AI model releases, November 2026 so far: what changed

Every major AI model released from 1 to 19 November 2026, dated and priced from vendor pages, plus what each one means for a school on Moodle 4.5 or 5.3.

Drawing of a person in a blue sweater unpacking small robots from cardboard boxes and lining them up on a long shelf beside older machines.

In the first 19 days of November 2026, Anthropic, Google, OpenAI, Meta, Alibaba and Tencent all shipped new models. Google alone put three Gemini 4 models on general release in eight days. Moonshot published the weights of Kimi K3.1 Pro.

This is the follow-up to our October roundup. Same approach. One table of releases, then what changed and what it means if you run Moodle for a school or a university.

Everything below is as of 19 November 2026 and covers text models only, so image, video and speech releases are left out. Dates and prices come from the vendor’s own blog, changelog, docs or model card, except Tencent’s, which come from Tencent’s own press releases on PR Newswire. All are linked in the footnotes. Benchmark numbers are the vendors’ own. Treat them as claims, not measurements.

The November lineup

Prices are US dollars per million input and output tokens on each vendor’s standard API tier. “Open” means the weights can be downloaded. It does not mean open source, and each licence has its own terms.

VendorModelReleasedWeightsPrice in / outBuilt for
AnthropicClaude Fable 5.29 Nov 2026Closed$8 / $40Top tier reasoning and long agent jobs12
MetaMuse Ember 1.09 Nov 2026Closed (no weights mentioned)$0.10 / $0.50 (prompts up to 200K)High volume agent steps: tool calls, extraction3
GoogleGemini 4 CarbonAnnounced 28 Oct, general release 10 Nov 2026Closed$2 / $10 to 30 Nov, then $4 / $20Hardest, long-horizon work45
GoogleGemini 4 Hydrogen12 Nov 2026Closed$1.50 / $7.50, also on the free tierFast everyday coding and agent work67
GoogleGemini 4 Neon18 Nov 2026Closed$3 / $15Middle of the Gemini 4 family87
AlibabaQwen4-Max16 Nov 2026Closed (API)$1.50 / $4.50 (Singapore region)Most capable Qwen, 1M context910
TencentHy44 Nov 2026Weights promisedAbout $0.90 / $2.70Tencent’s main model11
TencentHy4 flash17 Nov 2026Open (Tencent Hy License)About $0.070 / $0.281Cheap everyday work, 1M context1213
OpenAIGPT-6.2 Astra17 Nov 2026Closed$8 / $40Demanding reasoning and coding1415
OpenAIGPT-6.2 Sol17 Nov 2026Closed$2 / $10Everyday coding and agent workflows1416
MoonshotKimi K3.1 ProWeights early Nov 2026Open (Kimi K3.1 Pro License)n/a2.8T parameters, 104B active, 1M context17

Notes on the table:

  • Gemini 4 Carbon was first offered only to “a subset of cyber defenders” on 28 October.4 Developers could buy it from 10 November.5
  • Hy4 flash is dated 17 November in Tencent’s release, datelined Shenzhen.12 Several trackers list 16 November. The release says Tencent “plans to open-source” the model, but its Hugging Face model card already lists the weights as released.13
  • Hy4 prices and dates come from Tencent’s release on PR Newswire, which Tencent issued. That release said weights and system cards would follow “in the coming days”.11 On 19 November Tencent’s Hugging Face page had no Hy4 repository, only Hy4 preview and Hy4 flash.
  • Kimi K3.1 Pro weights sit on Moonshot’s Hugging Face page. The repository’s first commit is dated 2 November and its last 5 November. That is when the repository was created, not proof of when it went public.17
  • OpenAI’s launch post for GPT-6.2 blocked our reader. The dates, prices and limits here come from OpenAI’s changelog, pricing page and model pages instead.

One follow-up from October. Mistral has now released the weights of Mistral Large 4 (“Le Chonk”), a 1-trillion-parameter model, under its own open-weight licence, at $0.70 / $2.10 on its API. Its news page dates the post 29 October.18

What actually changed

The top tier got 20% cheaper

Claude Fable 5.2 costs $8 / $40. Fable 5.1 was $10 / $50.19 GPT-6.2 Astra also costs $8 / $40, against $10 / $50 for GPT-6 Astra.15 Both cuts are 20%, and they landed eight days apart.

Anthropic says Fable 5.2 answers up to 30% faster than Fable 5.1 and reports 49.1% on FrontierCode 2.0, against 45.5% for Opus 5.5.1 Neither vendor is telling most people to use the top model. Anthropic’s docs say to start with Opus 5.5 for most work, and to use Fable 5.2 for demanding reasoning or when Opus 5.5 at higher effort still falls short in your tests.2 OpenAI’s model page calls Sol “our go-to model for everyday work”.16

Google filled in the Gemini 4 family

Our October post had one Gemini 4 model, Argon. Now there are four. Google’s own descriptions put them in order: Carbon “tackles your hardest, long-horizon projects”, Neon sits in the middle, and Hydrogen “handles fast everyday execution”.8 Neon can direct a set of Hydrogen sub-agents in Google Antigravity.8

Hydrogen is the one to watch for cost. It costs $1.50 / $7.50 at any prompt length, has a 1M token context window, and is on the free tier of the Gemini API.67

Cheap models got serious

Two new models sit at the very bottom of the price list.

  • Muse Ember costs $0.10 / $0.50, the same as Claude Haiku 5.5. Meta reports 40.9% on Terminal-Bench 4.0, against 39.2% for Haiku 5.5.3
  • Hy4 flash costs about $0.070 / $0.281. Tencent’s release backs it with its own internal benchmark and a blind test with 85 internal users of its WorkBuddy app.12

Both claims are self-reported. The price is not, and it is now low enough that cost is no longer the reason to switch AI features off.

The gated versions kept coming

The pattern from October continued. Claude Mythos 5.2 is the same model as Fable 5.2 without the cyber and biology blocks, open only to vetted organisations.1 OpenAI added GPT-6 Cyber on 5 November for authorised vulnerability work.14 Carbon and Neon both have cyber variants reached through Google’s Fairwind Program.58 None of this matters for a course site, beyond knowing that the press release model and the model your key reaches are not always the same.

Prices still come with conditions

  • Gemini 4 Carbon doubles to $4 / $20 after 30 November.5
  • Muse Ember triples to $0.30 / $1.50 for prompts over 200K tokens.3
  • OpenAI charges more once input passes 272K tokens, and adds 10% for regional data processing.15
  • Hy4 flash is half price on Tencent Cloud TokenHub until 30 November.12
  • Qwen4-Max costs $1.50 / $4.50 in Singapore and $1.413 / $4.24 in Beijing, Frankfurt, Virginia and the other listed regions.10

One change goes the other way. Since 17 November OpenAI lets you wait two or eight hours for an answer, at 30% or 40% off.14 Overnight jobs, such as summarising a term’s forum posts, do not need an instant reply.

What this means for Moodle

First, update Moodle

On 12 November Moodle HQ shipped unscheduled releases for every supported branch, including 5.3.1 and 4.5.16, with security fixes whose details come about a week later.202122 None of the listed fixes touch the AI subsystem. Install them before you change any AI setting.

Also worth knowing if you are on 4.5: general support ended on 6 October 2025. Security fixes continue to 4 October 2027.20

Which new model fits which Moodle

Moodle 4.5 shipped with two AI providers: OpenAI and Azure AI.23 Moodle 5.3 lists Anthropic Claude (new in 5.3), Amazon Bedrock, Azure AI, DeepSeek, Gemini API, Ollama, OpenAI, and “Other providers” for services with an OpenAI-compatible API.2425

The mapping below is our view, based on those provider lists. Moodle’s docs do not say which model names each provider accepts, so test any new model on a staging site first.

ModelMoodle 4.5Moodle 5.3
GPT-6.2 Sol, GPT-6.2 AstraOpenAI providerOpenAI provider
Claude Fable 5.2Not in coreAnthropic Claude or Amazon Bedrock provider
Gemini 4 Carbon, Neon, HydrogenNot in coreGemini API provider
Qwen4-Max, Muse EmberNot in corePossibly “Other providers”, if the vendor’s endpoint is OpenAI-compatible. Check first
Hy4 flash, Kimi K3.1 Pro (self-hosted)Not in core”Other providers” through your own inference server

For Hy4 flash, the model card shows how to run your own vLLM server and call it through an OpenAI-compatible API.13 We did not find the same statement for Muse Ember. Meta’s post mentions only the Responses API.3

Cost per student

The same rough illustration as October, with our assumptions stated: a student makes 20 summarise or explain requests a month, each about 3,000 tokens in and 500 out. That is 60,000 input and 10,000 output tokens a month.

ModelCost per student per month2,000 students
Hy4 flash (standard price)about $0.007about $14
Muse Ember 1.0about $0.011about $22
Qwen4-Max (Singapore)about $0.14about $270
Gemini 4 Hydrogenabout $0.17about $330
GPT-6.2 Solabout $0.22about $440
Gemini 4 Carbon (to 30 Nov / from 1 Dec)about $0.22 / $0.44about $440 / $880
Gemini 4 Neonabout $0.33about $660
Claude Fable 5.2 or GPT-6.2 Astraabout $0.88about $1,760

Arithmetic on the published prices in the lineup table. Real usage varies a lot. The spread is still more than 100 times between the cheapest and the top tier, for tasks the cheap tier handles fine. Fable 5.2 is now 20% cheaper than Fable 5.1, and it is still the wrong model for summarising a course page.

Data location

When a student clicks “Summarise”, the page text goes to whichever provider is configured. Moodle’s providers page says nothing about data privacy,24 so check the vendor’s terms yourself:

  • Anthropic: the Claude API routes globally by default. US-only processing costs 1.1 times the normal price. On Bedrock and Google Cloud, regional endpoints cost 10% more than global ones.19 Anthropic says “eligible customers” can use Fable 5.2 with zero data retention, and that other customers have 30-day retention for safety monitoring.1
  • OpenAI: regional processing endpoints cost 10% more.15
  • Alibaba: Qwen4-Max is listed in Frankfurt, Virginia and Tokyo, but those rows are marked “Global” scope on the pricing page.10 Do not read a Frankfurt price as an EU data promise until Alibaba’s deployment documentation says so.
  • Meta: the Muse Ember post says nothing about data retention or region. It does say the model is available on Google Cloud and Oracle Cloud.3

Open weights, honestly sized

Two models this month have weights you can download. Neither is small.

Hy4 flash has 220 billion parameters, 12 billion active per token.13 Tencent’s release says 224 billion, so the numbers do not quite match.12 Tencent’s own serving recipes run the FP8 version split across eight GPUs. A GGUF version exists, but Tencent says it is “for end-to-end testing purposes”.13 Kimi K3.1 Pro has 2.8 trillion parameters.17 For a school, that is a data centre project, not an Ollama install.

Read the licences too. Both are permissive with extra commercial terms. The Tencent Hy License requires a separate licence for model-as-a-service businesses above US$50 million in revenue or 100 million monthly users, and exempts purely internal use, where neither the model nor its outputs reach third parties.13 Whether serving your students counts as internal use is a question for your legal team, not for us.

Where custom plugins come in

Core Moodle covers OpenAI, Anthropic, Gemini and a few others. Connecting a model that core does not support, a private endpoint, or a self-hosted open-weight model to a placement your teachers actually want is provider or placement plugin work. That is the kind of plugin we build. More on that in Moodle AI features and custom plugins.

Sources

  1. Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2 (9 November 2026). ↑ ↑ ↑ ↑

  2. Anthropic, Models overview. ↑ ↑

  3. Meta, Introducing Muse Ember (9 November 2026). ↑ ↑ ↑ ↑ ↑

  4. Google, Gemini 4 Carbon: our next frontier model (28 October 2026). ↑ ↑

  5. Google, Gemini 4 Carbon is now available (10 November 2026). ↑ ↑ ↑ ↑

  6. Google, Introducing Gemini 4 Hydrogen (12 November 2026). ↑ ↑

  7. Google, Gemini API changelog. ↑ ↑ ↑

  8. Google, Introducing Gemini 4 Neon (18 November 2026). ↑ ↑ ↑ ↑

  9. Alibaba Cloud, Alibaba unveils Qwen4-Max (16 November 2026). ↑

  10. Alibaba Cloud, Model Studio model inference pricing (updated 18 November 2026). ↑ ↑ ↑

  11. Tencent via PR Newswire, Tencent launches Hy4 (4 November 2026). ↑ ↑

  12. Tencent via PR Newswire, Tencent launches Hy4 flash (17 November 2026). ↑ ↑ ↑ ↑ ↑

  13. Tencent, Hy4 flash model card and licence on Hugging Face. ↑ ↑ ↑ ↑ ↑ ↑

  14. OpenAI, API changelog. ↑ ↑ ↑ ↑

  15. OpenAI, API pricing. ↑ ↑ ↑ ↑

  16. OpenAI, GPT-6.2 Sol model page; GPT-6.2 Astra model page. ↑ ↑

  17. Moonshot AI, Kimi K3.1 Pro on Hugging Face and its commit history. ↑ ↑ ↑

  18. Mistral AI, Mistral Large 4 open weights; date from the Mistral news index. ↑

  19. Anthropic, Claude API pricing. ↑ ↑

  20. Moodle HQ, Releases and support dates. ↑ ↑

  21. Moodle HQ, Moodle 5.3.1 release notes. ↑

  22. Moodle HQ, Moodle 4.5.16 release notes. ↑

  23. Moodle HQ, Moodle 4.5 release notes. ↑

  24. Moodle Docs, AI providers. ↑ ↑

  25. Moodle HQ, Moodle 5.3 release notes. ↑

Need a Moodle plugin built or rescued?

Tell the scoping assistant what you need. You get a fixed price and a timeline in a few minutes, and an engineer reviews it within one business day.

Get a fixed quote