This Week's Releases: Claude Browser Use, AgentCore Payments, DeepSeek Off-Peak Pricing and a Gemini Student Hub
Updated: 6 days ago
Title: This Week's Releases: Claude Browser Use, AgentCore Payments, DeepSeek Off-Peak Pricing and a Gemini Student Hub
Date: 20 August 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1417 words
Reading time: 6 min
Published: 20-08-2026
Anthropic made its computer use tool generally available on 19 August and added a separate tool for operating websites, so developers can now have Claude work a desktop or a browser as a supported feature. The day before, AWS released a service that lets AI agents pay for things on their own, within spending limits set in advance. DeepSeek brought in half-price off-peak rates for its largest model, Anthropic explained the invisible watermark it now puts in Claude's writing, and Google added a student hub to the Gemini app.
Claude can operate a desktop or a browser
Computer use lets Claude operate a computer much as a person does, by looking at the screen and then clicking and typing. It had been a beta feature. It is now generally available on the Claude API as computer_toolset_20260801 and no longer needs a beta header in the request. It can carry out several actions in one turn and has zoom switched on by default. Earlier beta versions remain available, but moving to the new version changes the shape of the request, so existing integrations need a code change.
The browser use tool, browser_toolset_20260801, is new. It drives a web browser hosted by the customer's own application. As well as taking screenshots and clicking, it reads the page's accessibility tree, the structured outline of buttons, links and fields that screen readers rely on, along with the page's elements, forms and open tabs. It can fill in forms, manage tabs and report downloads, and it can upload files if the developer switches that on.
In plain terms, operating a website on someone's behalf is now a supported Claude feature. Both tools work with Claude Fable 5, Mythos 5, Opus 5, Sonnet 5 and Opus 4.8, and both are available on Google Cloud from 20 August. The release notes give no new pricing.
Other Anthropic changes
Several other Anthropic features left beta on 19 August. The Files API, for uploading files to use with Claude, gained file expiry and pagination. Agent Skills and the Skills API are also generally available; skills are reusable instructions that an agent can pick up when a task calls for them. So are the user management endpoints in the Claude Enterprise Admin API, used for managing users.
In Managed Agents, Anthropic's hosted service for running agents, the web_search and web_fetch tools can now be restricted with lists of allowed or blocked domains, and self-hosted sandboxes can attach memory stores. The Console Workbench was renamed Playground on 18 August, and the Python SDK reached version 1.0 on 20 August.
AWS lets agents pay for services
Amazon Bedrock AgentCore Payments became generally available on 18 August. It lets an AI agent find, access and pay for paid APIs, MCP servers and content by itself. MCP servers are the connectors that give an AI assistant access to other software and data, and some of them charge for use. The service had been in preview since May and was built with Coinbase and Stripe.
Payments come from wallets run by Coinbase or Stripe Privy, using short-lived tokens, and the AWS console has a Quick Create option for setting up Coinbase credentials. Two payment protocols are supported. One is x402, which now includes an 'upto' scheme for pay-per-inference and dynamic pricing; the other is the Machine Payment Protocol, or MPP.
Each payment session has a maximum spend, in a stated currency, and an expiry time. AWS enforces both at the infrastructure layer, outside the agent, so the agent cannot raise its own limit. Spending can be followed in AgentCore Observability and CloudWatch dashboards. The service works with Claude Code, Kiro and Codex, and is set up through the AgentCore CLI or the console.
AWS names Anchor Browser, Travala (which books hotels across 2.2 million properties), BlockRun/SpreadX (Incarna), Elsa AI and Heurist AI as customers. The announcement gives no prices and points to the AgentCore pricing page instead. It does not list regions either, so availability in London or elsewhere in Europe needs checking against AWS's regional services list.
DeepSeek halves its off-peak prices
DeepSeek's new peak and off-peak pricing took effect at 16:00 UTC on 16 August, and outside peak hours its V4-Pro model now costs half the peak rate. Per million tokens, input that hits the cache costs $0.044 at peak and $0.022 off-peak. Input that misses the cache costs $1.32 and $0.66, and output costs $3.96 and $1.98. On the usual rule of thumb of three-quarters of a word per token, a million tokens is about 750,000 words.
The cache rate applies when part of a request repeats text the service has processed recently, such as a long set of standing instructions sent with every query.
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. In UK summer time that is 02:00 to 05:00 and 07:00 to 11:00. Everything else, weekends included, is off-peak, so a UK business running batch work in the afternoon or evening pays the lower rate.
The price change followed V4-Pro leaving preview on 13 August with what DeepSeek described as major upgrades for agents, and a choice of reasoning effort for both V4-Pro and V4-Flash. On the web and in the app it appears as Expert Mode, and on the API the model names are unchanged. The API now supports OpenAI's Responses API natively, with Codex integration in mind. V4-Pro handles up to a million tokens of context and up to 384,000 tokens of output, though it cannot read images, which only Flash can.
DeepSeek's official documentation is at api-docs.deepseek.com. The similarly named deepseek.ai is not DeepSeek's site.
How Claude's text watermark works
On 14 August Anthropic published an explanation of the watermark it now puts into Claude's written output. Models launched from 2 August 2026 onward carry it, and older models are in a transition period, with the rollout running over the coming months. The 2 August date is the day the EU AI Act's transparency rules for AI-generated content began to apply.
The method works on individual word choices. At low-stakes points in a sentence, where several words would be equally valid, the model's choice is nudged by a secret key combined with the words that came before. Anthropic says readers will not be able to tell watermarked text from unwatermarked text. The technique is derived from Google DeepMind's SynthID-Text, which was published in Nature in 2024.
Anthropic says the watermark cannot separate text written by Claude from text heavily edited by Claude. It is weak on short samples and on factual passages that need exact terms, there is very little of it in code, and a full rewrite removes it. No accuracy figures have been published.
Detection runs through an API in private preview. For now it is limited to organisations covered by EU law, such as regulators, law enforcement, media, fact-checkers, researchers, educational institutions and civil society groups, and to enterprises that need verification. Anthropic plans to widen access, but outside those groups there is no tool yet for checking whether a document came from Claude.
A student hub in the Gemini app
Google's Gemini app gained a student hub on 19 August, where students can manage their classes and build personalised study notebooks from course materials such as syllabi. Study notebooks, which launched on desktop in June, are now on mobile as well.
Eligible college students get a year of Google AI Pro free in the US, or Google AI Plus free in more than 140 countries. The release note does not list the countries, so whether UK students qualify has to be checked on Google's eligibility page.
Before switching anything on
Browser use drives a browser that the customer hosts, so the choices about which sites an agent may visit and what it is allowed to submit sit with the organisation deploying it. For agents run on Managed Agents, the new domain lists limit what web search and web fetch can reach, and they should be set before an agent goes into use.
For AgentCore Payments, the per-session spending cap and expiry time are the settings to agree with the finance team before any agent is allowed to spend money. The regional list and the pricing page need checking at the same time.
For DeepSeek, moving scheduled work outside peak hours halves the cost of the same output at the current rates.
The computer use and browser use tools are listed in the Claude Platform release notes under 19 August.
SAASiQ - Intelligent Solutions for SaaS ©
