In this issue
OpenAI shipped GPT-6 Astra on Thursday and led the announcement with one capability. The model operates a computer the way a person does. It fills in forms, updates CRM records, works spreadsheets, and builds a website from an empty folder. Greg Brockman said "welcome to the AGI era." I would park the AGI argument and read the rest of the week instead.
Anthropic cut Claude's price by roughly a quarter two days earlier, except it did not cut the price. The headline rate is identical to Fable 5. The whole discount sits on cache reads, which matters enormously if your agents reuse context and not at all if they do not. That is the pattern in this issue: the headline and the invoice disagree. Salesforce made Claude its default reasoning model, Cowork grew its own browser, Brussels started handing out fines, and a third of companies stopped buying software. Let's get into it.
Topics of the day:
GPT-6 Astra can operate your computer
Claude's 25% price cut is cache only
Claude lands in Salesforce and gets a browser
Curated reads on GTM agents and loops
The EU can now fine model providers
A third of firms skipped a SaaS buy
The Shortlist: cheaper Gemini, Astra's specs, Europe's model
OpenAI ships a model that operates your computer
What's happening: OpenAI started rolling out GPT-6 Astra on Thursday, and the headline skill is computer use. It fills online forms, updates CRM records, manipulates spreadsheets, works in Power BI, and creates websites from scratch. Greg Brockman said Astra "can really do anything a human can do with a computer."
In practice:
Price the job you hand over, because Astra runs $10 per million input and $50 per million output, and fast mode doubles both.
Hand over one repetitive screen job first, form filling and CRM updates are where OpenAI pointed the demos.
Check the work before you wire it to anything live, Astra scores 72.6% on OSWorld 2.0, so roughly one task in four still misses.
Note the cyber gate, OpenAI rated Astra Critical for cyber capability and gave Daybreak defenders first access.
Bottom line: The question stopped being which model is smartest and became which screen job you hand over first.
Claude's 25% price cut applies to cache reads only
What's happening: Anthropic released Claude Fable 5.1 on September 1 and said it costs around 25% less than Fable 5. The headline rate did not move. Anthropic's own pricing page lists $10 per million input tokens and $50 per million output for both models, and identical batch rates of $5 and $25. What changed is the cache read price, down from $1 to $0.25 per million.
Two prices did genuinely move this month. OpenAI cut GPT-5.6 Sol on August 21, from $5 to $4 in and $30 to $20 out, promotional through at least November 21. And Anthropic cancelled the September increase for Sonnet 5, so $2 and $10 is now its standard rate. Anthropic also reports 60% fewer cybersecurity false positives and 85% fewer refusals on benign biology questions.
In practice:
Check whether your workload reuses context, the saving lands on cache hits only, and fresh tokens cost what they did last week.
Switch prompt caching on before you count the discount, Fable 5.1 reads cache at 2.5% of base input, every other Claude model reads at 10%.
Expect the top of the range only on agent work, Anthropic puts typical savings near 25% and agentic savings near 45%.
Compare the tier below before you commit, Claude Opus 5 runs $5 in and $25 out, half of Fable 5.1 on both sides.
Bottom line: The discount lands on repeated context, so check your cache hit rate before you budget the saving.
Claude becomes the default reasoning model inside Salesforce
What's happening: Salesforce and Anthropic announced Claudeforce on August 26, making Claude the default reasoning model across Agentforce, Agent Builder, Slackbot and Slack AI. The other half of the deal is Salesforce in Claude, a plugin carrying 37 prebuilt sales skills for pipeline review, deal health and meeting prep.
In practice:
Wait for the September open beta before you plan around it, Salesforce in Claude is pilot-only today.
Check what Slackbot answers now, because Claude became the default model across all of Slack.
Map which of the 37 skills replace a report you build by hand, pipeline review and deal health lead the list.
Set write permissions first, because these skills take governed action on the records they touch.
Bottom line: Your CRM is becoming a place where an agent does the work.
Claude Cowork grows its own browser
What's happening: Anthropic built a browser into Claude Cowork on the desktop app. When a task needs a website, a browser opens in the side panel and Claude reads the page, clicks and types. There is no extension to install, and nothing is shared from your own browser unless you choose it.
In practice:
Start with the web chore you repeat, checking a portal, pulling the same report, filling one form every Monday.
Keep it away from your logged-in sessions at first, the built-in browser shares nothing from your own unless you allow it.
Check your plan, it rolls out on Pro, Max and Team, and on Enterprise where an owner switched it on.
Read it next to Astra, both labs now sell the same product, an agent that works a screen for you.
Bottom line: Your agent now works in a browser of its own.
Read Later
treg - One token that hands a coding agent 2,630 marketing endpoints across 47 providers, priced per result, such as $0.006 per Semrush keyword lookup.
loopy - A library of practical agent loops plus an installable skill that finds, audits and adapts repeatable workflows for Claude Code, Codex and Cursor.
KAMI - Open source go-to-market agents for early-stage founders that plan a campaign, delegate to specialists, and publish only after you approve.
Origin Code Hosting - Cursor now hosts your repos and pull requests next to your agents, with two-way GitHub sync, in early beta on every paid plan.
The EU can now fine the labs
What's happening: The AI Act's enforcement powers switched on on 2 August 2026. The Commission's AI Office can request documentation, run its own model evaluations, demand corrective measures, and restrict a model's availability in the EU. Penalties for general-purpose AI obligations reach €15 million or 3% of worldwide annual turnover, whichever is higher.
In practice:
Note who this lands on, the general-purpose AI fines target the model providers themselves.
Expect vendor paperwork to improve, providers now face evaluations and information requests they cannot quietly ignore.
Keep your own Article 50 duties on the list, telling people they are talking to an AI, and labelling AI-generated content.
Diarise 2 December 2026, providers of gen-AI models released before 2 August 2026 must watermark output by then.
Bottom line: Your AI vendors now answer to a regulator that can fine them, and that changes what they will put in writing.
A third of companies skipped a software purchase
What's happening: McKinsey surveyed 1,719 business leaders for its State of AI 2026 report. It found that 32% decided against buying at least one software product or feature, because they could build it internally with agentic coding tools. Among high performers, the 6% who credit AI with at least 5% of EBIT, nearly half skipped a purchase.
In practice:
Read the number as a decision to skip a purchase, because nobody was asked whether the replacement passed an audit.
Run the build-versus-buy math at your next renewal, technology firms hit 41% and healthcare payers and providers 39%.
Price the run cost as well, because an internal tool still needs an owner, a backlog and someone on call.
Start with the small feature you pay a vendor for and barely use, that is where the 32% is coming from.
Bottom line: Building software got cheap. Owning it still costs what it always did.
The Shortlist
Google shipped Gemini 3.7 Flash at half the per-million-token cost of 3.6 Flash, aimed at coding and agents, alongside Gemini 3.5 Transcribe for real-time speech to text.
OpenAI published Astra's API specs: a 1,050,000-token context window, 128,000 max output tokens, $1 per million cached input, and computer use plus web search as callable tools.
Anthropic kept the sharper twin behind a gate, Mythos 5.1 is the same model with lighter safeguards, and only registered cybersecurity and life-sciences partners can use it.
Multiverse Computing released Quasar 438B, which scores 43 on the Artificial Analysis index, the highest of any European model, and runs through the CompactifAI API.
This newsletter is where I (Kwadwo) share products, articles, and links that I find useful and interesting, mostly around AI. I focus on tools and solutions that bring real value to people in everyday jobs, not just tech insiders.
