Autoresearch: OpenAI / Anthropic pricing power and marks since Kimi K3 (as of 2026-09-19)

Generated by /autoresearch on 2026-09-19. Synthesized across 3 rounds from 11 web pages, anchored by the Grokipedia OpenAI entry. See Provenance. Treat as raw material — review before promoting into a project or thread. Context: vault/projects/stock-market

Scope of this pass: research only. This file is written outside the vault. It does not update wiki/mechanisms/open-weight-sputnik-to-frontier-lab-derate and does not assign a buy, sell, or size.

Summary

As of 2026-09-19, the Ismail "~75% derate" of frontier-lab marks remains unobserved at any company-organized print. OpenAI’s August 10, 2026 employee tender was reported at $852 billion — flat versus the March 2026 primary (TechCrunch, Bloomberg-first). Anthropic’s last first-party mark is still the May 28 Series H at $965 billion post-money with run-rate revenue above $47 billion (Anthropic). Microsoft, the clean listed proxy, booked FY2026 net gains on its OpenAI investment of $4.963 billion and a discrete $3.2 billion Q4 gain on Anthropic (Microsoft Q4 FY2026), and disclosed $24.1 billion of FY2026 revenue from OpenAI commercial arrangements including revenue-sharing (Microsoft 10-K note).

Revenue growth continued after Kimi K3 on the prints that can be dated. Pricing is the mixed half of the Step 3 falsifier: OpenAI did cut GPT-5.6 Luna (−80%) and Terra (−20%) on July 30 and put Sol on a time-boxed promo on August 21 (OpenAI, API changelog), while Anthropic held Fable 5 / 5.1 at $10 / $50 per MTok and converted Sonnet 5’s $2 / $10 intro rate into the standard price rather than taking the scheduled September 1 hike (Anthropic pricing, release notes, Aug 10). That is mid-tier list-price compression plus a cancelled hike — not a collapse of the frontier SKU, and not a marked 75% valuation derate.

Step 3 of open-weight-sputnik-to-frontier-lab-derate should stay open. The falsifier is conjunctive (sustain revenue growth and pricing). Revenue growth is the stronger observed leg; pricing is mixed; the valuation-premium derate has no company-organized print. Do not move the step to confirmed or to partial on Ismail’s magnitude.

Findings

Company-organized marks did not print a post-Kimi 75% derate

Grokipedia’s OpenAI primer (fetched 2026-09-19, 3,017 characters) still frames the company as private, with valuation “approximately $850 billion (March 2026)” and Microsoft at a 27% major-shareholder figure in the infobox (Grokipedia: OpenAI). That March print is the last encyclopedic anchor, not a post-July re-rate.

The only dated post-Kimi organized OpenAI mark located this pass is the August 10, 2026 employee tender. TechCrunch, attributing Bloomberg and stating OpenAI did not comment, reports OpenAI bought back $7 billion of employee shares at $852 billion — the same valuation as the March primary that added $122 billion (TechCrunch). A flat tender after Kimi K3 is the opposite of Ismail’s finger-in-the-air “about a quarter of what they were three months ago.” It is also not a step-up. Absence of a step-up is a weaker signal than a markdown; it is still not a 75% derate.

Anthropic has no first-party post-May 2026 valuation print in the pages fetched this pass. The official Series H (May 28, 2026) is $65 billion raised at $965 billion post-money, led by Altimeter, Dragoneer, Greenoaks, and Sequoia, and states run-rate revenue “crossed $47 billion earlier this month” (Anthropic). That mark is pre-Kimi K3. Treating it as a post-July derate would be a date error.

Informal September 2026 secondary composites that imply OpenAI near $475 billion (~44% below the March/August $852 billion print) appeared in search results on tracker sites. Those pages were not selected for fetch (aggregator / non-primary; not a completed company tender). They are recorded here only as an unfetched contradiction to the August 10 $852 billion print — not as an observed mark.

Listed-proxy accounting: Microsoft marked the relationship up through June 30, 2026

Microsoft is the listed proxy with a first-party OpenAI line. The FY2026 Q4 release (period ended June 30, 2026; released July 29 — ~10 days after the Kimi K3 Moonshots pod) reports:

• FY2026 revenue $331.8 billion, +18% (Microsoft Source).
• FY2026 net gains from investments in OpenAI of $4,963 million ($0.67 of diluted EPS); Q4 alone $480 million / $0.07. FY2025 had been a $3,620 million loss on the same line (Microsoft Source).
• A discrete Q4 item: “a $3.2 billion gain from our investment in Anthropic” among items that added $0.27 to diluted EPS versus April 29 guidance (Microsoft Source).
• Azure and other cloud services +43%; Microsoft Cloud $59.3 billion, +27%; commercial remaining performance obligation $678 billion, +84% (Microsoft Source).

The FY2026 Form 10-K accounting-policies note is the primary related-party disclosure. Microsoft holds “an approximate 25% interest on an as-converted basis,” signed a new definitive agreement in October 2025, and “recorded revenue from commercial arrangements with OpenAI, inclusive of revenue-sharing payments, of $24.1 billion” for FY2026, with $6.0 billion accounts receivable from OpenAI as of June 30, 2026. Funding commitments on the investment are $13.0 billion, of which $11.9 billion was funded. The same note says FY2026 ownership decreased in the October 2025 recapitalization “and we recorded dilution gains” (Microsoft 10-K, Note 1 / R11).

Two constraints on that $24.1 billion: it is Microsoft revenue from arrangements with OpenAI (Azure consumption + revenue-share + other commercial), not OpenAI’s own ARR, and the 10-K does not break out margin. The investment-gain line is an accounting mark, not a secondary-market print. Even so, a listed proxy that books multi-billion FY gains on OpenAI and a $3.2 billion Q4 Anthropic gain is not, through June 30, 2026, showing the Ismail derate. The Q4 release is dated July 29, so it can include some post-Kimi-announcement days but is not a September 19 mark.

Official API list prices: mid-tier cuts at OpenAI; frontier SKUs held at Anthropic

This is the only half of Step 3 with dated, first-party list-price movement after Kimi K3.

OpenAI, July 30, 2026 (official blog): GPT-5.6 Luna “will cost 80% less”; GPT-5.6 Terra “20% less.” Starting July 30, API pricing is $2 / $12 per million input/output tokens for Terra and $0.20 / $1.20 for Luna. “Sol pricing remains unchanged” on that day. ChatGPT and Codex subscription prices and quota budgets “remain unchanged,” while Terra and Luna consume fewer credits (OpenAI, July 30). The API changelog repeats the same July 30 Luna/Terra cuts (OpenAI API changelog).

OpenAI, August 21, 2026 (changelog): GPT-5.6 Sol moves to $4 / $20 per million tokens — “20% lower input and 33% lower output.” The cut is labeled promotional, “available at least through November 21, 2026” (OpenAI API changelog). The live pricing table fetched this pass shows gpt-5.6-sol at $4 / $20 short-context, gpt-5.6-terra at $2 / $12, gpt-5.6-luna at $0.20 / $1.20, and a later frontier SKU gpt-6-astra at $10 / $50 (OpenAI API pricing). Astra’s $10 / $50 is a new premium list, not a cut of an existing ID. Changelog dates Astra’s API release to September 8, 2026 (OpenAI API changelog).

Anthropic’s fetched price card still lists Claude Fable 5 and Fable 5.1 at $10 / $50 per MTok (Mythos 5 / 5.1 the same). Opus 5 is $5 / $25. Sonnet 5 is $2 / $10. The page states that the $2 / $10 Sonnet 5 rate, “announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3 / $15 … on September 1, 2026 will not occur” (Anthropic pricing). Release notes date that conversion to August 10, 2026 (Anthropic API release notes).

Read for the falsifier: OpenAI did compress list prices on the high-volume GPT-5.6 IDs in the two–five weeks after Kimi K3, and framed the July 30 cuts as passing through efficiency (“intelligence too cheap to meter” is a customer quote OpenAI chose to publish — OpenAI, July 30). Anthropic did not cut Fable. It forwent a planned 50% Sonnet 5 hike. That is pricing restraint, not a Fable collapse. Subscription sticker prices at OpenAI were explicitly held on July 30 (OpenAI, July 30).

Revenue / ARR prints: growth continued; the biggest post-July figures are leaks, not 10-Qs

OpenAI’s last first-party ARR series on a fetched page is Sarah Friar’s January 18, 2026 essay: “$2B ARR in 2023, $6B in 2024, and $20B+ in 2025,” tracking compute 0.2 → 0.6 → ~1.9 GW (OpenAI / Friar). That is pre-Kimi.

The August 14, 2026 “~$40 billion annualized revenue, enterprise overtook consumer” print is not on openai.com. The fetched syndication is a Value Add Pulse recap of CNBC, quoting Friar via a meeting attendee: “We entered the year at 60-40, but enterprise has accelerated much faster than expected and those lines have now crossed,” with “annualized revenue now running at roughly $40 billion” and “business-customer revenue rising 32% in a single month” (Value Add Pulse, Aug 14, citing CNBC — CNBC itself was not fetched; it is Persistent-failure in SOURCE_RELIABILITY.md). Treat $40B / +32% enterprise as leak-grade, CNBC-attributed, consistent with growth if the attendee quotes hold, not as an audited print. Direct Yahoo Finance syndication of the same meeting timed out this pass.

Anthropic’s last first-party run-rate is the Series H “crossed $47 billion” (Anthropic, May 28). Search results describe later investor-update leaks: $65 billion annualized at end-July (Bloomberg Aug 17 / CNBC Aug 17) and >$100 billion annualized by mid-September (New York Times Sept 18, Bloomberg/Axios syndications). Axios, Forbes, and Bloomberg were not successfully fetched this pass (timeout / empty / paywall). Those later figures stay in Provenance as unfetched leaks. They point with revenue growth, not against it, but they are not first-party and must not be used to flip a wiki step.

Microsoft’s $24.1 billion OpenAI-arrangement revenue is a FY2026 (year ended June 30) related-party total, so it is mostly pre-Kimi as a run-rate and cannot by itself prove post-July OpenAI growth (Microsoft 10-K R11).

What this does — and does not — do to Step 3

The mechanism’s Step 3 claim is: frontier-lab pricing power and API-revenue growth compress → the terminal-value premium in OpenAI/Anthropic (~$1T private marks) derates. The written falsifier is: “OpenAI/Anthropic sustain revenue growth and pricing through 2026.”

Sacks’s prior contra (closed models winning paid enterprise-spend share, 19% → 11% open-source) is not re-tested this pass; the Friar leak’s enterprise-over-consumer crossover is directionally consistent with paid wallet staying at closed labs, but it is a mix shift inside OpenAI, not an open-vs-closed spend share. SemiAnalysis Aug 21 (open-vs-closed gap halves each era; still prefers Fable for real work) was not re-fetched; it remains the already-logged partial attach on Step 1, not new Step 3 evidence.

Recommendation for the wiki (this file does not write it): leave Step 3 open. A partial tag on mid-tier list-price compression only would be defensible if a later editor wants to split the step; it would still be wrong to treat Ismail’s 75% mark, or a terminal-value collapse, as observed. Confirmed would require either (a) a company-organized or 13F/10-K impairment-quality markdown of the premium, or (b) a clean first-party print that both revenue and frontier list prices broke. Neither is on the table as of 2026-09-19.

Contradictions and open questions

• Flat $852B tender vs unfetched ~$475B September secondary composites. If a later pass can fetch a named secondary venue (Nasdaq Private Market, Forge, etc.) with dated trades, that could put a real discount on the March/August primary. Until then, the organized print controls.
• $40B OpenAI ARR and $65B / $100B Anthropic run-rates are leaks. CNBC/Bloomberg/NYT were not fetched (paywall / Persistent-failure / timeout). A first-party blog, S-1, or Microsoft 10-Q that repeats them would graduate those figures.
• Pricing vs revenue can both be true. OpenAI cut unit prices and (if the Friar leak holds) grew ARR. That is Jevons / mix-shift, not by itself a terminal-value derate. The Step 3 falsifier’s “and pricing” clause is the unresolved piece.
• Microsoft’s $24.1B is not OpenAI ARR, and FY2026 ends June 30 — mostly pre-Kimi as a growth test.
• No Anthropic first-party mark after May 28. IPO-valuation talk ($2T class) appeared in search; Forbes returned an empty body this pass. Do not treat a rumored IPO range as a mark.
• Paid open-source enterprise-spend share (Sacks 19% → 11%) and SemiAnalysis Fable-preference were not re-measured here.
• Labs remain private. There is still no public equity print that can confirm or kill Ismail’s magnitude.

Provenance

Rounds run: 3 (full)

Sub-questions by round:

Round 1 (broad survey):

1. What post-July 2026 private-market valuation marks exist for OpenAI and Anthropic?
2. Have OpenAI or Anthropic published official API list-price changes since July 2026?
3. What first-party or Microsoft-disclosed revenue/ARR prints exist since July 2026?
4. What does Microsoft’s latest 10-K / earnings recap say about OpenAI economics?
5. Have listed proxies shown a derate attributable to frontier-lab pricing since July?

Round 2 (drill-down):

1. What does Microsoft’s FY2026 10-K disclose as OpenAI commercial revenue and the investment mark? — targeted the failed Neowin 10-K recap
2. Did Anthropic cut Fable/Opus list prices after Kimi K3, or only freeze Sonnet 5? — targeted the pricing-power half
3. Can the Sept 18 Anthropic $100B ARR / $2T IPO talk be sourced to a fetchable syndication? — targeted post-May marks (Forbes timed out / empty)
4. Can the Aug 14 OpenAI $40B ARR / enterprise-crossover print be sourced off CNBC? — targeted the revenue-growth half (Yahoo timed out; Value Add Pulse fetched in R3)

Round 3 (resolve remaining uncertainty):

1. Can a fetchable page confirm the Aug 14 OpenAI $40B ARR / July growth print? — resolved as leak-grade via Value Add Pulse, not first-party
2. Can a fetchable page confirm Anthropic’s July $65B / Sept $100B run-rate and a post-May official mark? — not resolved; Axios timeout; no post-May official mark found
3. Does the OpenAI changelog date the Sol promo as a mid-life list-price cut? — yes: Aug 21, promotional through at least Nov 21, 2026

Anchor source (Grokipedia, fetched before round 1):

• OpenAI — 3,017 chars extracted via _lib/grokipedia.py (source: api, 365 citations on the live page) — primer: private PBC, ~$850B March 2026, Microsoft 27% in infobox, ChatGPT 900M WAU by March 2026. Encyclopedic only; not used as the sole source for contested 2026 marks.

URLs fetched (11 successful, 5 failed):

Round 1:

• Advancing the price-performance frontier with GPT-5.6 — official blog — July 30 Luna −80% / Terra −20%; Sol unchanged that day; subscription stickers held
• Pricing | OpenAI API — official docs — live $4/$20 Sol, $2/$12 Terra, $0.20/$1.20 Luna, $10/$50 Astra; Sol promo footnote through Nov 21
• Anthropic raises $65B Series H at $965B — official — last first-party mark ($965B) and $47B run-rate (May 28, pre-Kimi)
• Microsoft Cloud and AI strength fuels fourth quarter results — official IR — FY26 OpenAI investment gain $4.963B; Q4 Anthropic gain $3.2B; Azure +43%
• [Failed: https://www.neowin.net/news/microsoft-reveals-it-generated-241-billion-in-revenue-from-openai-in-fiscal-2026/] — Cloudflare challenge; routed to SEC 10-K in round 2
• OpenAI reportedly completed a $7 billion employee tender offer — news (Bloomberg-first) — $7B buyback at $852B, flat vs March; company no comment

Round 2:

• Anthropic model pricing — official docs — Fable 5/5.1 still $10/$50; Sonnet 5 $2/$10 now standard; scheduled $3/$15 hike cancelled
• Claude Platform release notes — official — Aug 10, 2026 dated conversion of Sonnet 5 intro price to standard
• Microsoft 10-K Note 1 excerpt (R11) — SEC primary — $24.1B OpenAI commercial revenue, $6.0B AR, ~25% as-converted, $13.0B / $11.9B funding, dilution gains
• A business that scales with the value of intelligence — official (Friar, Jan 18, 2026) — first-party $2B / $6B / $20B+ ARR path through 2025
• [Failed: https://finance.yahoo.com/markets/stocks/articles/openai-investors-must-consider-latest-171700821.html] — timeout (Friar $40B syndication)
• [Failed: https://www.forbes.com/sites/jimosman/2026/09/17/anthropic-ipo-could-hit-2-trillion-and-put-public-investors-last/] — empty body / timeout (IPO-range rumor)

Round 3:

• Changelog | OpenAI API — official — Jul 30 Luna/Terra cuts; Aug 21 Sol $4/$20 promo through at least Nov 21; Sep 8 GPT-6 Astra
• OpenAI enterprise revenue overtakes consumer: $40B ARR — VC recap of CNBC — leak-grade $40B ARR + enterprise/consumer crossover + 32% monthly business-revenue claim
• [Failed: https://www.axios.com/2026/09/18/anthropic-100-billion-revenue] — timeout (NYT $100B syndication; search snippet visible, not treated as fetched)

Search-visible, not fetched (disclosed so they are not laundered as primary):

• CNBC 2026-08-14 Friar investor meeting ($40B ARR) — Persistent-failure domain; cited only as the URL Value Add Pulse points at
• Bloomberg / CNBC 2026-08-17 Anthropic $65B end-July run-rate — paywall / Persistent-failure
• New York Times 2026-09-18 Anthropic >$100B annualized (Bloomberg and Axios syndications) — not fetched
• Private Market View / Crypto Briefing September composite implying OpenAI ~$475B — aggregator; not fetched

Tools used: WebSearch, WebFetch, _lib/grokipedia.py fetch "OpenAI" --max-chars 3000. No --include-x. Priors skipped per user. Generated: 2026-09-19 (America/Los_Angeles date on the task; fetch timestamps UTC).