Open-weight AI adoption
How much real AI workload is moving from closed to open-weight models, and where does the economic value of that shift go?
Edited by Finquill Research · Updated Oct 27, 2026
Review cadence: Weekly review of OpenRouter's open/closed chart; re-checked on the monthly Vercel index, flagship model or licence releases, and named earnings dates. Not daily.
Reviewed by the editor · Analytics, not advice
Free account. See what changed since your last visit, and the dates to watch on your calendar. Following never means agreeing.
Short answer (data to 21 Oct 2026). On the two routing platforms that publish usage, open-weight models now carry most tokens but a minority of the money: about 72% of OpenRouter tokens while closed-weight models still take 73% of its spend, and on Vercel's AI Gateway 71.5% of September tokens against 23% of spend. Most of that volume runs on Chinese labs' models, and a growing share of the dollars spent on them goes to US-based hosts. Flagship licences are adding commercial conditions, and measured capability gaps run from about four to eight months. These are platform figures, not market share: no hyperscaler or enterprise-wide split of open versus closed use has been published.
What changed in this update
The first update, published Oct 27, 2026.
What's next
Dated tests · set down in advance, graded by the editor
Due Oct 30, 2026 · awaiting the editor's grade · AMZN
Amazon's Q3 2026 earnings release names at least one open-weight model family (for example DeepSeek, Qwen, Kimi, GLM, Gemma, gpt-oss or Nemotron) among models added to Amazon Bedrock. Recorded as a pre-registered test, not a forecast.
Jun 30, 2027 · META
By 30 Jun 2027, Meta publishes downloadable weights for a new Muse-family or Llama-successor language model on Hugging Face or an official Meta site. Recorded as a pre-registered test, not a forecast.
Jun 30, 2027 · BABA
The next Qwen flagship language model released with downloadable weights after Qwen3.8-Max ships under a licence with a model-as-a-service condition, not Apache-2.0 or MIT. Recorded as a pre-registered test, not a forecast.
Open questions
- DeepSeek's per-request billing for Search and Fetch tool calls (from 1 Nov): does DeepSeek's share of OpenRouter tokens or requests change in November?
- Vercel Production Index with October data (mid-November, date unannounced): does open-weight spend share move from 23%, and does the US-provider share of spend on Chinese models stay near 70%?
- AWS re:Invent (30 Nov to 4 Dec): does Bedrock add a DeepSeek V4-series model? On 21 Oct it lists only V3.2, V3.1 and R1.
- Weekly OpenRouter chart: does open-weight token share stay above the 62% level OpenRouter's newsletter cites (72.0% in the week of 12 Oct)?
Already happened
These dates have passed; the questions are kept as they were asked.
- Oct 28, 2026 · Microsoft FY27 Q1 (28 Oct): does the release or 10-Q mention open-weight models in Foundry, which listed DeepSeek V4 variants and Kimi 2.6 via Fireworks in May?
- Oct 29, 2026 · Amazon Q3 (29 Oct): does the release name an open-weight family among Bedrock additions, as Q2's named Gemma 4? Alphabet Q3 (date unconfirmed): does it update the 22 billion Gemini API tokens a minute?
The analysis
Tokens and spend where they are published
- OpenRouter: open-weight models took 72% of tokens in the week of 14 Sep, Chinese labs 69%. Its weekly chart, which excludes stealth models, shows 72.0% for the week of 12 Oct, out of 164.0 trillion tokens, and 66.5–75.2% in every full week since early July (read from the chart data).
- Closed-weight models took 73% of OpenRouter spend in the second week of September, down from 89% in the week of 5 Jan. Monthly spend on Moonshot AI, Z.ai and DeepSeek rose tenfold or more from January to August, though together they stay in the high single digits of all spend.
- Vercel AI Gateway, September: open-weight share of spend more than quadrupled since June. Since December, open-weight spend grew more than seventyfold and frontier-model spend tenfold.
- DeepSeek's models processed 58% of Vercel tokens for 6.6% of spend; Moonshot's Kimi took 3.6% of tokens and 13.2% of spend.
Who earns on Chinese open-weight models
- Nearly 70% of Vercel's September spend on Chinese models went through US-based providers, up from less than 10% in December. US-based providers served nine in ten Kimi tokens. DeepSeek served more than four in five of its own tokens, yet two in five dollars spent on its models went to US-based providers.
- Cloud catalogues differ by lab. On 21 Oct, AWS Bedrock lists Kimi K3 and GLM 5.3 but, for DeepSeek, only V3.2, V3.1 and R1. Microsoft said in May that DeepSeek V4 variants are in Foundry, with DeepSeek V4 Pro and Kimi 2.6 also offered through Fireworks. Amazon's Q2 release named Google DeepMind's Gemma 4 among 10+ models added to Bedrock.
- Alibaba opened the weights of its flagship Qwen3.8-Max (2.4 trillion parameters) in August and reported AI-related product revenue of RMB12.4bn for the quarter to June, a twelfth quarter of triple-digit growth.
Licences are diverging
- DeepSeek published V4.1-Flash-1017 on 17 Oct with the repository and weights under the MIT License.
- Flagship licences add conditions for companies that sell model access as a service. Kimi K3 (27 Jul) requires a separate agreement with Moonshot AI above $20m of revenue over 12 months. Qwen3.8-Max (12 Aug) requires a separate licence above $50m, and also covers AI Work Assistant businesses. GLM-5.3 (28 Aug) requires a Z.ai security review above $10bn. The Qwen Community License on Qwen3.8-Flash-Next requires a separate licence for such businesses at any revenue.
- Meta's Q2 10-Q says it makes AI technology available under open source, commercial and non-commercial licences, and its Llama 4 licence requires a licence from Meta above 700 million monthly active users. Its Q3 release (21 Oct) names Muse and Muse Wonder as its first frontier model and describes no new downloadable weights.
How far behind
- Epoch: about four months, or 8 index points, on average from January to May 2026. Stanford's AI Index: 49 Arena points (3.4%) in March. NIST's CAISI: about 8 months for DeepSeek V4 in its April evaluation, which includes non-public benchmarks.
- Cyber-only measures differ by evaluator: CAISI put GLM-5.3 about four months behind the US frontier; Epoch puts the open-weight cyber frontier 6–7 months behind.
Price lists
- DeepSeek V4.1-Flash lists at $0.30 per million input tokens (cache miss) and $1.20 per million output at peak, half that off-peak. Claude Haiku 5.5 lists at $0.10 and $0.50 for prompts up to 100,000 tokens. Tokenizers differ, so per-token prices are not like-for-like.
- From 1 Nov, DeepSeek bills per request for Search and Fetch tool calls made through its Harness, Claude Code or Codex; they are free until then.
Related. The debt, leases and capex behind the hosts are covered in Financing the AI buildout. Chip exports appear here only as context: a BIS rule of 15 Jan moved H200-class exports to China from a presumption of denial to case-by-case review, subject to conditions, and NVIDIA's outlook for its quarter to late October assumes no China data-centre compute revenue.
Strongest alternative
Router data may overstate how much paid work has moved. Both platforms skew to developers, startups and agent loops. OpenRouter says cached prompt tokens were just under 90% of agentic tokens in September but only about a third of the bill in its worked example, and that its rankings show traffic routed through OpenRouter. Vercel estimates spend at list prices and broadened its open-weight definition this month, so its series breaks. On spend, closed-weight models still take about three-quarters on both platforms (73% on OpenRouter; about 77% on Vercel, derived).
The largest channels are not in the data. Alphabet said in July that Gemini models process 22 billion API tokens a minute, about 222 trillion a week (derived), more than OpenRouter's whole 164.0 trillion in the week of 12 Oct. No hyperscaler splits open from closed use, so enterprise traffic on direct vendor APIs and clouds may still be mostly closed.
Closed vendors also list low-price tiers (Claude Haiku 5.5 at $0.10 input and $0.50 output per million tokens), and a four-to-eight-month capability lag still matters for frontier work. As flagship licences add commercial conditions, the value of open-weight models may flow to the labs' own clouds and licensing deals rather than to independent hosts. AWS lists Kimi K3 and GLM 5.3, whose licences require an agreement or a review for large model-as-a-service businesses; the terms AWS operates under are not public.
Developments
Oct 21, 2026
Meta's Q3 release names Muse as its first frontier model; no new weights described
Meta's Q3 2026 results release quoted its CEO on launching Muse and Muse Wonder, "our first frontier model". The release does not mention downloadable weights; the Llama 4 licence that governs Meta's most recent downloadable models took effect in April 2025.
Tests whether the largest US publisher of open weights is keeping its new frontier models API-only; the Meta watch-claim tracks this to mid-2027.
1 report · one development
Reports sharing an origin are coverage, not independent confirmation.
Oct 17, 2026
DeepSeek publishes V4.1-Flash-1017 under MIT; per-request tool billing set for 1 Nov
DeepSeek published DeepSeek-V4.1-Flash-1017 with its weights under the MIT License. Its price list says Search and Fetch tool calls made through DeepSeek Harness, Claude Code or Codex are free until per-request billing starts on 1 Nov.
Keeps the most-used open-weight family under a permissive licence, and sets a dated price change to check against November platform data.
2 reports · one development
Reports sharing an origin are coverage, not independent confirmation.
Oct 15, 2026
Vercel: open-weight models at 71.5% of tokens and 23% of spend
Vercel's October Production Index reported that open-weight models processed 71.5% of September gateway tokens and took nearly one in four dollars, and that nearly 70% of spend on Chinese models went through US-based providers. Vercel broadened its open-weight definition in this report.
A second platform shows the same token-versus-spend split and adds where the money on Chinese models goes.
1 report · one development
Reports sharing an origin are coverage, not independent confirmation.
Sep 22, 2026
OpenRouter: open-weight models at 72% of tokens, closed models at 73% of spend
OpenRouter's first data newsletter reported that open-weight models carried 72% of tokens in the week of 14 Sep, with Chinese labs at 69%, while closed-weight models took 73% of spend, down from 89% in early January.
Sets the token-versus-spend split at the centre of the question, on one platform.
1 report · one development
Reports sharing an origin are coverage, not independent confirmation.
Aug 20, 2026
Alibaba opens Qwen3.8-Max weights; flagship licences add commercial conditions
Alibaba's results filing said it opened the weights of Qwen3.8-Max (2.4 trillion parameters). The licence published with them on 12 Aug requires a separate licence for model-as-a-service and AI Work Assistant businesses above $50m of revenue. Z.ai's GLM-5.3 weights followed on 28 Aug under a licence requiring a Z.ai security review above $10bn.
Shows flagship open weights shipping under commercial conditions rather than permissive licences, the divergence the explanation tracks.
3 reports · one development
Reports sharing an origin are coverage, not independent confirmation.
Evidence
Sources: Vercel (AI Gateway Production Index, October 2026) ×6 · OpenRouter (Inferred 01 newsletter) ×4 · DeepSeek (API Docs, Models & Pricing) ×2 · Epoch AI (Data Insight) ×2 · NIST Center for AI Standards and Innovation (CAISI) ×2 · OpenRouter (LLM Rankings, Open / Closed chart) · and 18 more
OpenRouter (Inferred 01 newsletter) · Sep 22, 2026
69% of tokens on OpenRouter last week were from Chinese labs (vs 72% being open-weight overall).
Scope and limits: A single week (14 Sep 2026) of traffic routed through OpenRouter, excluding reseller activity. It is platform data, not market share, and does not cover direct vendor APIs, hyperscaler clouds or self-hosted use.
Retrieved Oct 21, 2026.
OpenRouter (Inferred 01 newsletter) · Sep 22, 2026
Open-weight model token share crossed 50% of tokens briefly in February and more permanently in April of this year It reached 79% in early July, its high point, and has held above 62% since.
Scope and limits: OpenRouter's own description of its weekly series to mid-September 2026. Its rankings chart, which excludes stealth models, peaks at 75.2% (week of 13 Jul) rather than 79%, so the two OpenRouter series differ by definition.
Retrieved Oct 21, 2026.
OpenRouter (Inferred 01 newsletter) · Sep 22, 2026
Closed weight models (from about 30 different labs, though dominated by the big 2 labs plus Google) saw 73% of spend in the second week of September as opposed to 89% the week of Jan 5.
Scope and limits: Spend is measured on OpenRouter only and excludes reseller activity; it is not industry revenue. Two weeks (5 Jan and the second week of September 2026) are compared, not a full trend.
Retrieved Oct 21, 2026.
OpenRouter (Inferred 01 newsletter) · Sep 22, 2026
3 key open-weight labs have seen their monthly spend increase tenfold or more from January to August 2026 (Moonshot AI, Z.ai, and DeepSeek). When viewed as a percentage of all spend the three labs are still only in the high single digits, but the growth rates are impressive.
Scope and limits: Growth rates on OpenRouter only; the newsletter gives no dollar amounts for the three labs, and "high single digits" is its own approximation of their combined share of spend.
Retrieved Oct 21, 2026.
OpenRouter (LLM Rankings, Open / Closed chart) · Oct 20, 2026
Share of tokens processed by open-weight vs. closed-weight models on OpenRouter, by week. Stealth models are excluded. […] These rankings show traffic routed through OpenRouter.
Scope and limits: The excerpt is the chart caption and scope line. Figures cited are read from the chart data on 21 Oct 2026 (usage data through 20 Oct): open-weight share 72.0% in the week of 12 Oct (118.1 trillion of 164.0 trillion tokens), 72.4% in the week of 14 Sep, between 66.5% and 75.2% in every full week from 6 Jul to 12 Oct; the week of 19 Oct was incomplete. OpenRouter states its rankings data is licensed CC BY 4.0 and that the rankings measure neither users nor spend.
Retrieved Oct 21, 2026.
OpenRouter (Inferred 03 newsletter) · Oct 8, 2026
In September, cached prompt tokens were just under 90% of all agentic tokens. […] Cached tokens are nearly nine in ten agentic tokens, but only about a third of the bill.
Scope and limits: The "third of the bill" figure is a worked example at one closed model's list prices (Claude Sonnet 5 on OpenRouter, 7 Oct 2026), not a platform-wide spend measure. Agentic versus human classification is OpenRouter's own.
Retrieved Oct 21, 2026.
Vercel (AI Gateway Production Index, October 2026) · Oct 15, 2026
Open weight takes 71.5% of tokens and 23% of spend […] Open-weight models now process more than seven in ten gateway tokens and account for nearly one in four dollars. Their share of spend has more than quadrupled since June.
Scope and limits: September 2026 traffic routed through Vercel AI Gateway only. Spend is estimated at list prices, and the open-weight definition was broadened this month, so the share is not comparable with earlier reports without restatement.
Retrieved Oct 21, 2026.
Vercel (AI Gateway Production Index, October 2026) · Oct 15, 2026
Since December, spend on frontier models has grown tenfold; spend on open-weight models has grown more than seventyfold.
Scope and limits: Growth multiples on one gateway from December 2025 to September 2026; they say nothing about absolute dollar amounts or other channels.
Retrieved Oct 21, 2026.
Vercel (AI Gateway Production Index, October 2026) · Oct 15, 2026
DeepSeek’s models processed 58% of all gateway tokens for 6.6% of spend. More than four in five of those tokens were served by DeepSeek itself, yet two in five dollars spent on its models went through US-based providers.
Scope and limits: One gateway, one month (September 2026). "US-based" means provider company headquarters, not where the servers run.
Retrieved Oct 21, 2026.
Vercel (AI Gateway Production Index, October 2026) · Oct 15, 2026
Nearly 70% of September’s gateway spend on Chinese models went through US-based providers, up from less than 10% in December.
Scope and limits: Routing of spend on one gateway (December 2025 and September 2026). "US-based" means provider company headquarters. It does not show the providers' margins, licence terms or any payment to the model labs.
Retrieved Oct 21, 2026.
Vercel (AI Gateway Production Index, October 2026) · Oct 15, 2026
Moonshot AI, the Beijing-based lab behind Kimi, accounted for 3.6% of gateway tokens and 13.2% of spend in September. US-based providers served nine in ten of those tokens.
Scope and limits: One lab on one gateway in September 2026; Kimi K3 is not split from other Kimi models.
Retrieved Oct 21, 2026.
Vercel (AI Gateway Production Index, October 2026) · Oct 15, 2026
Spending is estimated at published list prices; actual bills may differ. […] Open-weight classifications follow the current AI Gateway model list, a broader definition than in earlier reports.
Scope and limits: Methodology notes only; Vercel also says prior months may be revised as methodology is updated.
Retrieved Oct 21, 2026.
DeepSeek (Hugging Face model card, DeepSeek-V4.1-Flash-1017) · Oct 17, 2026
DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks. […] This repository and the model weights are licensed under the MIT License.
Scope and limits: Licence and release statement only. The model card's benchmark claims are DeepSeek's own and are not used here; Hugging Face download counters include automated pulls.
Retrieved Oct 21, 2026.
Moonshot AI (Kimi K3 License, Hugging Face) · Jul 27, 2026
If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars […] the Licensee must enter into a separate agreement with Moonshot AI
Scope and limits: Clause 2 of the licence as published with the weights (file committed 27 Jul 2026); the omitted words are "(or the equivalent in other currencies) in total over any consecutive 12 months,". Clause 4 exempts internal use and access through Moonshot AI's official products or certified inference partners. Whether any host has signed a separate agreement is not public.
Retrieved Oct 21, 2026.
Z.ai (GLM-5.3 License, Hugging Face) · Aug 28, 2026
If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 10 billion US dollars […] the Licensee must pass Z.AI's security review
Scope and limits: Clause 2 of the licence on the flagship GLM-5.3 weights (file committed 28 Aug 2026); the omitted words are "(or the equivalent in other currencies) in total over any consecutive 12 months,". The smaller GLM-5.3-Flash is published under MIT. The licence does not define the review beyond saying Z.ai determines its scope and method.
Retrieved Oct 21, 2026.
Qwen (Qwen3.8-Max License, Hugging Face) · Aug 12, 2026
If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, and the aggregate revenue of the licensee and its affiliates exceeds US$50,000,000 […] the licensee shall obtain a separate license from Qwen
Scope and limits: Condition 2 of the licence on the Qwen3.8-Max weights (file committed 12 Aug 2026); the omitted words are "(or the equivalent amount in any other currencies) during any consecutive twelve (12) months,". Internal use that exposes nothing to third parties is exempt.
Retrieved Oct 21, 2026.
Qwen (Qwen Community License 1.0, Qwen3.8-Flash-Next, Hugging Face) · Aug 26, 2026
If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose.
Scope and limits: Condition 2 of the licence on Qwen3.8-Flash-Next (file committed 26 Aug 2026). Which other Qwen releases use this licence is not established here.
Retrieved Oct 21, 2026.
Meta (Llama 4 Community License Agreement) · Apr 5, 2025
If, on the Llama 4 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta
Scope and limits: The Llama 4 licence, effective 5 Apr 2025. It does not show whether Meta will publish weights for later models, including Muse.
Retrieved Oct 21, 2026.
Meta Platforms, Inc. (SEC Form 10-Q, Q2 2026) · Jul 29, 2026
We are also developing AI technology that we make available via open source, commercial, and non-commercial license agreements to third-parties that can use this technology for use in their own products and services.
Scope and limits: A risk-factor statement of intent; it names no model, licence or release date and gives no revenue from licensing.
Retrieved Oct 21, 2026.
Meta Platforms, Inc. (SEC Form 8-K, Exhibit 99.1, Q3 2026 results) · Oct 21, 2026
"We hit major milestones this quarter toward delivering personal superintelligence for everyone - including launching Muse and Muse Wonder, our first frontier model," said Mark Zuckerberg, Meta founder and CEO.
Scope and limits: The CEO's statement in the results release. The release does not say whether Muse or Muse Wonder weights will be published; the Q3 10-Q had not been filed on 21 Oct 2026.
Retrieved Oct 21, 2026.
Amazon.com, Inc. (SEC Form 8-K, Exhibit 99.1, Q2 2026 results) · Jul 30, 2026
Added 10+ fully managed foundation models to Amazon Bedrock, including OpenAI’s GPT-5.6, Anthropic’s Claude Opus 5, Google DeepMind’s Gemma 4, and SpaceXAI’s Grok 4.3. Amazon Bedrock provides the best selection of leading models, at superior performance, and with the governance and security controls that companies need, and it’s continuing to grow quickly—hundreds of thousands of customers now use Bedrock, more customers were added in the last six months than in the first two years after launch, and customers spent more in Q2 than all prior quarters combined.
Scope and limits: Names four of the "10+" models added in Q2 2026; it gives no split of Bedrock usage or spend between open-weight and closed models.
Retrieved Oct 21, 2026.
Alphabet Inc. (SEC Form 8-K, Exhibit 99.1, Q2 2026 results) · Jul 22, 2026
Gemini models now process 22 billion API tokens per minute and the Gemini App has 950 million monthly active users.
Scope and limits: A rate stated by the CEO in July 2026, for Gemini (closed-weight) models only; Alphabet gives no open-weight figure. The weekly conversion (about 222 trillion tokens) is derived and assumes the rate holds all week.
Retrieved Oct 21, 2026.
Alibaba Group Holding Ltd. (SEC Form 6-K, Exhibit 99.1, quarter ended 30 Jun 2026) · Aug 20, 2026
In August, we launched our flagship foundation model Qwen3.8-Max within three months of its prior version, and we opened its model weights with 2.4 trillion parameters. […] AI-related product revenue continued to show strong momentum, achieving RMB12,376 million (US$1,824 million) and delivering the twelfth consecutive quarter of triple-digit year-over-year growth.
Scope and limits: AI-related product revenue is for Alibaba's AI Cloud and Compute Services segment and is not attributed to Qwen or to open-weight models. The licence on the opened weights is the separate Qwen3.8-Max License.
Retrieved Oct 21, 2026.
NVIDIA Corporation (SEC Form 8-K, Exhibit 99.1, Q2 fiscal 2027 results) · Aug 26, 2026
NVIDIA is not assuming any Data Center compute revenue from China in its outlook. […] Launched a local AI initiative with optimizations for top open models — DeepSeek v4 Flash, Diffusion Gemma, Nemotron 3.5 Lightning and Qwen 3.8 — and agent harnesses Hermes Agent and OpenClaw across RTX and DGX platforms.
Scope and limits: Outlook for the quarter ending late October 2026 and a product announcement; neither quantifies revenue from open-weight models.
Retrieved Oct 21, 2026.
Amazon Web Services (Amazon Bedrock User Guide, Models at a glance) · Oct 21, 2026
DeepSeek V3.2, DeepSeek-V3.1, DeepSeek-R1 […] Gemma 4: Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 E2B […] Kimi K3, Kimi K2.5, Kimi K2 Thinking […] GLM 4.7, GLM 4.7 Flash, GLM 5, GLM 5.3
Scope and limits: Catalogue rows for DeepSeek, Google (Gemma), Moonshot AI and Z.AI as read on 21 Oct 2026; the page is undated and changes. It shows availability only, not usage, and not the licence terms AWS operates under.
Retrieved Oct 21, 2026.
Microsoft (Microsoft Foundry Blog, What's new in Microsoft Foundry, May 2026) · May 31, 2026
DeepSeek V4 variants are now available in Microsoft Foundry. […] DeepSeek V4 Pro and Kimi 2.6 arrive via Fireworks for high-performance open-model inference.
Scope and limits: A May 2026 product update; it does not say whether DeepSeek V4.1 or Kimi K3 has since been added, and gives no usage figures.
Retrieved Oct 21, 2026.
DeepSeek (API Docs, Models & Pricing) · Oct 21, 2026
Note: when you use the DeepSeek API through DeepSeek Harness, Claude Code, or Codex, server-side tool calls already hit the Search and Fetch APIs. These calls are currently free of charge; per-request billing starts at 00:00 Beijing time on November 1, 2026.
Scope and limits: A price-list notice read on 21 Oct 2026; Search and Fetch are listed at $0.75 per 1,000 successful requests each. It does not say how much traffic these tool calls carry.
Retrieved Oct 21, 2026.
DeepSeek (API Docs, Models & Pricing) · Oct 21, 2026
(2) Off-peak rates are half of the peak rates. Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays.
Scope and limits: The excerpt is the peak/off-peak footnote. Prices cited are read from the table on 21 Oct 2026 for deepseek-flash (DeepSeek-V4.1-Flash-1017), per 1M tokens: input cache miss $0.30 peak / $0.15 off-peak; output $1.20 / $0.60; input cache hit $0.006 / $0.003. List prices only; third-party hosts set their own.
Retrieved Oct 21, 2026.
Anthropic (Claude API Docs, Pricing) · Oct 21, 2026
Claude Haiku 5.5 is priced by prompt length: a request whose prompt is over 100,000 tokens pays higher prices.
Scope and limits: The excerpt is the long-context rule. Prices cited are read from the model pricing table on 21 Oct 2026, per 1M tokens: Claude Haiku 5.5 $0.10 input / $0.50 output for prompts up to 100,000 tokens ($0.50 / $2.50 above), cache hits $0.01; Claude Opus 5.5 $4 / $20. The page also says Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text, so per-token prices across vendors are not like-for-like.
Retrieved Oct 21, 2026.
Epoch AI (Data Insight) · May 29, 2026
Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months in the Epoch Capabilities Index (ECI), our aggregate measure of model capability. The average ECI gap was 8 points, similar to the gap between GPT-5 and GPT-5.5.
Scope and limits: Average over 1 Jan to 28 May 2026 on Epoch's composite index; Epoch says the estimate grows to six months under a stricter test. Not updated for models released after May. Epoch states its work may be reused with credit under a Creative Commons BY licence.
Retrieved Oct 21, 2026.
Epoch AI (Data Insight) · Oct 16, 2026
Z.ai’s GLM-5.3 is the highest-scoring open-weight model on the Cyber ECI, a cybersecurity-focused version of our Epoch Capabilities Index (ECI). […] This places the open cyber frontier roughly where the closed frontier stood 6–7 months ago.
Scope and limits: A cybersecurity-only index using data as of 8 Oct 2026; three of GLM-5.3's results come from CAISI's September assessment. It is a capability measure, not a security assessment.
Retrieved Oct 21, 2026.
Stanford HAI (AI Index Report 2026, Chapter 2: Technical Performance) · Jul 15, 2026
As of March 2026, the top closed-weight model, Claude Opus 4.6 (1,503), led the top open-weight model GLM‑5 (1,454) by 49 points (3.4%).
Scope and limits: One leaderboard (Arena) at one date (March 2026). The date shown is the chapter PDF's embedded creation date; the report's publication date is not stated in the file.
Retrieved Oct 21, 2026.
NIST Center for AI Standards and Innovation (CAISI) · May 1, 2026
In April 2026, the Center for AI Standards and Innovation (CAISI) evaluated the open-weight AI model DeepSeek V4 Pro (“DeepSeek V4”). CAISI evaluations indicate that DeepSeek V4’s capabilities lag behind the frontier by about 8 months (Figure 1).
Scope and limits: CAISI's own measure, which includes non-public benchmarks; CAISI says DeepSeek V4 scored better on DeepSeek's self-reported evaluations, which put it about 2 months behind. One model, evaluated in April 2026; DeepSeek V4.1 models are not covered.
Retrieved Oct 21, 2026.
NIST Center for AI Standards and Innovation (CAISI) · Sep 17, 2026
Z.ai publicly released the model’s weights two weeks later. […] GLM-5.3 lags the capability level of the U.S. frontier by about four months in an aggregate measure of performance across CAISI cyber benchmarks (Figure 1).
Scope and limits: A cyber-only capability measure on four benchmarks; it is not a general capability measure. CAISI dates GLM-5.3's release to 14 Aug 2026, with weights two weeks later. Findings are CAISI's own.
Retrieved Oct 21, 2026.
Bureau of Industry and Security, U.S. Department of Commerce (Federal Register final rule 2026-00789) · Jan 15, 2026
The Bureau of Industry and Security (BIS) is revising its license review policy for exports of certain semiconductors to China and Macau—changing it from a presumption of denial to a case-by-case review. The semiconductors covered by this rule are the Nvidia H200 and its equivalents
Scope and limits: The rule's summary. Case-by-case review is subject to exporter certifications and third-party testing set out in the rule; it does not show any licence granted or any shipment.
Retrieved Oct 21, 2026.
Coverage scope: Global usage of text and multimodal language models through APIs and self-hosting, with labs in the US (OpenAI gpt-oss, Meta, Google Gemma, NVIDIA Nemotron), China (DeepSeek, Alibaba Qwen, Moonshot Kimi, Z.ai GLM) and the EU (Mistral). Sources: usage publications from OpenRouter and Vercel AI Gateway (platform data, not market share); licence files and model cards on Hugging Face and Meta's site; price lists (DeepSeek, Anthropic); cloud catalogues (AWS Bedrock, Microsoft Foundry); SEC filings (Meta, Amazon, Alphabet, Alibaba, NVIDIA); NIST CAISI evaluations and a BIS rule; Epoch AI and the Stanford AI Index. Out of scope: data-centre capex, capacity and financing (see Financing the AI buildout and the AI Infrastructure collection), consumer chatbot share, image and video models, and security assessments beyond attributed government documents.
Update history
- Update 1 · Oct 27, 2026 · this page
Free account. See what changed since your last visit, and the dates to watch on your calendar. Following never means agreeing.
Analytics, not investment advice. How Finquill tracks narratives · Report an issue