This Week's Releases: Anthropic Ships Claude Fable 5.2 at Lower Prices, Meta Launches Muse Ember, OpenAI Releases GPT Image 2.6 and Gemini Makes 15-Second Videos

Title: This Week's Releases: Anthropic Ships Claude Fable 5.2 at Lower Prices, Meta Launches Muse Ember, OpenAI Releases GPT Image 2.6 and Gemini Makes 15-Second Videos
Date: 11 November 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1556 words
Reading time: 6 min
Published: 11-11-2026
Anthropic released Claude Fable 5.2 on 9 November and cut its standard price by 20 per cent. Meta launched Muse Ember, its smallest Muse model, on 9 November. OpenAI released GPT Image 2.6 on 4 November, in a fast version and a higher-quality version. Google released Gemini Omni 1.2 Flash on 5 November, which makes video clips of up to 15 seconds. Anthropic opened a Personal Finance beta in Claude on 5 November. Mistral added managed agents to Mistral Studio on 5 November. Moonshot published the weights of Kimi K3.1 Pro on 4 November.
Claude Fable 5.2
Anthropic released Claude Fable 5.2 on Monday 9 November, with Claude Mythos 5.2. Fable 5.2 is on all paid Claude plans, in Claude Code, on the Claude Platform and on the main cloud platforms. Mythos 5.2 is for vetted organisations only, through Anthropic's Life Sciences and Cyber Verification Programs.
The two models are built on the same underlying model. The public Fable 5.2 carries extra safeguards on cyber security and biology to limit misuse. Coverage of the launch says Fable 5.2 blocks about 55 per cent fewer requests than Fable 5.1 did.
The standard API price is now $8 per million input tokens and $40 per million output tokens. Fable 5.1 cost $10 and $50. Batch requests, which run without an immediate answer, cost half that: $4 and $20. A new fast mode generates text up to 2.5 times faster, at double the standard price. It runs in Claude Code, on the Claude Platform, on Google Cloud and in Microsoft Foundry. Anthropic says ordinary output is up to 30 per cent faster than the 5.1 models.
Anthropic says Fable 5.2 at its High effort setting matches Claude Opus 5.5 at maximum effort on coding, at about half the cost per task. It gives a FrontierCode 2.0 score of 49.1 per cent at the highest effort. On the same test it lists 45.5 per cent for Opus 5.5 and 42.8 per cent for OpenAI's GPT-6 Astra. These are Anthropic's figures.
Anthropic also says Fable 5.2 solved two open problems in theoretical physics, on black holes and quantum gravity. No independent check of that claim had appeared in the coverage we found. A customer quoted on Anthropic's product page says the model made the phonon code in Quantum ESPRESSO, a physics simulation package, about 36 times faster on average on the same GPU. Anthropic says the UK AI Security Institute and METR were among the outside groups that tested both models before launch.
Meta's Muse Ember
Meta launched Muse Ember on 9 November. Most reports give that date, and one model tracker lists 10 November. Ember is the smallest and cheapest model in Meta's Muse range. Meta built it for the repeated steps inside agent work: calling tools in order, reading files, running tests, making small code fixes and pulling data out of documents. Meta's plan is for a larger model, such as Muse Wonder, to do the planning and hand those steps to Ember.
Ember reads text, images, audio and video, and takes up to one million tokens at once. It costs $0.10 per million input tokens and $0.50 per million output tokens, through the Meta Model API and Muse Code. Meta is keeping the weights private. Ember is the fourth Muse model, after Spark in April, its open-weights twin Glimmer in August and Wonder in October.
Meta says Ember gives more capability per dollar than Anthropic's Claude Haiku 5.5 or OpenAI's GPT-6 Luna. Independent results are mixed. On the FrontierCode 2.0-AA coding test, Haiku 5.5 scored 34.9 per cent and Ember 27.2 per cent, according to coverage of the launch.
OpenAI's GPT Image 2.6
OpenAI released GPT Image 2.6 on Wednesday 4 November, as two models in its API. GPT Image 2.6 Flare is the fast one. OpenAI says it can return an image in as little as six seconds. GPT Image 2.6 Sunburst takes longer and gives more detail and more consistent results across a series of edits.
The release adds animated GIFs, export to SVG, a vector format that designers can edit, and reference images that stay attached to a conversation. Users can also mark up an image to show what should change. OpenAI says the new version is in ChatGPT and Codex on every plan.
The API price per token is the same as GPT Image 2.5: $8 per million image tokens in and $30 per million out. The Batch API, which halves the price for requests that can wait, was not supported at launch. Analysts disagree on the real cost per picture. One says the new version uses about four times as many tokens at low and medium quality. Another found that token counts matched the old version to within 3 per cent. Artificial Analysis ranked Sunburst first for text-to-image on its leaderboard on 7 November.
Gemini Omni 1.2 Flash
Google released Gemini Omni 1.2 Flash, a video model, on Thursday 5 November. It makes clips of up to 15 seconds. The previous version, 1.1, stopped at 10 seconds. Users can set any length from three to 15 seconds, in one-second steps.
The price in the Gemini API is about $0.07 per second of video at 720p, down from about $0.10 for version 1.1. A full 15-second clip at that resolution costs about $1.05. The model is generally available in the Gemini API and on Google Cloud. In the Gemini app, 15-second clips go first to Google AI Ultra subscribers. Google Vids, the video tool in Workspace, offers 15-second clips at no extra charge within its monthly AI limits.
Users can now edit a whole 15-second clip over several turns of a conversation. Reviewers report that faces, clothing and product labels stay more consistent from shot to shot.
Personal Finance in Claude
Anthropic opened a beta of Personal Finance in Claude on 5 November. It is for Pro, Max and Ultra subscribers in the US. Free, Team and Enterprise accounts cannot use it.
Users link their bank and card accounts through Plaid, a service that connects apps to banks. The user signs in to the bank on Plaid's screen, so Claude never sees the password. The user chooses which accounts to share. Claude can then see the bank's name, each account's name and type, balances, credit limits and up to two years of transactions.
Claude can total spending by category, find unexpected fees and repeat charges, and draw charts. Access is read-only, so Claude cannot move or invest money. Anthropic says the feature does not give investment advice, and that conversations in the Personal Finance area are not used to train its models. Most coverage describes bank and card accounts only. One outlet says loans and investment accounts are included, which we could not confirm.
Managed agents in Mistral Studio
Mistral added managed agents and apps to Mistral Studio, its platform for building on its models, on 5 November. Both are in public preview. Mistral also announced that its workflow builder is generally available on Mistral Cloud.
Developers write agents in Python with Mistral's agents SDK. The same agent can run inside the developer's own program, on a server the customer runs, or on Mistral's hosting. Apps are internal tools built in Studio, with hosting, sign-in and access controls included. Mistral says private cloud and on-premises versions follow at the end of November. It has not published prices for managed agents, apps or the workflow builder.
Muse by phone and video
Meta made voice and video calls with its Muse assistant available to everyone in the US on 5 November. Alexandr Wang, Meta's chief AI officer, announced the rollout.
A user taps the phone icon in a Muse conversation to start a call. Muse answers in a synthetic voice, with an animated avatar whose face and lips follow the conversation. Calls work in the Muse app on mobile, on the web and on Mac. Meta says two new models drive it: Muse Realtime Voice and Muse Realtime Avatar. Voice is due on Meta's AI glasses later, with no date given.
Kimi K3.1 Pro's weights
Moonshot AI published the weights of Kimi K3.1 Pro on Hugging Face on 4 November. Moonshot had promised them by that date when it launched the model in late October. SAASiQ covered the launch on 28 October.
Kimi K3.1 Pro has about 2.8 trillion parameters in total, with 104 billion active for each word it produces. It reads text and images and takes just over one million tokens of context. The download is about 1.56 terabytes, which means it needs several linked servers to run. Coverage describes the licence as a modified MIT licence, with conditions for very large products.
Before switching
Most of this week's benchmark figures come from the companies that made the models. The FrontierCode result that puts Claude Haiku 5.5 ahead of Muse Ember is one of the few independent comparisons.
Teams that run agents at volume can compare Fable 5.2's new prices, and Ember's, against what they pay now, using their own logs. In SAASiQ's view, the price changes this week matter more to most organisations than the benchmark tables, because an agent's bill grows with every step it takes.
Mistral's private cloud and on-premises versions of Studio are due at the end of November. Microsoft Ignite runs from 17 to 20 November in San Francisco.
SAASiQ - Intelligent Solutions for SaaS ©


