The AI models that matter in October 2026, and what changed

Between 1 September and 26 October 2026, Anthropic, OpenAI, Google, xAI, Meta and DeepSeek all shipped new flagship or near-flagship models. Mistral and Reflection announced trillion-scale and 500-billion-scale models with weights to follow. Alibaba said Qwen 4 is in training.

That is a lot of names in eight weeks. Most coverage lists them. This post tries to do something more useful: put them on one page, say what actually changed, and explain what it means if you run Moodle for a school, a university or a training team.

All benchmark numbers below are the vendors’ own unless stated otherwise. Treat them as claims, not measurements.

How we got here, in one paragraph

In May 2020 OpenAI described GPT-3, a 175 billion parameter model, in a paper called Language Models are Few-Shot Learners.¹1 arXiv, Language Models are Few-Shot Learners (submitted 28 May 2020). It was a research result with an API waiting list. ChatGPT, built on a GPT-3.5 model, launched on 30 November 2022²2 Wikipedia, ChatGPT. and was estimated to reach 100 million monthly users two months later, according to a UBS note reported by Reuters.³3 Euronews (Reuters), ChatGPT sets record for fastest-growing user base. Four years on, the products are less about chatting and more about models that run tools, write and test code, and operate a computer for minutes or hours at a time.

The current 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.

Gaps worth knowing about. Google’s pricing page²¹21 Google, Gemini Developer API pricing. lists no Gemini 3.5 Pro; TechCrunch reported in July that the model had been delayed, and Google moved on to 4 Argon. Alibaba’s 22 September press release says only that Qwen 4 “is currently in training”.²²22 Alibaba Cloud, Alibaba unveils roadmap on full-stack AI strategy. OpenAI’s open-weight models are still the gpt-oss pair from August 2025.²³23 OpenAI, Introducing gpt-oss.

What actually changed

Models now do the work, not just answer

The headline benchmarks have moved from quiz questions to tasks: run commands in a terminal, click through a desktop, finish a multi-step job. The jumps in one generation are large. Anthropic reports Sonnet 5.5 at 70.6% on Terminal-Bench 4.0, against 10.3% for Sonnet 5. Haiku 5.5 scores 39.2%, where Haiku 4.5 scored 0.0%.⁶⁷ OpenAI pitches GPT-6.1 Terra as “the everyday model” for reading screens, filling in forms and clicking through apps, and reports 50.3% on the same test, up from 21.5%.⁹

The same SiliconANGLE report carries the warning that comes with this. In OpenAI’s own testing, Terra went against its restrictions to finish a task in nearly 30% of simulated runs.⁹ An agent that can act can also act wrongly.

The most capable versions are gated

Four vendors made the same move within weeks. Claude Mythos 5.1 is the same model as Fable 5.1 with looser safeguards, limited to trusted access programs.⁴ OpenAI said GPT-6 Astra was its first model to reach its “Critical” internal cybersecurity threshold and gave early access to its Daybreak program first, CNBC reported.⁸ Google is rolling Argon and Gemini 3.8 Flash Cyber out to cyber defenders through its Fairwind Program first.¹² xAI limits Grok 4.7 red-team access to invited security partners.¹⁴

For most buyers this changes little. It does mean the model in the press release and the model your API key reaches are not always the same thing.

Prices fell, but the bill got conditional

The direction is down. Anthropic says Opus 5.5 costs 40% less to run than Opus 5.⁵ Terra costs about half of what its predecessor did.⁹ Fable 5.1 kept its base price but cut cache reads, which Anthropic says saves about 25% on typical work.⁴ DeepSeek cut prices again with V4.1-Flash.²⁴24 DeepSeek, API change log.

The catch is that the list price is now a starting point:

• Gemini 4 Argon is $2 / $10 until 30 November, then doubles. Gemini 3.8 Flash doubles on 1 January 2027.²¹
• OpenAI charges more once input passes 272K tokens, and adds 10% for regional data processing.¹⁰
• Haiku 5.5 is five times more expensive for prompts over 100K tokens.²⁵25 Anthropic, Claude pricing.
• DeepSeek’s rates double during peak hours, which are defined in UTC.¹⁷

If you budget from the headline number, you will be wrong in at least one direction.

Long context is normal now

A million tokens of context was a selling point two years ago. It is now listed almost by default. Anthropic’s pricing page says its models support up to 1M tokens, varying by model.²⁵ Muse Spark Open, Kimi K3, Beam and DeepSeek’s API all list 1M.^15162017 Gemini 4 Argon raised its output limit from 64K to 1M tokens.¹¹ What changed is the pricing of that context, as above, not its availability.

Open weights are close behind

Kimi K3’s own benchmark table puts it within a point of GPT-5.6 Sol and Claude Fable 5 on several tests.¹⁶ DeepSeek says its MIT-licensed V4.1-Flash-1017 solves 77% of DeepSWE v1.1 tasks;¹⁷ Google reports 77.9% for Gemini 4 Argon on the same benchmark.¹¹ Both numbers are self-reported, and SiliconANGLE notes Claude Opus 5 still led DeepSeek on seven of eleven published benchmarks.¹⁷

The American side is also back in open weights. Meta calls Muse Spark Open the most capable US open-weight model, though running it locally needs around 1 TB of RAM, per TechCrunch.²⁶26 TechCrunch, Meta launches an open-weight AI model that runs on Macs. Reflection’s Beam and Mistral Large 4 both promise weights; neither had shipped them in the coverage we found.²⁰¹⁹

What this means for Moodle

Moodle’s AI subsystem, release by release

Moodle added an AI subsystem in 4.5. The pieces are providers (who runs the model) and placements (where users see AI in Moodle). From the release notes:

Release dates are from Moodle’s release table.³²32 Moodle HQ, Releases and support dates. The placements themselves are still small: generate text and generate image in the editor, and summarise or explain the text of a course page. All are off by default.³³33 Moodle Docs, AI placements.

Two lines in the providers documentation matter more than the rest.³⁴34 Moodle Docs, AI providers. “Other providers” covers any service with an OpenAI-compatible API. And the Ollama provider talks to models you install on your own server. Between them, almost every model in the table above can be connected to Moodle 5.3 in some way.

Choosing a provider is a data decision first

When a student clicks “Summarise”, the text of that course page goes to whichever provider is configured. That raises three questions for a school.

Where does the data go, and is it kept? Anthropic offers Opus 5.5 and Sonnet 5.5 with zero data retention.⁵⁶ OpenAI sells regional processing at a 10% uplift.¹⁰ AWS Bedrock and Azure keep traffic inside a cloud account the institution may already have under contract.³⁴ Moodle’s own providers page has no data privacy notes, so this is your homework, not Moodle’s.

What does it cost per student? A rough illustration, 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.

Arithmetic on the published prices above.²⁵¹⁷²¹ Real usage varies a lot, and the conditional pricing rules change these numbers. The point is the spread: a factor of 100 between the cheap and top tiers, for a task the cheap tier handles fine. Moodle 5.0’s provider instances let you route small requests to a cheap model and large ones to a bigger one.³⁴

Can you run it yourself? Open weights make self-hosting legal and possible. They do not make it cheap. Muse Spark Open needs around 1 TB of RAM.²⁶ A 27B model like Qwen3.8-27B, with an Apache 2.0 licence, is a far smaller thing to host.¹⁸ For a school with strict data rules, a small open model behind Ollama keeps student text on its own hardware.

Academic integrity, briefly

One detail from the Fable 5.1 launch is worth a teacher’s attention. MacRumors reports that Anthropic now adds invisible watermarks to text Claude generates, with a detection API in private preview, “as required by EU law”.⁴ That covers one vendor’s text, not all AI text. It is not a plagiarism checker, and nobody should treat it as one yet.

Where custom plugins come in

Moodle’s core providers and placements cover the common cases. Connecting a specific model, a private endpoint or a self-hosted open-weight model to a new placement, such as feedback on a quiz attempt or a course-specific tutor, is provider or placement plugin work, and that is the kind of plugin we build.