-5.6 Terra.concurrency null (leave unset) concurrent requests https://developers.openai.com/api/docs/models/gpt-5.6-terra first-party 2026-07-10 No model-specific public concurrency tier identified.
GPT-5.6 Luna.contextTokens 1050000 tokens https://developers.openai.com/api/docs/models/gpt-5.6-luna first-party 2026-07-10 Official endpoint page.
GPT-5.6 Luna.maxOutputTokens 128000 tokens https://developers.openai.com/api/docs/models/gpt-5.6-luna first-party 2026-07-10 Official endpoint page.
GPT-5.6 Luna.longContextThresholdInputTokens 272000 input tokens https://developers.openai.com/api/docs/models/gpt-5.6-luna first-party 2026-07-10 Above 272K input, the entire request reprices at 2× input and 1.5× output.
GPT-5.6 Luna.price.inputMiss.short 1 USD/million input tokens https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Official standard short-context rate.
GPT-5.6 Luna.price.inputHit.short 0.1 USD/million cached-input tokens https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Official cached-input short-context rate.
GPT-5.6 Luna.price.cacheWrite.short 1.25 USD/million cache-write tokens https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Official cache-write short-context rate.
GPT-5.6 Luna.price.output.short 6 USD/million output tokens https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Official standard short-context output rate.
GPT-5.6 Luna.price.inputMiss.long 2 USD/million input tokens https://developers.openai.com/api/docs/models/gpt-5.6-luna
https://developers.openai.com/api/docs/pricing
first-party 2026-07-10 Official >272K whole-request tariff.
GPT-5.6 Luna.price.inputHit.long 0.2 USD/million cached-input tokens https://developers.openai.com/api/docs/models/gpt-5.6-luna
https://developers.openai.com/api/docs/pricing
first-party 2026-07-10 Official >272K whole-request tariff.
GPT-5.6 Luna.price.cacheWrite.long 2.5 USD/million cache-write tokens https://developers.openai.com/api/docs/models/gpt-5.6-luna
https://developers.openai.com/api/docs/pricing
first-party 2026-07-10 Official >272K whole-request tariff.
GPT-5.6 Luna.price.output.long 9 USD/million output tokens https://developers.openai.com/api/docs/models/gpt-5.6-luna
https://developers.openai.com/api/docs/pricing
first-party 2026-07-10 Official >272K whole-request tariff.
GPT-5.6 Luna.price.cacheStoragePerMtokHour null (leave unset) USD/million cached tokens-hour https://developers.openai.com/api/docs/pricing first-party 2026-07-10 No persistent cache-storage tariff shown.
GPT-5.6 Luna.batchDiscount 0.5 fraction https://developers.openai.com/api/docs/pricing first-party 2026-07-10 Batch and Flex are exactly half Standard rates.
GPT-5.6 Luna.concurrency null (leave unset) concurrent requests https://developers.openai.com/api/docs/models/gpt-5.6-luna first-party 2026-07-10 No model-specific public concurrency tier identified.
Gemini 3.5 Flash.contextTokens 1048576 tokens https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash first-party 2026-07-10 Official model/API limit.
Gemini 3.5 Flash.maxOutputTokens 65536 tokens https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash
https://deepmind.google/models/model-cards/gemini-3-5-flash/
first-party 2026-07-10 Official model/API limit.
Gemini 3.5 Flash.price.inputMiss 1.5 USD/million input tokens https://ai.google.dev/gemini-api/docs/pricing first-party 2026-07-10 Official rate card checked 2026-07-10.
Gemini 3.5 Flash.price.inputHit 0.15 USD/million cached-input tokens https://ai.google.dev/gemini-api/docs/pricing first-party 2026-07-10 Official context-caching tariff.
Gemini 3.5 Flash.price.output 9 USD/million output tokens https://ai.google.dev/gemini-api/docs/pricing first-party 2026-07-10 Official output tariff; output includes thinking tokens.
Gemini 3.5 Flash.price.cacheWrite null (leave unset) USD/million cache-write tokens https://ai.google.dev/gemini-api/docs/pricing first-party 2026-07-10 No separate cache-write tariff; creation is billed through input plus storage.
Gemini 3.5 Flash.price.cacheStoragePerMtokHour 1 USD/million cached tokens-hour https://ai.google.dev/gemini-api/docs/pricing first-party 2026-07-10 Official explicit-caching storage rate.
Gemini 3.5 Flash.batchDiscount 0.5 fraction https://ai.google.dev/gemini-api/docs/pricing first-party 2026-07-10 Batch input, cached input and output are half Standard rates.
Gemini 3.5 Flash.concurrency null (leave unset) concurrent requests https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash
https://ai.google.dev/gemini-api/docs/pricing
first-party 2026-07-10 No model-specific public concurrency tier identified.

Model architectures and parameterization — 32 rows

parameter value unit source (URL) source class as-of date note
DeepSeek V4-Flash.architecture MoE + token-wise compression + DSA text https://api-docs.deepseek.com/news/news260424 first-party 2026-07-10 Official V4 release and linked technical report.
DeepSeek V4-Flash.totalParamsB 284 billion parameters https://api-docs.deepseek.com/news/news260424 first-party 2026-07-10 Official V4-Flash disclosure.
DeepSeek V4-Flash.activeParamsB 13 billion active parameters https://api-docs.deepseek.com/news/news260424 first-party 2026-07-10 Official V4-Flash disclosure.
DeepSeek V4-Flash.precisionMult 1.85 multiplier https://api-docs.deepseek.com/news/news260424 assumption 2026-07-10 Selective low-precision calculator scenario; not a disclosed end-to-end production speedup.
GLM-4.7.architecture MoE + MTP; interleaved/preserved thinking text https://github.com/zai-org/GLM-4.5/blob/main/README.md first-party 2026-07-10 Official repository and serving recipes.
GLM-4.7.totalParamsB 355 billion parameters https://github.com/zai-org/GLM-4.5/blob/main/README.md first-party 2026-07-10 Official GLM-4.7 model download table.
GLM-4.7.activeParamsB 32 billion active parameters https://github.com/zai-org/GLM-4.5/blob/main/README.md first-party 2026-07-10 Official GLM-4.7 model download table.
GLM-4.7.precisionMult 1 multiplier https://github.com/zai-org/GLM-4.5/blob/main/README.md assumption 2026-07-10 BF16 and FP8 weights are published, but hosted numerical format is unknown; retain dense-8-bit baseline.
GPT-5.6 Terra.architecture sparse/MoE-family scenario — UNKNOWN text https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 OpenAI does not disclose architecture; scenario only.
GPT-5.6 Terra.totalParamsB.central 1000 billion parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Project OpenAI deep-dive scenario.
GPT-5.6 Terra.totalParamsB.min 250 billion parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Subjective scenario range.
GPT-5.6 Terra.totalParamsB.max 5000 billion parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Subjective scenario range.
GPT-5.6 Terra.activeParamsB.central 50 billion active parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Project OpenAI deep-dive scenario.
GPT-5.6 Terra.activeParamsB.min 20 billion active parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Subjective scenario range.
GPT-5.6 Terra.activeParamsB.max 110 billion active parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Subjective scenario range.
GPT-5.6 Terra.precisionMult 1.35 multiplier https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Mixed FP8/MXFP4-equivalent calculator scenario; production numerical format is undisclosed.
GPT-5.6 Luna.architecture sparse/MoE-family scenario — UNKNOWN text https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 OpenAI does not disclose architecture; scenario only.
GPT-5.6 Luna.totalParamsB.central 250 billion parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Project OpenAI deep-dive scenario.
GPT-5.6 Luna.totalParamsB.min 50 billion parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Subjective scenario range.
GPT-5.6 Luna.totalParamsB.max 1500 billion parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Subjective scenario range.
GPT-5.6 Luna.activeParamsB.central 20 billion active parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Project OpenAI deep-dive scenario.
GPT-5.6 Luna.activeParamsB.min 8 billion active parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Subjective scenario range.
GPT-5.6 Luna.activeParamsB.max 50 billion active parameters https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Subjective scenario range.
GPT-5.6 Luna.precisionMult 1.35 multiplier https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Mixed FP8/MXFP4-equivalent calculator scenario; production numerical format is undisclosed.
Gemini 3.5 Flash.architecture natively multimodal reasoning; topology undisclosed text https://deepmind.google/models/model-cards/gemini-3-5-flash/ first-party 2026-07-10 Official model card; topology is not disclosed.
Gemini 3.5 Flash.totalParamsB.central 600 billion parameters https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Google-dive scenario, not a leak.
Gemini 3.5 Flash.totalParamsB.min 400 billion parameters https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Scenario bracket.
Gemini 3.5 Flash.totalParamsB.max 1200 billion parameters https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Scenario bracket.
Gemini 3.5 Flash.activeParamsB.central 20 billion active parameters https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Google-dive scenario.
Gemini 3.5 Flash.activeParamsB.min 10 billion active parameters https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Scenario bracket.
Gemini 3.5 Flash.activeParamsB.max 50 billion active parameters https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Scenario bracket.
Gemini 3.5 Flash.precisionMult 1 multiplier https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Dense-8-bit calculator baseline; hosted production precision is undisclosed.

Hardware rates by procurement class — 23 rows

parameter value unit source (URL) source class as-of date note
xAI cash-marginal.costBasisMode cash-marginal mode https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Strict short-run economic lens: sunk hardware; only incremental power, cooling, maintenance and operations are marginal.
xAI cash-marginal.chipHourUSD 0.6 USD/accelerator-hour https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Central strict short-run cash value.
xAI cash-marginal.rentMult 1 multiplier https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Direct chip-hour override; do not apply a second procurement scalar.
xAI cash-marginal.derived.outputCostUSDPerMtok 0.6666666666666667 USD/million output tokens https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 0.60 ÷ 3600 ÷ 250 × 1,000,000.
xAI opportunity-cost.costBasisMode external-opportunity-value mode https://content.spacex.com/cms-assets/FINAL_Documents%20and%20Updates/SpaceX%20-%20EU%20Prospectus%20%28Approved%20by%20Bafin%29%20-%20June%205%2C%202026.pdf
https://inference-margins.pages.dev/research/xai-gptpro
credible-secondary 2026-07-10 Allocation lens defined from the Anthropic capacity agreement; this is foregone external revenue, not xAI production cost.
xAI opportunity-cost.chipHourUSD 5.27 USD/accelerator-hour https://content.spacex.com/cms-assets/FINAL_Documents%20and%20Updates/SpaceX%20-%20EU%20Prospectus%20%28Approved%20by%20Bafin%29%20-%20June%205%2C%202026.pdf credible-secondary 2026-07-10 $1.25B/month ÷ 325,000 named GPUs ÷ 730 h/month = $5.267…; the contract bundle also includes CPUs, storage and networking.
xAI opportunity-cost.rentMult 1 multiplier https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Direct chip-hour override; do not apply a second procurement scalar.
xAI opportunity-cost.derived.outputCostUSDPerMtok 5.855555555555555 USD/million output tokens https://content.spacex.com/cms-assets/FINAL_Documents%20and%20Updates/SpaceX%20-%20EU%20Prospectus%20%28Approved%20by%20Bafin%29%20-%20June%205%2C%202026.pdf
https://inference-margins.pages.dev/research/xai-gptpro
assumption 2026-07-10 5.27 ÷ 3600 ÷ 250 × 1,000,000.
China public-cloud on-demand.baseAnnualChipHourUSD.H800 1.75 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro community 2026-07-10 Calculator annual-commit central value used as the H800 denominator.
China public-cloud on-demand.baseAnnualChipHourUSD.H20 1 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro community 2026-07-10 Calculator annual-commit central value used as the H20 denominator.
China public-cloud on-demand.baseAnnualChipHourUSD.Ascend910C 1.95 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro credible-secondary 2026-07-10 Huatai one-year private-cloud procurement midpoint used as the annual 910C baseline.
China public-cloud on-demand.rateCard.Tencent.H800.HCC.monthly 10.615 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro first-party 2026-07-10 Midpoint of $10.55–$10.68 per card-hour for the eight-card monthly HCC product.
China public-cloud on-demand.rateCard.Tencent.H20.PNV6.onDemand 4.48 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro first-party 2026-07-10 Tencent PNV6 posted on-demand normalization.
China public-cloud on-demand.rateCard.Tencent.H20.HCC.onDemandMid 5.08 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro first-party 2026-07-10 Midpoint of the $3.91–$6.25 per-card-hour HCC-PNV6 bundle spread.
China public-cloud on-demand.rateCard.Alibaba.H20.96G.onDemand 7.4 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro first-party 2026-07-10 Alibaba RDS Custom AI H20-96G normalization.
China public-cloud on-demand.rateCard.Alibaba.H20.141G.onDemand 10.62 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro first-party 2026-07-10 Alibaba RDS Custom AI H20-141G normalization.
China public-cloud on-demand.chipHourUSD.H20.representative 6.24 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro community 2026-07-10 Median of [4.48, 5.08, 7.40, 10.62] = (5.08 + 7.40) ÷ 2.
China public-cloud on-demand.rentMult.H20 6.24 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro community 2026-07-10 6.24 ÷ annual H20 default 1.00.
China public-cloud on-demand.rentMult.H800 6.065714285714286 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro community 2026-07-10 10.615 ÷ annual H800 default 1.75.
China public-cloud on-demand.rentMult.globalFallback 6.152857142857143 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Mean of H20 multiplier 6.24 and H800 multiplier 6.065714285714286.
China public-cloud on-demand.rentMult 6.152857142857143 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Compatibility alias for rentMult.globalFallback when the calculator supports only one procurement multiplier.
China public-cloud on-demand.rentMult.Ascend910CProxy 6.152857142857143 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Apply the global public-cloud fallback because no reproducible public on-demand 910C tariff was found.
China public-cloud on-demand.chipHourUSD.Ascend910CProxy 11.998071428571428 USD/card-hour https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 1.95 × 6.152857142857143; explicit scalar proxy, not a reproduced Huawei public retail price.

Hardware fleet routing and blend assumptions — 20 rows

parameter value unit source (URL) source class as-of date note
xAI cash-marginal.hardwareFleetBlend.GB300 1 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Central 8×GB300 Grok 4.5 replica scenario.
xAI opportunity-cost.hardwareFleetBlend.GB300 1 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Central 8×GB300 Grok 4.5 replica scenario.
Ant Group production H20 (SLO replay).hardwareFleetBlend.H20 1 fraction https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Operator benchmark hardware.
DeepSeek V4-Flash.hardwareFleetBlend.Ascend910C_proxy_for_950 0.5 fraction https://inference-margins.pages.dev/research/deepseek-gptpro
https://inference-margins.pages.dev/research/chinese-accel-gptpro
assumption 2026-07-10 DeepSeek 2026 serving prior; 910C is an explicit calculator proxy for newer Ascend hardware.
DeepSeek V4-Flash.hardwareFleetBlend.H800 0.35 fraction https://inference-margins.pages.dev/research/deepseek-gptpro assumption 2026-07-10 DeepSeek 2026 serving prior; no audited shares.
DeepSeek V4-Flash.hardwareFleetBlend.H20 0.15 fraction https://inference-margins.pages.dev/research/deepseek-gptpro assumption 2026-07-10 DeepSeek 2026 serving prior; no audited shares.
GLM-4.7.hardwareFleetBlend.Ascend910C_proxy_for_domestic_other 0.7 fraction https://inference-margins.pages.dev/research/zhipu-gptpro
https://inference-margins.pages.dev/research/chinese-accel-gptpro
assumption 2026-07-10 Domestic-accelerator scenario; proxy for a broader domestic pool.
GLM-4.7.hardwareFleetBlend.H800 0.2 fraction https://inference-margins.pages.dev/research/zhipu-gptpro assumption 2026-07-10 Mixed-fleet scenario.
GLM-4.7.hardwareFleetBlend.H20 0.1 fraction https://inference-margins.pages.dev/research/zhipu-gptpro assumption 2026-07-10 Mixed-fleet scenario.
GPT-5.6 Terra.hardwareFleetBlend.H100 0.25 fraction https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Partner-hosted Hopper/Blackwell routing scenario.
GPT-5.6 Terra.hardwareFleetBlend.H200 0.2 fraction https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Partner-hosted Hopper/Blackwell routing scenario.
GPT-5.6 Terra.hardwareFleetBlend.B200 0.2 fraction https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Partner-hosted Hopper/Blackwell routing scenario.
GPT-5.6 Terra.hardwareFleetBlend.GB200 0.35 fraction https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Partner-hosted Hopper/Blackwell routing scenario.
GPT-5.6 Luna.hardwareFleetBlend.H100 0.35 fraction https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Partner-hosted Hopper/Blackwell routing scenario.
GPT-5.6 Luna.hardwareFleetBlend.H200 0.3 fraction https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Partner-hosted Hopper/Blackwell routing scenario.
GPT-5.6 Luna.hardwareFleetBlend.B200 0.15 fraction https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Partner-hosted Hopper/Blackwell routing scenario.
GPT-5.6 Luna.hardwareFleetBlend.GB200 0.2 fraction https://inference-margins.pages.dev/research/openai-gptpro assumption 2026-07-10 Partner-hosted Hopper/Blackwell routing scenario.
Gemini 3.5 Flash.hardwareFleetBlend.TPUv7 0.6 fraction https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Google-dive serving-generation scenario.
Gemini 3.5 Flash.hardwareFleetBlend.TPUv6e 0.3 fraction https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Google-dive serving-generation scenario.
Gemini 3.5 Flash.hardwareFleetBlend.TPUv5e 0.1 fraction https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Google-dive serving-generation scenario.

Throughput, MFU and latency constants — 36 rows

parameter value unit source (URL) source class as-of date note
xAI cash-marginal.referenceActiveParamsB 200 billion parameters https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Central calibration point within the dive’s 100–500B active-parameter range.
xAI cash-marginal.aggregateOutputTokPerSecPerGPU 250 output tokens/s/GPU https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 80 user tok/s × 25 streams ÷ 8 GPUs; only the 80 tok/s user-stream speed is disclosed.
xAI cash-marginal.effectiveMFUOverride 0.02 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 250 × 2 × 200B ÷ 5.0 PFLOP/s = 0.0200; all-in busy-GPU operating point.
xAI cash-marginal.interactivityMult 1 multiplier https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Use with effectiveMFUOverride=0.02; latency and continuous batching are already absorbed.
xAI cash-marginal.fallback.interactivityMultVsGB300Anchor 0.15748031496062992 multiplier https://inference-margins.pages.dev/research/xai-gptpro community 2026-07-10 0.0200 ÷ calculator GB300 anchor MFU 0.127; use only when an MFU override is unavailable, never together with it.
xAI cash-marginal.stackEfficiency 1 multiplier https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 The all-in MFU override already absorbs the stack operating point; no second stack haircut.
xAI cash-marginal.latencyRegime 80 user tok/s; 25 concurrent streams; 8-GPU replica text https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Launch stream speed plus central saturation scenario.
xAI cash-marginal.fallback.assumedGB300AnchorEffectiveMFU 0.127 fraction https://inference-margins.pages.dev/research/xai-gptpro community 2026-07-10 Calculator GB300 effective-MFU anchor used only to derive the fallback interactivity multiplier.
xAI opportunity-cost.referenceActiveParamsB 200 billion parameters https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Central calibration point within the dive’s 100–500B active-parameter range.
xAI opportunity-cost.aggregateOutputTokPerSecPerGPU 250 output tokens/s/GPU https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 80 user tok/s × 25 streams ÷ 8 GPUs; only the 80 tok/s user-stream speed is disclosed.
xAI opportunity-cost.effectiveMFUOverride 0.02 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 250 × 2 × 200B ÷ 5.0 PFLOP/s = 0.0200; all-in busy-GPU operating point.
xAI opportunity-cost.interactivityMult 1 multiplier https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Use with effectiveMFUOverride=0.02; latency and continuous batching are already absorbed.
xAI opportunity-cost.fallback.interactivityMultVsGB300Anchor 0.15748031496062992 multiplier https://inference-margins.pages.dev/research/xai-gptpro community 2026-07-10 0.0200 ÷ calculator GB300 anchor MFU 0.127; fallback only and never together with the override.
xAI opportunity-cost.stackEfficiency 1 multiplier https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 The all-in MFU override already absorbs the stack operating point.
xAI opportunity-cost.latencyRegime 80 user tok/s; 25 concurrent streams; 8-GPU replica text https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Launch stream speed plus central saturation scenario.
xAI opportunity-cost.fallback.assumedGB300AnchorEffectiveMFU 0.127 fraction https://inference-margins.pages.dev/research/xai-gptpro community 2026-07-10 Calculator GB300 effective-MFU anchor used only to derive the fallback multiplier.
China public-cloud on-demand.stackEfficiency 1 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Preserve the selected model/hardware preset’s stack setting; this lens changes procurement and occupancy only.
China public-cloud on-demand.latencyRegime enterprise-interactive text https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 On-demand rate cards are most relevant to elastic interactive use, not throughput-optimized reserved batch pools.
China public-cloud on-demand.interactivityMult 1 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 No extra latency multiplier; preserve the selected operating point.
Ant Group production H20 (SLO replay).model DeepSeek-R1 model https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Operator benchmark model.
Ant Group production H20 (SLO replay).activeParamsB 37 billion active parameters https://github.com/deepseek-ai/DeepSeek-V3 first-party 2026-07-10 DeepSeek-R1/V3 activated parameter count.
Ant Group production H20 (SLO replay).precisionMult 1 multiplier https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ assumption 2026-07-10 Measured MFU is derived against the H20 dense 8-bit peak; do not apply a second precision multiplier.
Ant Group production H20 (SLO replay).dense8bitFLOPS_PF 0.296 PFLOP/s https://inference-margins.pages.dev/research/chinese-accel-gptpro first-party 2026-07-10 H20 dense FP8/INT8 peak used by the project hardware table.
Ant Group production H20 (SLO replay).tier.Base.outputTokPerSecPerGPU 714 output tokens/s/GPU https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Batch 48, TTFT <2 s, TPOT <70 ms.
Ant Group production H20 (SLO replay).tier.Base.effectiveMFU 0.1785 fraction https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Exact arithmetic: 714 × 2 × 37B ÷ 296 TFLOP/s.
Ant Group production H20 (SLO replay).tier.Pro.outputTokPerSecPerGPU 675 output tokens/s/GPU https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Batch 32, TTFT <1.5 s, TPOT <50 ms.
Ant Group production H20 (SLO replay).tier.Pro.effectiveMFU 0.16875 fraction https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Exact arithmetic: 675 × 2 × 37B ÷ 296 TFLOP/s.
Ant Group production H20 (SLO replay).tier.Max.outputTokPerSecPerGPU 423 output tokens/s/GPU https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Batch 12, TTFT <1 s, TPOT <30 ms.
Ant Group production H20 (SLO replay).tier.Max.effectiveMFU 0.10575 fraction https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Exact arithmetic: 423 × 2 × 37B ÷ 296 TFLOP/s.
Ant Group production H20 (SLO replay).representativeTier Pro tier https://inference-margins.pages.dev/research/chinese-accel-gptpro community 2026-07-10 The annex identifies 675–700 output tok/s/GPU as the neutral production anchor.
Ant Group production H20 (SLO replay).effectiveMFUOverride 0.16875 fraction https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Use the measured Pro-tier equivalent directly.
Ant Group production H20 (SLO replay).interactivityMult 1 multiplier https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ assumption 2026-07-10 Use with the MFU override; the measured point already includes SLO, batch, MTP and production software.
Ant Group production H20 (SLO replay).fallback.interactivityMultVsH20Anchor 0.9926470588235294 multiplier https://inference-margins.pages.dev/research/chinese-accel-gptpro community 2026-07-10 0.16875 ÷ calculator H20 anchor MFU 0.17; use only when no MFU override exists.
Ant Group production H20 (SLO replay).fallback.assumedH20AnchorEffectiveMFU 0.17 fraction https://inference-margins.pages.dev/research/chinese-accel-gptpro community 2026-07-10 Calculator H20 effective-MFU anchor used only to derive the fallback multiplier.
Ant Group production H20 (SLO replay).stackEfficiency 1 multiplier https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ assumption 2026-07-10 Measured effective MFU already includes the production software and latency/batch regime.
Ant Group production H20 (SLO replay).latencyRegime Pro: TPOT <50 ms text https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Published middle SLO tier.

Workload, utilization and accounting conventions — 26 rows

parameter value unit source (URL) source class as-of date note
xAI cash-marginal.utilization 1 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Per-occupied, saturated decode-GPU unit-cost convention; not a whole-fleet occupancy claim.
xAI cash-marginal.batchShare 0 fraction https://inference-margins.pages.dev/research/xai-gptpro first-party 2026-07-10 No published Grok 4.5 asynchronous API batch discount; internal continuous batching is already in the 25-stream assumption.
xAI cash-marginal.negotiatedDiscount 0 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 List-price unit-economics perimeter.
xAI cash-marginal.freeTrafficShare 0 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Billed marginal-token perimeter; free and subscription traffic excluded.
xAI opportunity-cost.utilization 1 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Per-occupied, saturated decode-GPU allocation convention; not a whole-fleet occupancy claim.
xAI opportunity-cost.batchShare 0 fraction https://inference-margins.pages.dev/research/xai-gptpro first-party 2026-07-10 No published Grok 4.5 asynchronous API batch discount; internal batching already appears in throughput.
xAI opportunity-cost.negotiatedDiscount 0 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 List-price unit-economics perimeter.
xAI opportunity-cost.freeTrafficShare 0 fraction https://inference-margins.pages.dev/research/xai-gptpro assumption 2026-07-10 Billed marginal-token perimeter; free and subscription traffic excluded.
China public-cloud on-demand.utilization 0.35 fraction https://inference-margins.pages.dev/ assumption 2026-07-10 Enterprise elastic capacity is peak-provisioned; the report uses 35% versus 70% as a salient occupancy contrast.
China public-cloud on-demand.batchShare 0 fraction https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Infrastructure rental does not imply a provider-model asynchronous batch tariff.
China public-cloud on-demand.negotiatedDiscount 0 fraction https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Use posted/public rates; negotiated commitments belong in a separate lens.
China public-cloud on-demand.freeTrafficShare 0 fraction https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Infrastructure procurement lens only.
Ant Group production H20 (SLO replay).utilization 1 fraction https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ assumption 2026-07-10 Per-occupied decode-GPU benchmark convention; do not divide measured throughput by fleet occupancy again.
Ant Group production H20 (SLO replay).batchShare 0 fraction https://www.lmsys.org/blog/2025-09-26-sglang-ant-group/ first-party 2026-07-10 Published batch size 32 is internal continuous batching, not a discounted asynchronous API tariff.
Ant Group production H20 (SLO replay).negotiatedDiscount 0 fraction https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Operating-point preset only.
Ant Group production H20 (SLO replay).freeTrafficShare 0 fraction https://inference-margins.pages.dev/research/chinese-accel-gptpro assumption 2026-07-10 Operating-point preset only.
DeepSeek V4-Flash.nativeInputOutputRatio 3.619047619047619 input tokens/output token https://github.com/deepseek-ai/open-infra-index
https://inference-margins.pages.dev/
community 2026-07-10 Transfer of DeepSeek’s 2025 production trace: 608 input units ÷ 168 output units.
DeepSeek V4-Flash.nativeCacheHitRate 0.5625 fraction https://github.com/deepseek-ai/open-infra-index community 2026-07-10 Transfer of the approximately 56.3% V3/R1 production cache-hit rate, represented as the project’s exact trace fraction.
GLM-4.7.nativeInputOutputRatio 8 input tokens/output token https://inference-margins.pages.dev/research/zhipu-gptpro assumption 2026-07-10 Transferred GLM production-workload scenario; not GLM-4.7 telemetry.
GLM-4.7.nativeCacheHitRate 0.41 fraction https://inference-margins.pages.dev/research/zhipu-gptpro assumption 2026-07-10 Transferred GLM production-workload scenario; not GLM-4.7 telemetry.
GPT-5.6 Terra.nativeInputOutputRatio 9 input tokens/output token https://inference-margins.pages.dev/research/openai-gptpro community 2026-07-10 Transfer of Sol-dive workload: 7 cached + 2 fresh input : 1 output.
GPT-5.6 Terra.nativeCacheHitRate 0.7777777777777778 fraction https://inference-margins.pages.dev/research/openai-gptpro community 2026-07-10 7 cached ÷ 9 total input, transferred from the Sol dive.
GPT-5.6 Luna.nativeInputOutputRatio 9 input tokens/output token https://inference-margins.pages.dev/research/openai-gptpro community 2026-07-10 Transfer of Sol-dive workload: 7 cached + 2 fresh input : 1 output.
GPT-5.6 Luna.nativeCacheHitRate 0.7777777777777778 fraction https://inference-margins.pages.dev/research/openai-gptpro community 2026-07-10 7 cached ÷ 9 total input, transferred from the Sol dive.
Gemini 3.5 Flash.nativeInputOutputRatio 15 input tokens/output token https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Provider-native traffic scenario; not a disclosed Gemini 3.5 traffic trace.
Gemini 3.5 Flash.nativeCacheHitRate 0.6 fraction https://inference-margins.pages.dev/research/google-gptpro assumption 2026-07-10 Provider-native traffic scenario; not a disclosed Gemini 3.5 traffic trace.

Delta notes