This Week's Releases: GPT-6 Reaches Every ChatGPT User, Claude Haiku 5.5, Google's Gemini Agent and Meta's Muse Wonder
Title: This Week's Releases: GPT-6 Reaches Every ChatGPT User, Claude Haiku 5.5, Google's Gemini Agent and Meta's Muse Wonder
Date: 14 October 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1427 words
Reading time: 6 min
Published: 14-10-2026
OpenAI began putting GPT-6 into ChatGPT for paying users on 7 October. Free and Go users followed on 8 October. Anthropic released Claude Haiku 5.5 on 7 October at $0.10 per million input tokens. Google Cloud announced a Gemini agent for work on 8 October. The Gemini agent is in private preview. Meta launched Muse Wonder on 12 October, and sells its larger version only inside a coding subscription. JetBrains and Liquid AI released smaller open models on 7 and 8 October.
GPT-6 in ChatGPT
OpenAI published GPT-6 and Intelligent UI for everyone on Wednesday 7 October. GPT-6 started rolling out that day in the Chat tab of ChatGPT for Plus, Pro, Business and Enterprise users. Free and Go users began getting it on Thursday 8 October. In Enterprise workspaces, administrators' settings decide when staff see it.
Paying users get GPT-6 Sol. Free and Go users get GPT-6 Luna, the smaller model. Both are tuned for everyday conversation. GPT-6.1 Sol, which OpenAI released at its DevDay conference on 29 September, stays in ChatGPT Work and Codex and is not yet in Chat.
Intelligent UI changes what an answer looks like. ChatGPT can now reply with charts, buttons, forms and small working tools mixed in with the text. It chooses the format for each question: a side-by-side comparison, a diagram, a map or a plain paragraph. One outlet reports that users will be able to reduce the amount of visual material.
OpenAI says that, for questions which need a web search, GPT-6 Instant starts answering 44 per cent sooner on average than GPT-5.6 Instant. That figure covers web-search questions only. It is OpenAI's own measurement.
Claude Haiku 5.5
Anthropic released Claude Haiku 5.5 on Wednesday 7 October. It is the smallest and fastest model in the Claude 5.5 family. Anthropic describes it as built for subagents and for high-volume work where cost matters. A subagent is a model that a larger agent hands smaller jobs to.
The price is $0.10 per million input tokens and $0.50 per million output tokens, for prompts of up to 100,000 tokens. Above that, the rate rises to $0.50 and $2.50, and the higher rate applies to the whole request. Haiku 4.5 cost $1 and $5. Haiku 5.5 keeps a context window of one million tokens and can write up to 128,000 tokens in one answer. One analysis reports that Haiku 5.5's new tokenizer counts about 30 per cent more tokens for the same text, so the real saving is smaller than the list prices suggest.
Anthropic's own figures put Haiku 5.5 at 72.4 per cent on the OSWorld 2.1 offline subset, a test of operating a computer, against 15.7 per cent for Haiku 4.5. On Terminal-Bench 4.0 it scores 39.2 per cent, well below Sonnet 5.5 at 70.6 per cent. Anthropic still points customers to Sonnet 5.5 and Opus 5.5 for complex coding agents.
Haiku 5.5 is generally available through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. AWS says it is also in GovCloud (US).
Google's Gemini agent
Thomas Kurian, chief executive of Google Cloud, announced a Gemini agent for work at the Gemini at Work event on Thursday 8 October. Google describes it as one agent for chat, longer pieces of autonomous work and coding. A user gives it an objective that may run over several days, and the agent plans and carries out the steps.
The agent works across Google Workspace, Microsoft 365 and Slack. Agents set up as persistent coworkers get their own Workspace account, with a calendar, a Drive and an entry in the company directory. Their addresses take the form @agents.company.com. "Every Gemini agent has an identity of its own, separate from your identity," Kurian said. Every action is logged against the agent's identity rather than a person's. VentureBeat reports that Google has not said whether every agent gets an account, or only the persistent ones.
Administrators set company rules once for all agents and can put a hard spending cap on each project. They can also send individual jobs to more than 200 models, including Anthropic's Claude as well as Gemini. Versions for financial services and legal work are in preview, with government, healthcare and retail versions to follow.
The agent is in private preview. Google has not published a date for general availability. Reports on price conflict. One cites a Google spokesperson saying there will be no additional charge where Gemini Enterprise is available, including existing Workspace business and enterprise plans. Other outlets say Google has not published pricing, or said whether long autonomous tasks will be charged by use.
Meta's Muse Wonder
Meta launched Muse Wonder on Monday 12 October. It is the newest model from Meta Superintelligence Labs, after Muse Spark in April. Muse Wonder comes in two sizes, and Meta sells them in different ways.
The standard Muse Wonder is available through Meta's Model API at $4 per million input tokens and $20 per million output tokens, with cached input at $0.20. It has a context window of about one million tokens. It accepts text, images, audio, video and PDFs. Meta has not published the weights.
The larger Muse Wonder Full is not in the API. CNBC reports it is available only to developers on Meta's Muse Code coding plans, at $50 or $200 a month. Reports differ on whether the $50 plan includes Wonder Full or only the standard model.
Artificial Analysis, an independent benchmarking firm, gives Wonder Full a score of 56 on its Intelligence Index. That ties Claude Sonnet 5.5 and is ahead of GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1, at 53 each. Only Claude Opus 5.5 scores higher among these, at 58. The standard Muse Wonder, the one developers can call through the API, scores 52. Most of Meta's headline benchmark figures are for Wonder Full.
Smaller open models
JetBrains, the maker of developer tools, released Mellum2.1 on Thursday 8 October. It is a mixture-of-experts model with 12 billion parameters, of which about 2.5 billion work on each token. It is meant for coding agents and fast subagents. The weights are on Hugging Face under the Apache 2.0 licence, which allows commercial use. The context window is 131,000 tokens, and the smallest downloadable build is 7 GB, for use with llama.cpp, Ollama or LM Studio.
JetBrains says the improvement came from reinforcement learning in real code repositories rather than a new design. On SWE-bench Verified, a test of fixing real software bugs, it reports 47.0 for Mellum2.1, against 2.0 for Mellum2. Alibaba's Qwen3.5-9B scores 50.0 on the same test. Mellum2.1 leads on LiveCodeBench v6 and the BFCL v4 tool-calling test. These are JetBrains' figures.
Liquid AI released two decision models on Wednesday 7 October. A decision model picks between answers set in advance and returns a probability for each. Cloudflare and Amazon released models of this kind on 1 October. Liquid AI's d1-3B has about 3.1 billion parameters and reads text and images. The experimental d1-omni-600M has about 600 million and reads text with images or audio.
One report says d1-3B answers a short question in 8 milliseconds on an Nvidia RTX 4090 graphics card. The smaller model scored 15.95 on the Decision Index. The weights are on Hugging Face under Liquid AI's own LFM Open License. One report says companies with annual revenue of $10m or more need separate permission for commercial use, so the licence text needs checking before deployment.
Before switching
Most of the performance figures above come from the companies that built the models. The Artificial Analysis scores for Muse Wonder are the main independent numbers this week.
For prompts under 100,000 tokens, Haiku 5.5's list price is a tenth of Haiku 4.5's. Because of the new tokenizer, the cost of an existing workload is worth measuring on real prompts before a budget is set. The higher rate on prompts over 100,000 tokens applies to the whole request.
Google's coworker agents get their own accounts and directory entries. In SAASiQ's view, that gives identity and access teams something they can manage like any other account: who owns the agent, what it can reach, and when its access is reviewed.
Developers who want the Muse Wonder that ties Sonnet 5.5 on the Intelligence Index cannot reach it through the API, so API tests will show the lower-scoring model.
Google has not set a general availability date or a price for the Gemini agent. Skills were due to reach the Gemini app from 13 October, and Gems move into the app's Settings panel on 17 November.
SAASiQ - Intelligent Solutions for SaaS ©
