AI October is a disruptors dream
OpenAI's AI wrote 722 maths papers at once, GPT-6 went free and a leading AI model got 90 percent cheaper. This is the first fortnight of October, explained in plain language.
October 2026 started at a frantic pace. On 6 October, OpenAI posted 722 maths papers written by an unreleased AI model, and mathematicians are still arguing over what it means; one called the flood a "slopocalypse". In the same fortnight, OpenAI made GPT-6 free for everyone who uses ChatGPT, Anthropic cut the price of its smallest Claude model by 90 percent in under a year, and Google launched an AI that carries out office tasks by itself. This guide explains the eight biggest releases in plain language, and how to use them.
Eight big AI releases in 13 days
- 1 OctMicrosoftMAI-Transcribe-2: live speech to text
- 6 OctOpenAI722 maths papers written by AI
- 6 OctOpenAIDecisions API: cheap message sorting
- 6 OctMistralMistral Large 4: a model you can host yourself
- 6 OctGoogleNano Banana 2.1: images with readable text
- 7 OctAnthropicClaude Haiku 5.5: 90% cheaper
- 7 to 8 OctOpenAIGPT-6 free in ChatGPT
- 8 OctGoogleGemini agent: AI that does office work
Release 1 of 8OpenAI's AI wrote 722 maths papers in one go and left mathematicians reeling
- When
- 6 October
- Made by
- OpenAI, using an AI model it has not released
- Who can see it
- Anyone; the papers are free to read on GitHub, a public website for sharing files
- Cost
- Free to read
A mathematical proof is a step-by-step argument that shows a statement is true beyond doubt. Proofs are how mathematics moves forward, and one important proof can take a human expert years.
On 6 October, OpenAI published new results on long-standing maths problems, written by an internal AI model it has not released to the public. Axios counted 722 papers, grouped into 372 families of related findings; Nature, the science journal, put the figure at more than 700.
OpenAI says many of the proofs come with versions written for Lean, a computer tool that checks a proof line by line so that a mistake cannot slip through. It says the average result took about three hours of AI computing, and that it took advice from an independent group at the Institute for Advanced Study in Princeton before publishing.
Some results touch famous problems. Alex Kontorovich of Rutgers University said one result linked to the Riemann hypothesis, among the best-known unsolved problems in mathematics, would earn a Fields Medal, the field's top prize, if a person had produced it. Nature reported that none of the papers solves one of the remaining Millennium Prize Problems, the famous challenges that each carry a US$1 million prize.
What mathematicians said about OpenAI's 722 papers
“This is completely superhuman mathematics!”
Constantin Kogler, Institute for Advanced Study, on X
“If a human did this, it would be an instant Fields Medal, no questions asked.”
Alex Kontorovich, Rutgers University, on X
“All mathematicians should take a brief holiday to recover.”
Thomas Bloom, University of Manchester, on X
“So I will not be reading these papers.”
Michael Harris, Columbia University, who called the process illegitimate, to Nature
“The problem is most of those theorems are not ones that anybody will care about.”
Stephen Wolfram, computer scientist, at a National Museum of Mathematics event, reported by Axios
“Tossing the community a messy first draft”
Ursula Martin, University of Oxford, describing OpenAI's approach, to Nature
The alarm is about volume as much as accuracy. One mathematician quoted by Nature called the flood a "slopocalypse", a play on AI slop, the name for low-effort content made by machines. Experts now have to check hundreds of papers they did not ask for, and some researchers found that projects they had worked on were beaten to the finish by the AI, Nature reported.
Why it matters
Maths is the cleanest test of reasoning there is, because a proof is either right or wrong and a computer can check it. Axios noted that AI made the same leap in computer programming, another field where answers can be checked automatically. Our reading: the same capacity for long, careful reasoning is heading for engineering, medicine, finance and law next.
For Caribbean students sitting CSEC and CAPE mathematics, the lesson is that producing an answer is getting cheap, and judging whether an argument holds up is becoming the scarce skill. For the region's universities, it is an opening: a small mathematics department at the University of the West Indies or the University of Guyana can now test ideas with tools that once needed a large research team, provided someone checks every step.
The flood also previews a problem every Caribbean institution will meet: more AI-made work than people can review. A ministry that receives 700 AI-written reports faces the same question the mathematicians face this month, which is who checks them, and how.
Release 2 of 8GPT-6 is now free for everyone who uses ChatGPT
- When
- 7 October for paying users, 8 October for free users
- Made by
- OpenAI, the company behind ChatGPT
- Who can use it
- Anyone with ChatGPT, worldwide
- Cost
- Free; paid plans get a version called GPT-6 Sol
GPT-6 is the newest model behind ChatGPT. A model is the AI system that writes the answers; ChatGPT is the app you type into. On 7 October, OpenAI started giving GPT-6 to every ChatGPT plan, and free users received a version called GPT-6 Luna the next day. OpenAI says the rollout is worldwide, with no countries left out.
The update also changes what an answer looks like. OpenAI calls this Intelligent UI. Instead of only paragraphs of text, ChatGPT can now reply with charts, forms, buttons you can tap, and small working tools such as calculators and checklists. It still uses plain text when text is clearer.
Why it matters
An upgrade to the free plan reaches students, market vendors and small business owners who may not pay for a premium subscription. For them, the best version of ChatGPT they can use just improved at no cost.
The interactive answers help most for people who learn by seeing. Ask for a repayment calculator on a JMD 500,000 loan at 12 percent over three years, and ChatGPT can now build one you can adjust. Treat the numbers as a starting point and check them against your bank's figures, because a neat chart can still contain a wrong number.
Release 3 of 8Google's Gemini agent carries out office work by itself
- When
- 8 October
- Made by
- Who can use it
- Businesses first; ordinary users later
- Cost
- Not announced yet
Most AI chatbots answer questions. An agent goes further: you give it a goal, and it works out the steps and carries them out using your email, documents and spreadsheets. On 8 October, Google announced the Gemini agent, which it describes as "a universal agent for work." According to TechCrunch, it connects to Google's office tools, Microsoft's office tools, Slack and many business databases.
Each agent gets its own work account and email address, so everything it does is recorded under its own name instead of a staff member's. It can also switch between AI models from different companies, including Anthropic's Claude. Google is offering it to businesses first and has not published a price.
Why it matters
This is AI moving from giving advice to doing the work. A small importer in Kingston could one day ask the agent to check stock levels, draft reorder emails to suppliers and update the stock sheet, then review what it did.
That makes oversight the first question for any manager. An agent that can send email on a company's behalf needs the checks a new employee gets: what it may touch, who approves its work and how mistakes get caught. The record of its actions is what lets a manager, an auditor or a bank's compliance officer answer those questions.
Release 4 of 8Claude Haiku 5.5 does more than last year's version for 90 percent less
- When
- 7 October
- Made by
- Anthropic, the company behind Claude
- Who can use it
- Businesses and developers building their own AI tools
- Cost
- US$0.10 to read and US$0.50 to write a million tokens
Haiku is the smallest, fastest and cheapest member of Anthropic's Claude family, built for jobs that run thousands of times a day, such as answering customer questions or summarizing documents. Haiku 5.5 arrived on 7 October. Businesses reach it through an API, a connection that lets developers plug an AI model into their own app or website, and through the cloud services of Amazon, Google and Microsoft.
AI companies charge by the token, a small piece of text about three quarters of a word long on average, according to OpenAI, so a million tokens is roughly 750,000 words. When Haiku 4.5 launched in October 2025, it cost US$1 to read a million tokens. Haiku 5.5 costs US$0.10, a cut of 90 percent. It also does more: on a test of whether an AI can operate a computer, opening programs and clicking through tasks, it scored 72.4 percent against 15.7 percent for the older model, according to Anthropic's own figures.
Why it matters
Price often stops a credit union, telecom or hotel in the region from building its own AI assistant. Take a support chatbot that handles 10,000 conversations a month. At Haiku 5.5's price, the AI fees come to about US$4.50 a month (our calculation, assuming each conversation reads 2,000 tokens and writes 500).
Building it well still costs money, because the chatbot has to be connected to accurate customer records and needs a quick way to pass hard cases to a person. The difference is that the budget now goes into doing the job properly instead of paying for the AI itself.
Release 5 of 8OpenAI's Decisions API sorts a hundred thousand messages for about US$5
- When
- 6 October
- Made by
- OpenAI
- Who can use it
- Developers, as an early test version
- Cost
- US$0.10 per million tokens read; nothing for the answers
Plenty of AI work needs no written reply. Sometimes you only need a decision: is this email a complaint, which department should get it, how urgent is it. OpenAI's Decisions API is a service built for that job. A developer writes the questions and the allowed answers in advance, and the AI picks an answer and says how confident it is.
It opened to all developers on 6 October as a public beta, an early version that is still being tested and may change. It runs on GPT-6 Luna, the same model free ChatGPT users get. According to a report by Let's Data Science, OpenAI charges US$0.10 per million tokens it reads and nothing for its answers.
Why it matters
Ministries, insurers and credit unions across the region deal with inboxes that staff still sort by hand. Sorting 100,000 messages of about 375 words each would cost about US$5 at that price (our calculation). That frees staff to deal with the messages instead of routing them.
Microsoft and Cloudflare released similar sorting models in the same fortnight, so organizations can compare prices and accuracy on their own messages before choosing one.
Release 6 of 8Mistral Large 4 lets an organization run a powerful AI model on its own computers
- When
- 6 October
- Made by
- Mistral AI, a French company
- Who can use it
- Anyone, as an early preview; the model files are published
- Cost
- US$0.68 to read and US$2.09 to write a million tokens through Mistral's own service (sale price)
Most popular AI models stay on their makers' computers. When you use them, your words travel over the internet to that company's servers, and the answer travels back. Mistral Large 4 is an open-weight model. The weights are the files that hold everything a model learned in training, and Mistral publishes them, so an organization can download the model and run it on its own servers.
It can read pictures as well as text, and it can take in about 750,000 words at once, enough for a long legal case file or a year of board papers. It launched as a public preview on 6 October. It is also large, with more than a trillion internal settings, so running it in-house needs powerful and expensive servers.
Why it matters
Banks, hospitals and government ministries often cannot send sensitive records to a company in another country. A model they can run on their own servers removes that obstacle and supports data sovereignty, the idea that a country should control where its citizens' data is kept and who can see it.
Most institutions in the region will start by testing it through Mistral's online service. Before planning a private setup, read the licence terms on the download page, which set out what you are allowed to do with the model.
Release 7 of 8Nano Banana 2.1 makes AI images with text you can actually read
- When
- 6 October
- Made by
- Who can use it
- Anyone with the Gemini app; also in Google Ads and developer tools
- Cost
- Not listed by Google
Nano Banana is the nickname of Google's image AI, the tool inside the Gemini app that creates and edits pictures from a written description. Earlier image AIs were known for garbled writing on posters and signs. Google says version 2.1, released on 6 October, is built for posters, diagrams and infographics, with clearer text in several languages.
Google's own tests support that. On a measure of whether the facts in AI-made infographics are correct, scored from 0 to 1, Nano Banana 2.1 scored 0.52, up from 0.18 for the previous version.
Why it matters
Small businesses, churches, schools and Carnival bands make flyers every week. A tool that writes dates, prices and place names correctly saves design time and printing money.
Google's own notes still list weaknesses, including small text, so check every phone number, price and local place name before anything is printed or posted.
Release 8 of 8Microsoft's MAI-Transcribe-2 types up speech in 60 languages as people talk
- When
- 1 October
- Made by
- Microsoft
- Who can use it
- Businesses and developers through Microsoft's cloud services
- Cost
- US$0.54 per hour of audio until the end of 2026
Transcription means turning spoken words into written text. MAI-Transcribe-2-Streaming does it live, so words appear while a person is still speaking, and it works out which of its 60 languages is being spoken without being told. Microsoft released it on 1 October alongside two voice models that speak 23 languages and can copy a person's voice from a few seconds of recording; Microsoft says it built in checks to require consent.
Why it matters
Live transcription can caption a press conference, produce notes from a court hearing or give a call centre a written record of every call. A Caribbean call centre transcribing 1,000 hours of calls would pay about US$540 at the introductory price (our calculation).
Microsoft does not list its 60 languages, so buyers should not assume it understands Jamaican Patois, Trinidadian Creole or Haitian Creole. Test it on recordings of your own customers, with their permission, before buying. The voice-copying models also make scam calls that imitate a relative easier to produce, which is one more reason for the family code word in our tips.
Six more AI releases from the first two weeks of October
- Meta, 12 October. The company behind Facebook, Instagram and WhatsApp released Muse Wonder and Muse Wonder Full, AI models designed to carry out tasks and remember details about the user across conversations. They are available to developers through Meta's own services; Meta's announcement does not mention WhatsApp.
- Anthropic, 6 October. An expanded Claude for Startups programme which, TechCrunch reported, includes a free year of Claude Team and US$1,000 in credits. When we checked on 13 October, the programme page said the current group was full and new applications were closed.
- Amazon, 5 October. Nova 2.5 Sonic, an AI for natural voice conversations such as customer phone lines. It works in seven languages Amazon does not name and is offered from four Amazon data centre regions, in the United States, Sweden and Japan.
- OpenAI, 1 to 6 October. ChatGPT can now accept audio files on paid plans and scan notes and documents with the iPhone camera. Its money-management feature, Finances, reached free users on 2 October, in the United States only.
- Google, 6 October. EmbeddingGemma 2, a free model that helps apps search through photos and documents on the phone or laptop itself, without sending them online.
- JetBrains, 8 October. Mellum2.1, a free AI model for writing computer code, released under a licence that allows businesses to use and change it.
Nine tips for using October's new AI tools
- Teach checking alongside answers. Teachers and parents: ask students to explain why an answer is right and to find the mistake in a worked example, because judging an argument is the skill AI makes more valuable.
- Give GPT-6 a real job this week. Ask the free version for something you need, such as a hurricane supplies checklist or a household budget calculator in your own currency, then check every figure it produces.
- Keep private information out of personal accounts. A credit union officer should not paste a member's statement into a free ChatGPT account. Use your employer's approved AI tools, or remove names and account numbers first.
- Agree a family code word. Voice copying is now cheap. Pick a word only your family knows, so a relative can confirm that an urgent call asking for money is real.
- Test speech tools on Caribbean voices before buying. Record ten minutes of real calls, with permission, in the accents your customers use, and count the mistakes. A company's list of languages will not tell you how its AI handles Bajan or Vincentian speech.
- Do the maths before you buy. Multiply how many times a month a task happens by the tokens it uses and the listed price, as we did above. When the AI costs a few dollars a month, spend the budget on clean records and staff training.
- Proofread every AI image. Check prices, dates, phone numbers and place names on any flyer an AI makes, because a misspelled Ocho Rios on 500 printed flyers costs more than the design time it saved.
- Treat previews as previews. Mistral Large 4 is an early preview and the Gemini agent has no published price. Try them on low-risk work, and keep a ministry's or bank's core process off them until the terms are in writing.
- Avoid depending on one company. This fortnight brought similar tools from OpenAI, Microsoft, Cloudflare and Google. Set up your systems so that you can switch providers without starting again.
What Caribbean leaders should watch before November
- Whether mathematicians confirm the most important of OpenAI's proofs, and whether OpenAI releases the model that wrote them.
- Google's price for the Gemini agent, which its announcement did not include.
- When OpenAI's Decisions API leaves its test phase, which OpenAI expects within weeks.
- Independent tests of the new speech tools on Caribbean English and creole speech, work that regional universities and call centres are well placed to do.
- Whether features limited to the United States, such as ChatGPT's Finances tool, reach Caribbean users.
The October lesson for Caribbean organizations
Questions readers ask about October's AI releases
The 21 sources behind this brief
- OpenAI, "Sharing AI progress in mathematics," 6 October 2026
- Axios, "OpenAI's math breakthrough points beyond math," 8 October 2026
- Nature, "OpenAI posts 700 maths preprints online: mathematicians are up in arms," 7 October 2026
- OfficeChai, mathematicians react to OpenAI's proofs (quotes from X), 7 October 2026
- OpenAI, "GPT-6 and Intelligent UI for everyone," 7 October 2026
- OpenAI, ChatGPT release notes, October 2026 entries
- OpenAI Help Center, "What are tokens and how to count them"
- Google, Gemini at Work 2026 announcement, 8 October 2026
- TechCrunch, "Google brings agentic AI to Gemini, starting with businesses," 8 October 2026
- Anthropic, Claude Haiku 5.5 announcement and pricing, 7 October 2026
- Anthropic, "Introducing Claude Haiku 4.5," 15 October 2025
- Let's Data Science, OpenAI Decisions API public beta, 7 October 2026
- OpenAI, DevDay 2026 recap, 29 September 2026
- Mistral AI, Mistral Large 4 documentation, 6 October 2026
- Google DeepMind, Nano Banana 2.1 model card, 6 October 2026
- Microsoft AI, MAI-Transcribe-2-Streaming and MAI-Voice-2.1, 1 October 2026
- Meta, Muse Wonder announcement, 12 October 2026
- TechCrunch, Anthropic expands Claude for Startups, 6 October 2026
- Anthropic, Claude for Startups programme page, checked 13 October 2026
- Amazon Web Services, Nova 2.5 Sonic, 5 October 2026
- No Code MBA, AI release tracker, October 2026
