News
Checked AI releases from September and October 2026, with the price, the limit, and what changes for someone who uses a chat box. A score below is the company's own figure unless the source names an outside test.
Latest news
Meta Launches Ember: Smallest, Most Efficient AI Model Yet
Meta has introduced Ember, a compact and efficient AI model designed for complex reasoning tasks and repetitive agent tasks like tool calls, coding, routing, extraction, and long-document analysis. Ember offers more intelligence per dollar than its predecessors, Claude Haiku 5.5 and GPT-6 Luna, and is available on Meta Model API.
AI Research Speeds Up, RSI Closer by 2027
A report from METR shows that AI research at Anthropic is improving at a rate nearly twice as fast as between March 2024 and October 2025. This acceleration is largely due to AI performing AI research, suggesting recursive self-improvement (RSI) could occur by 2027, possibly even within a year.
Fireworks Hit by Financial Cybercriminals in October Security Breach
In October 2026, a security breach at Fireworks exposed account and billing information, as well as partial datasets and unfinished models, to financially motivated cybercriminals. The incident was detected by a third-party security provider on October 23 and confirmed by Fireworks the next day. This breach highlights the ongoing risks of data exposure and the importance of robust security measures for both users and companies.
Fable 5.2: AI Model Near AGI Capabilities
Fable 5.2, a new AI model, has shown capabilities close to artificial general intelligence (AGI). It can create code, videos, and business ideas at an expert level, though it has some limitations in identifying issues in its solutions.
New AI Phenomenon Boosts Advanced Function Development Speed
A new phenomenon called computational extrapolation can accelerate the development of advanced AI functions and complicate the assessment of dangerous capabilities. This effect is measured by observing performance changes with increased calculations, such as more tokens in chain-of-thought or more agents in a multi-agent LLM swarm.
The author shares a method to handle burnout and loss of focus when working with AI tools like Claude Code or Codex. By organizing tasks into a prioritized queue and processing each issue one at a time, the approach helps maintain focus and manage context effectively.
Scott Aaronson and Andris Ambainis proposed the Aaronson-Ambainis conjecture in 2009. A startup, Physical Superintelligence, has now proven the conjecture using their Claude model, showing a variable influence of at least ~1/d^6, verified by human experts and Lean certification.
Despite their proficiency in writing regular programs, large language models (LLMs) are poor at designing AI systems. They tend to create overly complex symbolic systems instead of leveraging other LLMs, leading to inefficient solutions.
Alex Stamos, chief security officer at Cognition, has described the current cybersecurity landscape as a 'dark old age' due to AI-supported attacks. Recent months have seen significant vulnerabilities and breaches, including a major hack at Binance, as AI tools like ARTEX can now identify and exploit these weaknesses in real-time.
AI systems from Anthropic have reportedly achieved a dramatic acceleration in AI research, doubling their development capabilities compared to previous rates. This surge, observed between October 2025 and 2026, is largely attributed to AI automating AI research, potentially making research ten times more efficient.
Anthropic has developed a method for improving AI memory called 'dreams,' where the AI Claude organizes and enhances its memory based on past experiences. This process helps Claude perform better in tasks by creating and using verified rules, as seen in the Continual Learning Bench 1.0.
Beginner
What a model announcement is telling you
A model release is a new version of an AI system you can talk to, or that a program can call. Ask what you can do this week that you could not do last month, and what you pay.
Three terms help you read the headlines. A token is a small piece of text the system reads or writes, and companies bill per million tokens. A context window is how much text fits in one job, from a short email to a long contract. Open weights means the company publishes the model file, so someone can run it on a machine they control. A chat-only product keeps that file on the company's computers.
From 21 September to 17 October 2026, Anthropic shipped Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5. SpaceXSI shipped Grok 4.7 with a published token price. Mistral opened Large 4, trained in Europe, as a preview and promised downloadable weights by the end of October. DeepSeek published an updated open model and a web-search tool. Google put Gemini 4 Argon in the Gemini app for AI Pro and Ultra subscribers and opened it to paying developers. Meta opened Muse Wonder to developers.
Each note below gives the date, the change, and what it means in a chat box. Open the linked page and read its price table. A score is the company's own test unless that page names an outside lab.
Explain one AI announcement in plain language
Read the announcement I paste below. I may never have used an AI tool. In short sections, answer: what changed on the date of the post; what I can do differently this week if I only use a chat box; any price, limit, or subscription named in the post; and which sentence is the company's own claim rather than a measured fact. Quote the price exactly if one appears. If the post gives no price, say the price is not in the post. Skip slogans. End with three checks I can do myself on the linked page. Announcement: [paste the announcement here]
Intermediate
Prices, the next EU date, and a fair test
You can check the price yourself. Output tokens, the words the model writes, usually cost more than input tokens, the words you send. Cache reads are a discount for text the service kept from an earlier step. An agent browses, edits files, and calls tools in a loop, so those repeated reads dominate the bill.
Rates from the companies' own pages, in US dollars per million tokens. Grok 4.7 (21 Sep): from $2 input and $6 output, with faster serving at twice the price. Claude Opus 5.5 (22 Sep): $4 and $20, cache reads $0.20. Claude Sonnet 5.5 (28 Sep): $2 and $10, cache reads $0.10 from 7 Oct. Mistral Large 4 (6 Oct): $0.70 and $2.10 on Mistral's model page, preview, weights by the end of October. Claude Haiku 5.5 (7 Oct): $0.10 input and $0.50 output up to 100,000 prompt tokens, then $0.50 and $2.50.
DeepSeek's 17 October note keeps the old programming price and adds search at $0.75 per 1,000 calls. Google's launch post for Argon gives an introductory API rate of $2 input and $10 output until 30 November 2026, then $4 and $20. Meta's Muse Wonder post prints no token rate. Treat a missing rate as unknown.
The next EU date is 2 December 2026. The Commission AI Act Service Desk (https://ai-act-service-desk.ec.europa.eu/en/ai-act/eu-ai-act-implementation-timeline) says bans then apply to systems that generate non-consensual sexual deepfakes and child sexual abuse material, and some providers must finish Article 50(2) marking of synthetic content. High-risk rules for hiring and credit wait until 2 December 2027. Run one document you already understand on two models and record time, edits, and the invoice.
Compare two model prices on one real task
I will paste two vendors' pricing sections and describe one task I actually run. Using only the numbers in what I paste, estimate input tokens, output tokens, and cache reads. Show the arithmetic. Say which assumptions would make the estimate wrong, including a prompt length that crosses a published threshold. Then list a fair test: same source document, same instructions, and what I will record (time, edits I had to make, and the invoice). Do not invent prices that are absent from my paste. Pricing text: [paste] Task: [describe]
Expert
Active parameters, compressed files, and how to read a score
Several of these releases are mixture-of-experts models. A mixture of experts (MoE) stores many specialist subnetworks and runs only a few for each token. Mistral describes Large 4 as an MoE with about 1 trillion parameters and 52 billion active. Speed and the token bill follow the active slice. Memory still has to hold the full file unless you load a compressed copy.
Hugging Face lists Muse-Spark-Open-816B-A42B for download. Quantization stores each weight with fewer bits so the file fits in less memory, and it can change answers. Compare the full file and the compressed file on the same prompts.
Read the context window on the model card: maximum tokens in, maximum tokens out, and whether images share that budget. DeepSeek's 17 October post puts the new Flash weights on Hugging Face and keeps the API name deepseek-flash.
Vendor tables quote Terminal-Bench, DeepSWE, and GDPval-AA. Compare scores only when the harness, the effort setting, and the safeguards match. Anthropic says some Opus 5.5 scores used production safeguards, and blocked cyber or biology tasks were finished by an older model. The arXiv cs.AI recent list (https://arxiv.org/list/cs.AI/recent) shows hundreds of new submissions a day. Read the PDF protocol. Language-model methods are on the cs.CL list (https://arxiv.org/list/cs.CL/recent). Model files are indexed at https://huggingface.co/models and agent code shows up on https://github.com/trending.
Review a model release like a systems engineer
I will paste a model announcement and, if I have it, the model card. Write a technical note with: architecture (dense or mixture-of-experts, total parameters, active parameters); context window and output cap; modalities; license and whether weights are downloadable; quantization or reduced-precision files; API price per million tokens, including cache and long-context tiers if stated; which benchmarks name the harness, effort, and whether safety filters changed the score; and three failure modes for a production agent (prompt injection, silent truncation, cost runaway). Mark every figure as vendor-reported or independently measured. If a figure is missing, write unknown. Sources: [paste]
17 October 2026 — DeepSeek updates its open coding model and opens search
DeepSeek published DeepSeek-V4.1-Flash-1017 on its website, app, and programming interface. It says tasks finish faster and answers are shorter than on the previous V4.1 Flash, at the same programming price. Weights are on Hugging Face, so you can run the file yourself. The note also opens web search at $0.75 per 1,000 calls. Search and fetch calls made through DeepSeek Harness, Claude Code and Codex are free until 31 October 2026 at 16:00 UTC (17:00 Swedish time); after that they cost $0.75 per 1,000 calls. Source: https://api-docs.deepseek.com/news/news261017/
13 October 2026 — Gemini 4 Argon reaches paying Gemini app users
Google turned on Gemini 4 Argon in the Gemini app for AI Pro and Ultra subscribers. It says answers are more direct and show progress on long jobs. On Mac, chat can switch into Antigravity, Google's coding workspace. The Tasks feature in Gemini can keep working in the background after you leave the app. Google says the app connects to more than 40 products, including Instacart, OpenTable, Adobe, and Shopify. Free plans are outside this post, and no token price is printed. Source: https://blog.google/innovation-and-ai/products/gemini-app/gemini-4-argon-availability/
13 October 2026 — Paying developers can call Gemini 4 Argon
The same day, Google opened Gemini 4 Argon on the paid programming interface in Google AI Studio and Android Studio. Ultra subscribers can also try it in AI Studio and Antigravity. Google calls it the model for long software work, company research, and cybersecurity defense, including multi-step code migrations. A safety note by Anca Dragan and Raluca Ada Popa sits beside the post. This is the release that lets a product call Argon. According to Google's launch post and the Gemini API pricing page, the introductory price is $2 per million input tokens and $10 output through 30 November 2026, then $4 and $20. Source: https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-argon/ Price: https://ai.google.dev/gemini-api/docs/pricing
12 October 2026 — Meta opens Muse Wonder for multi-step work
Meta announced Muse Wonder and Muse Wonder Full for jobs with many steps, such as reading a lease or building a small app. Developers get Muse Wonder on the Meta Model API and in Muse Code. Muse Wonder Full is for Muse Code subscribers. Meta says the models carry relevant details across chats, keep going after errors, and ask before important actions. They are trained to refuse harmful requests and to withhold private data. No token price is printed. Subscribers get the consumer app later. Source: https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
8 October 2026 — Anthropic rewrites the rules for using Claude
Anthropic published a Usage Policy that takes effect on 12 November 2026. It says most edits only make existing rules clearer. A new section bans fake accounts, hidden authorship, and fabricated news sites. Elections language now targets deceiving voters. Weapons language covers guidance software. Surveillance language bans tracking people without consent and using Claude to choose who to investigate. Health, legal, and finance uses still need a person who can overrule the model. Read this before 12 November if you use Claude at work. Source: https://www.anthropic.com/news/2026-usage-policy-update
7 October 2026 — Claude Haiku 5.5 cuts the cost of short, repeated jobs
Anthropic released Claude Haiku 5.5 on the Claude apps and on Amazon, Google, and Microsoft clouds. The programming name is claude-haiku-5-5. Prompts up to 100,000 tokens cost $0.10 per million input tokens and $0.50 output. Longer prompts cost $0.50 and $2.50. Anthropic says a typical job costs about 75% less than Haiku 4.5. The same day, Sonnet 5.5 cache reads fell from $0.20 to $0.10 per million. Use Haiku for summaries and sorting. Leave hard judgment to a larger Claude model. Source: https://www.anthropic.com/claude-haiku-5-5
6 October 2026 — Mistral previews Large 4, trained in Europe
Mistral opened a public preview of Mistral Large 4 on Mistral Studio. It is a mixture-of-experts model with about 1 trillion parameters and 52 billion active, for text and images. Mistral's model page lists $0.70 per million input tokens and $2.10 output. Mistral says the weights come by the end of October, after more cybersecurity testing. Training used 3,800 NVIDIA Grace Blackwell chips in Mistral's European data centers. Training data covered more than 160 languages, including every official EU language. The post gives the weights date. Source: https://mistral.ai/news/mistral-large-4/ Price: https://docs.mistral.ai/models/mistral-large-4
28 September 2026 — Claude Sonnet 5.5 speeds up everyday work
Anthropic released Claude Sonnet 5.5 for routine documents and code. List price stays $2 per million input tokens and $10 output. Anthropic says fewer tokens per task make most work up to 30% cheaper than Sonnet 5, and output more than 30% faster. On its Terminal-Bench 4.0 run the score is 70.6%, against 10.3% for Sonnet 5. The programming name is claude-sonnet-5-5 on Anthropic's platform and on Amazon, Google, and Microsoft. Ordinary bug-fixing still runs. Higher-risk cyber requests fall back to Sonnet 5. Source: https://www.anthropic.com/claude-sonnet-5-5
22 September 2026 — Claude Opus 5.5 lowers the price of the largest Claude
Anthropic released Claude Opus 5.5. Input is $4 per million tokens and output is $20, down from $5 and $25 on Opus 5. Cache reads are $0.20, down from $0.50. Anthropic says a typical job costs about 40% less and output is more than 30% faster. The programming name is claude-opus-5-5. A faster mode costs $8 and $40 per million. Outside testers, including METR, reviewed it before release. Safeguards match those on Anthropic's most restricted models. Long sessions should cost less, but check the facts in its answers yourself. Source: https://www.anthropic.com/claude-opus-5-5
21 September 2026 — Grok 4.7 ships with a published token price
SpaceXSI released Grok 4.7 for coding and knowledge work, priced from $2 per million input tokens and $6 output, the same start as Grok 4.6. A faster option doubles speed and price. It is in Cursor, Grok Build, and the Grok programming interface. SpaceXSI says a longer training run improved self-checking, and it prints its own scores against Grok 4.6, GPT-5.6 Sol, and Claude Fable 5.1. Those figures are the vendor's. Set this rate beside Claude and Mistral, and read the docs for long prompts. Source: https://x.ai/news/grok-4-7
Polaris plans to raise money for a shared DGX Station in Sweden. If the station is bought, open models like DeepSeek-V4.1-Flash-1017 could be run and tested on it, with results going into the free library. About the shared station
Sources
- DeepSeek-V4.1-Flash-1017 announcementOpen-weight update, unchanged API price, and Search and Fetch rates.
- Gemini 4 Argon in the Gemini appWhich paid Gemini plans get Argon, tasks, and connected apps.
- Gemini 4 Argon for developersPaid API access in AI Studio and Android Studio, with no token price on the post.
- Muse Wonder announcementWhere Muse Wonder and Muse Wonder Full are available.
- Anthropic Usage Policy updateRules that take effect on 12 November 2026, and the full policy link.
- Claude Haiku 5.5Small-model prices and the Sonnet 5.5 cache-read cut.
- Mistral Large 4European training setup, preview price, and the weights timetable.
- EU AI Act implementation timelineCommission dates, including 2 December 2026 and the later high-risk deadlines.
- Introducing Gemini 4 ArgonGoogle's launch post, with the introductory API price and the later rate.
- Claude Opus 5.5Prices, the faster mode and outside testing.
- Claude Sonnet 5.5List price, speed and the Terminal-Bench run.
- Grok 4.7Token price and the vendor's own scores.
- Inception Mercury 3List price and launch price.