This Week's Releases: OpenAI's GPT-6.1 Terra, Anthropic's Claude Duet, Kimi K3.1 Pro and Canva's First Image Model
Title: This Week's Releases: OpenAI's GPT-6.1 Terra, Anthropic's Claude Duet, Kimi K3.1 Pro and Canva's First Image Model
Date: 28 October 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1577 words
Reading time: 6 min
Published: 28-10-2026
OpenAI released GPT-6.1 Terra on 26 October. OpenAI says its standard API prices are half those of GPT-5.6 Terra, the model it replaces. Anthropic opened a waitlist on 21 October for Claude Duet, an agent that checks a user's apps and flags what needs attention. Anthropic made Claude Security re-scan code as it changes, also on 21 October. Canva released Craft, its first in-house image model, on 22 October. Google expanded speech translation in Meet to 63 languages on 22 October. Moonshot AI released Kimi K3.1 Pro on 23 October. NVIDIA released Nemotron 3.5 Bolt E4B, a small open model for local assistants and robots, also on 23 October.
OpenAI's GPT-6.1 Terra
OpenAI released GPT-6.1 Terra on Monday 26 October. It is OpenAI's middle model. It sits between the cheaper GPT-6 Luna and the more capable GPT-6.1 Sol and GPT-6 Astra. OpenAI aims it at computer use, reading what is on a screen, and the writing and form-filling that make up much of an office day.
The API price is $1 per million input tokens and $5 per million output tokens, with cached input at $0.05. OpenAI says that is 50 per cent lower than GPT-5.6 Terra. OpenRouter also lists a batch rate of $0.50 and $2.50. The context window is about 1.05 million tokens, with up to 128,000 tokens of output.
In ChatGPT, Terra is rolling out to Plus, Pro, Enterprise and Edu users in ChatGPT Work and in Codex. It is not offered on the free tier, and it is not in the regular Chat tab. Developers call it as gpt-6.1-terra. The Codex command-line tool needs version 0.159.0 or later to select it.
The benchmark figures are mostly OpenAI's own. On Agents' Last Exam, a test of professional tasks, OpenAI reports 60.0 per cent at about $1.01 per task. It gives 50.7 per cent for GPT-5.6 Terra at $3.58, 58.6 per cent for GPT-6.1 Sol at $2.00 and 59.3 per cent for GPT-6 Astra at $7.24. Decrypt pointed out that all of those numbers come from OpenAI.
Terra is weaker on coding in a terminal. Kingy.ai reports 50.3 per cent on Terminal-Bench 4.0 at maximum effort, up from 21.5 per cent for GPT-5.6 Terra. That is below GPT-6.1 Sol at 55.9 per cent and Claude Sonnet 5.5 at 70.6 per cent. Artificial Analysis data, as reported by eesel.ai, shows Terra guessing when it does not know an answer 38 per cent of the time, down from 88 per cent for GPT-5.6 Terra.
Anthropic's Claude Duet
Anthropic opened Claude Duet as a research preview on Wednesday 21 October. Duet is an agent that works in the background. Each morning it sends a summary of the email, calendar entries and other items that need the user's attention. It then checks the connected apps every few hours and sends an alert when something new comes up. Anthropic says it runs on Claude Opus 5.5.
Duet uses the user's existing Claude account, including memory, skills and connectors. More than 6,000 connectors are available, among them Gmail, Slack, QuickBooks, Stripe, Gusto and HubSpot. Before Duet sends a message, makes a payment, signs a document or deletes a file, it shows what it plans to do and waits for approval.
Access is through a waitlist for Claude Pro and Max subscribers in the United States. Free accounts are excluded, and Team and Enterprise workspaces cannot switch it on yet. Duet is included in the Pro and Max prices, and its background checks count towards the plan's usage limits. Reports differ on the apps: 9to5Google describes the web and a standalone iOS app, while Beebom lists iOS and Android as well as the web.
Claude Security scans as code changes
Anthropic also updated Claude Security on 21 October. The tool has been in public beta for Claude Enterprise customers since 30 April. A company connects its code repositories once. Claude Security runs a full first scan and then scans again as changes land.
When a fix ships, Claude Security checks the changed code and marks the finding fixed, even if the team fixed it outside the tool. If the problem comes back, the finding reopens. Repositories connect through the Claude GitHub app, including GitHub Enterprise Server, or with an access token from GitLab and other Git providers. Findings can be passed to other tools through an API and an MCP connector. Anthropic says access for Team and Max plans is coming, without a date.
Canva's Craft
Canva released Craft on Thursday 22 October. It is the first image generation model built by Canva Research. Craft is in the Dream Lab tab for all Canva users, within each plan's allowance, and in Canva's public API.
Craft works in two stages. A planning model lays out the design, the position of objects and the text. A second model then renders the image. Output goes up to 2K resolution, with about 30 preset aspect ratios, and Canva says it can render text in 55 languages. Canva trained Craft mainly on its own library of designs made by people, rather than on photographs.
Developers reach it through Canva's Connect API under the name craft_v1_2k. Each request uses the account's AI credits. A Canva staff post put the time for one request at about 50 to 60 seconds. Canva has not published a price per image.
Craft entered the Design Arena image leaderboard in fourth place, with 1,341 points. The ranking is preliminary and based on relatively few votes. Canva says the result makes it the second-best image lab on the board, behind OpenAI. Cheaper versions called Craft-fast and Craft-ultrafast are planned, with no prices or dates.
Meet translates 63 languages
Google expanded Gemini speech translation in Google Meet to 63 languages on 22 October, on the web, Android and iOS. Until now, a meeting had to be set to a single pair of languages. Each participant now switches translation on from the toolbar and chooses their own language.
Meet detects which language is being spoken, including when a speaker changes language part way through.
What a meeting gets depends on the organiser's plan. Technobezz reports up to 90 minutes per meeting on eligible Workspace plans and Google AI Pro, and up to eight hours on Google AI Ultra and Google's expanded access plans. The rollout began on 22 October for Workspace domains on Rapid Release.
Moonshot's Kimi K3.1 Pro
Moonshot AI released Kimi K3.1 Pro on 23 October, which was 24 October in China. It is the first model in Moonshot's K3.1 series. Moonshot calls it a post-trained version of Kimi K3, with the same design and much more training after the base model was built. The context window is one million tokens.
The API price is $2 per million input tokens and $10 per million output tokens, down from $3 and $15 for Kimi K3. A faster version, K3.1 Pro Highspeed, costs $4 and $20. Moonshot says it produces output two to three times as fast. The model is on Kimi.ai, the Kimi apps, Kimi Code, the Kimi API and OpenRouter.
Moonshot reports 65.7 per cent on Terminal-Bench 4.0. On Moonshot's figures that is just behind Claude Opus 5.5 at 66.4 per cent and ahead of GPT-6 Astra at 58.2 per cent. The Decoder reports a score of 55 on the Artificial Analysis Intelligence Index, 11 points above Kimi K3. That figure appears in one report only.
The weights are not out yet. Moonshot has promised them by 4 November. It has not announced the licence.
NVIDIA's Nemotron 3.5 Bolt E4B
NVIDIA released Nemotron 3.5 Bolt E4B 4GB on 23 October. It is a small open model for assistants that run on a local machine, characters in games and robots that hold a conversation and use tools. NVIDIA built it from Google's open Gemma 4 E4B model, trained further with reinforcement learning only.
The model has 8 billion parameters, 4 billion of them effective, and a context window of 64,000 tokens. It handles text only. The 4GB in the name is the graphics memory NVIDIA says it needs. The download itself is 6.18GB, one report notes. It runs on GeForce RTX cards, Jetson boards, RTX Spark and DGX Spark. The weights are free on Hugging Face under the OpenMDW-1.1 licence, which allows commercial use.
NVIDIA's demonstration ran a Reachy Mini robot on an 8GB Jetson Orin Nano board, with Bolt handling the conversation. The release follows the larger Nemotron 3.5 Bolt 74B, released on 15 October.
Before switching
Most of the benchmark figures this week come from the companies that built the models. Artificial Analysis data on Terra and a single report of its score for Kimi K3.1 Pro are the main independent numbers.
On OpenAI's own figures, Terra scores about the same as GPT-6 Astra on professional tasks at about a seventh of the cost per task. Teams running GPT-5.6 Terra can test the new model on their own workloads before moving.
Duet waits for approval before it sends, pays, signs or deletes. For now it is limited to personal Pro and Max plans in the US.
In SAASiQ's view, Kimi K3.1 Pro is the release to watch for teams that want an open model, but only once the weights and the licence are published.
Moonshot has promised the Kimi K3.1 Pro weights by 4 November. Google's Meet translation reaches Scheduled Release domains from 2 November, a rollout that can take 15 days. Anthropic has given no date for wider access to Claude Duet.
SAASiQ - Intelligent Solutions for SaaS ©
