Write a procurement brief using ONLY the frozen official-source summaries supplied after this prompt. They describe October 8, 2026; do not browse or update them. The customer requires eligible EU inference residency and uses the Responses API. This is a source-synthesis exercise, not a live-web search or a measurement of actual latency.

Calculate two request shapes. Billed output already includes reasoning.
A: 80,000 cached input + 20,000 ordinary input + 5,000 billed output; no cache writes.
B: 280,000 ordinary input + 8,000 billed output; no cached input or cache writes.
Apply the stipulated 10% regional-processing uplift to BOTH tiers. For each shape calculate Standard cost, Ultrafast cost, extra cost, and break-even USEFUL waiting-time savings at USD60/hour. Calculate the token cost of 1,000 shape-A regional requests, with 900 Standard and 100 Ultrafast. Return Sol Ultrafast TPM allowances for Build, Launch and Grow.

Return exactly this JSON structure (descriptions below are placeholders, not literal values):
{
  "A": {"standard_usd": number, "ultrafast_usd": number, "premium_usd": number, "break_even_seconds": number},
  "B": {"standard_usd": number, "ultrafast_usd": number, "premium_usd": number, "break_even_seconds": number},
  "mixed_1000_usd": number,
  "sol_ultrafast_tpm": {"Build": integer, "Launch": integer, "Grow": integer},
  "memo": "250–450 words"
}

All monetary fields are USD; break_even_seconds is seconds; TPM is tokens per minute. The evaluator checks all 12 numeric fields against source-derived arithmetic, using absolute tolerance 0.00001 inclusive. Do not round small costs to cents; retain enough precision to meet that tolerance. Qualification purchases and monthly usage ceilings must not be described as monthly subscription fees or free spending.

EVERY acceptance requirement follows. All must pass; none is hidden.
O01: Return exactly one valid JSON object, without Markdown fences or surrounding commentary or duplicate object keys. Use exactly the keys and nested structures below. Numeric fields must be finite JSON numbers, never strings, booleans, null, NaN or Infinity. TPM values must be JSON integers. Memo must be a string.
O02: Memo length must be 250–450 words inclusive, counted by splitting its decoded string on whitespace. Include all F01–F12 requirements in that memo; information appearing only in numeric fields does not satisfy a memo requirement.

N01: A.standard_usd must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N02: A.ultrafast_usd must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N03: A.premium_usd must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N04: A.break_even_seconds must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N05: B.standard_usd must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N06: B.ultrafast_usd must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N07: B.premium_usd must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N08: B.break_even_seconds must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N09: mixed_1000_usd must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N10: sol_ultrafast_tpm.Build must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N11: sol_ultrafast_tpm.Launch must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.
N12: sol_ultrafast_tpm.Grow must match the source-derived calculation or allowance within absolute tolerance 0.00001; TPM fields remain integers.

F01: State that API token billing is separate from subscription access and allowances. A subscription does not supply API tokens.

F02: State that subscription Ultrafast is available on Pro USD500/month and eligible Enterprise/Edu plans, with Enterprise owner enablement required. State explicitly that Pro USD500 is not required for Responses API access; API access is subject to limits.

F03: Distinguish Sol US/EU residency and global processing from Astra Ultrafast US/global processing, with no EU option documented in this frozen pack. Explain that non-US residency requires eligibility/approval for abuse-monitoring controls and a Modified Retention amendment; customer location alone does not establish inference location. Name the EU endpoint and define Europe as EEA plus Switzerland.

F04: State that more than 272,000 input tokens puts the full request, including all input and output, on long-context rates; the higher rate is not confined to excess input.

F05: Explain that ordinary input, cached input and cache-write input are exclusive pricing categories. A cache-write rate replaces another input rate for the same tokens and is not an additive fee. No cache writes are present in shapes A/B; their stipulated cached counts are calculation inputs, not guaranteed future hits.

F06: State that billed output includes invisible reasoning and consumes the output budget. The supplied billed output counts already include reasoning; do not add another reasoning charge.

F07: Explain that TPM is a traffic/rate-limit allowance, not a single response's generation speed. Ultrafast limits are separate from Standard/Fast.

F08: State that the supplied official guide gives no numerical Sol throughput or task-speed guarantee. Do not invent measured latency, speed, quality, success rates or retry savings. Do not claim that no independent Sol measurement exists anywhere: this task is limited to the supplied archived pack.

F09: Attribute the up-to-8x token-generation claim specifically to Astra in Codex. Do not apply it to Sol or completed coding/research tasks.

F10: Distinguish subscription included usage at 8x Standard from purchased credits/Enterprise pay-as-you-go at 6x. These usage multipliers are not speed measurements or the API-key billing mechanism.

F11: Recommend Ultrafast ONLY when both conditions hold: (a) measured USEFUL waiting-time savings exceed the TOTAL extra task cost, including all attempts, tools and applicable regional charges; and (b) the accepted output meets the SAME factual-accuracy, citation-support and requirement-coverage checks in both tiers. Explicitly name those three quality checks in the memo. Treat this quality gate as the customer's acceptance policy, not an OpenAI vendor claim. Default to Standard if either condition is unmet.

F12: Support factual claims with the nearest applicable source IDs S1–S6 in the same or immediately adjacent sentence. IDs must actually support the claim; a decorative list is insufficient. Label cost/break-even figures as calculations rather than observed savings. Do not add unsupported claims, contradictory facts or outside sources. Apply the stipulated 10% regional uplift to both tiers and explain the calculation basis.

Acceptance requires valid structure and word count, correct values in all 12 numeric fields, and every F01–F12 memo requirement. A correct cost calculation alone does not satisfy the recommendation's quality gate. No exact wording is required: preserve the specified meaning and name the three F11 quality checks.


Frozen source pack:
# Frozen official-source summaries for the research pilot

Prepared October 8, 2026 from the archived official documentation in ../sources/. These are factual summaries for a controlled synthesis task, not live search results or measurements.

## S1 — October 8 API changelog
URL: https://developers.openai.com/api/docs/changelog

On October 8 OpenAI added Ultrafast to GPT-6.1 Sol in the Responses API. The model ID is gpt-6.1-sol and the request service tier is ultrafast. The announcement says all API users can access it subject to rate limits, with global processing and US and EU data residency. The underlying Sol model was released September 29.

## S2 — Ultrafast API guide
URL: https://developers.openai.com/api/docs/guides/ultrafast-mode

The guide supports HTTP and recommends persistent WebSockets for agents with repeated tool calls. Sol Ultrafast default TPM is 1,000,000 for Build, 4,000,000 for Launch and 40,000,000 for Grow. Ultrafast has limits separate from Standard and Fast. Sol Ultrafast supports US/EU residency and global processing; Astra Ultrafast supports US residency and global processing. The current guide supplies no numeric Sol throughput or task-speed guarantee.

## S3 — API pricing and Sol model pricing notes
URLs: https://developers.openai.com/api/docs/pricing?latest-pricing=ultrafast and https://developers.openai.com/api/docs/models/gpt-6.1-sol

All following prices are USD per million tokens. Sol Standard short context: ordinary input 2, cached input 0.10, cache write 2.50, output 10. Sol Ultrafast short context: ordinary input 12, cached input 0.60, cache write 15, output 60. Standard long context: ordinary input 4, cached input 0.20, cache write 5, output 15. Ultrafast long context: ordinary input 24, cached input 1.20, cache write 30, output 90. More than 272,000 input tokens puts the entire request on long-context rates. Regional processing adds 10% for eligible models released on or after March 5, 2026. GPT-6.1 Sol has 1,050,000 context, 922,000 max input, and 128,000 max output tokens. API reasoning effort supports low/medium/high/xhigh/max; medium is default.

## S4 — Cache accounting and reasoning
URLs: https://developers.openai.com/api/docs/guides/prompt-caching and https://developers.openai.com/api/docs/guides/reasoning

Cache writes are not an additive fee: input tokens use ordinary input, cached-input or cache-write pricing. A shared prompt prefix can reuse cached work, but a cache hit is not guaranteed. Invisible reasoning tokens are billed as output and consume the generated-output budget. Count all attempts and tools in task cost.

## S5 — Work/Codex speed and subscription pricing
URLs: https://learn.chatgpt.com/docs/agent-configuration/speed and https://learn.chatgpt.com/docs/pricing

Subscription Ultrafast access is on Pro USD500/month and eligible Enterprise/Edu plans. Other self-serve plans do not get subscription Ultrafast through purchased credits. Enterprise access is disabled by default and requires an owner to enable it. Eligible Enterprise agreements are credit-based or USD usage-based; eligible Edu uses credits. Legacy Enterprise rate-limit plans are unsupported. Included usage drains at 8x Standard; purchased credits and Enterprise pay-as-you-go bill at 6x. API-key usage instead uses API token pricing, separate from subscription allowances. Work and Codex share usage. The speed page's up-to-8x token-generation statement specifically names Astra in Codex, not Sol or completed coding/research tasks.

## S6 — Residency and API tiers
URLs: https://developers.openai.com/api/docs/guides/your-data and https://developers.openai.com/api/docs/guides/rate-limits

Europe for inference residency means EEA plus Switzerland. Non-US residency requires approval for abuse-monitoring controls and a Modified Retention amendment. Customer location alone does not establish inference location. Supported regional API base URLs are https://us.api.openai.com/v1 and https://eu.api.openai.com/v1. Build qualifies at USD5 cumulative credit purchases, Launch at USD100, Grow at USD500. Listed monthly usage limits are USD500/5,000/200,000 respectively, and these are not free included spending. Qualification thresholds are not monthly subscription fees. TPM allowances do not establish an individual response's generation speed.
