What does your AI budget actually buy?

An AI subscription can give you a quick summary or keep several agents busy on a substantial assignment. The monthly fee does not tell you how many usable results you will get. This week's announcements from Anthropic and Google make that distinction harder to ignore: more work can consume the same usage allowance faster, and a tool included at purchase can leave the bundle at renewal.

Hammer's thesis for week 38 is that AI budgets need to account for completed assignments, not just licensed seats. If that is right, the next purchasing decision should start with the work you want done. It might justify a more expensive solution. It might also reveal that buying more seats will not solve the problem.

The English podcast above draws on nine Hammer research reports dated September 14–20, 2026. This companion article examines the economics behind the week's product announcements.

Claude puts more work inside a shared allowance

Anthropic announced its new combined chat and Cowork experience on September 16. A conversation can handle quick questions as well as longer work on documents, spreadsheets, and presentations. The rollout is gradual, starting with Pro and Max. Anthropic explicitly says that longer tasks involving web searches, code execution, or file creation generally consume more of the plan's usage allowance than a quick question.

A redesigned Projects beta followed in Claude Code on September 17. A coordinator divides work among parallel threads, each of which is a full cloud session. An AI agent here is a system that can work through several steps and use tools, rather than just write a reply. Anthropic warns that projects can reach usage limits faster.

That is a useful capability. Different parts of a larger assignment can be tackled simultaneously. But five working threads do not mean five new subscriptions with separate allowances. Nor does the announcement establish a fixed relationship between thread count and completed assignments.

The Projects beta opened to selected Pro and Max users with cloud sessions and no existing projects on web or desktop. It is not a general offer for every business account.

Sources: Anthropic's guide to the combined experience and Projects redesigned.

An illustrative calculation: SEK 25 or SEK 100 per assignment

Assume an entirely imaginary monthly AI subscription budget of SEK 2,000. This is not a vendor price. Compare two assumptions about how many equivalent assignments are completed and usable:

• At 80 completed assignments, the fee allocates to SEK 25 per assignment.
• At 20 completed assignments, the same fee allocates to SEK 100 per assignment.

The calculation divides the monthly fee by completed assignments. It says nothing about the capacity Claude, Gemini, or another provider actually offers. It does not establish that parallel agents cause the difference. Review time, integrations, and any additional usage charges sit outside this example.

For a consultancy, the question gets more interesting when the assignments differ. Twenty substantial client analyses might be worth more than eighty short meeting summaries. You cannot then compare the counts as if the results were equivalent. Chasing the lowest cost per reply can steer work toward what is easy to produce rather than what the client needs.

Google changes what comes with the renewal

Google's September 17 announcement concerns a different part of the calculation. New Gemini Enterprise Standard and Plus subscriptions, and existing subscriptions renewed online, no longer include Gemini Code Assist features. Existing included access continues until the subscription term ends.

This concerns the named Gemini Enterprise product. It is not an announcement that every Google Workspace subscription is getting more expensive or that every Code Assist installation stops working immediately. Google points to Antigravity as an alternative for developer tools.

The purchasing consequence is still concrete: last year's list of included tools is not enough to budget for next year. The same number of seats need not buy the same bundle. The cost effect depends on the agreement and which capabilities you need to retain.

Source: Gemini Enterprise release notes for September 17.

More demanding AI work can be worth paying for

OpenAI introduced Astra for Law on September 17: GPT-6 Astra with legal instructions and a search index for U.S. legal authorities. Initial access is for selected law firms through Trusted Access; API availability is described as coming soon. It is an example of vendors packaging AI for more specialized work, not proof of profitability or reliability on Swedish legal matters.

Grok Build gained cross-session memory in SpaceX/xAI's September 16 announcement. The tool saves project facts and decisions as notes and retrieves them for later work. That might reduce repeated explanations. We do not, however, have measured evidence here that it lowers total costs.

Mistral's Cloudera partnership, included in this week's research material, was announced on September 10. It is background, not a week 38 launch. The partnership concerns AI in enterprises' own deployment environments and adaptation to their data. Its economics also involve infrastructure and operations, well beyond a simple comparison of monthly fees.

Sources: OpenAI: Astra for Law, SpaceXAI: Memory in Grok Build, and Mistral and Cloudera.

Subscriptions remain a reasonable way to buy AI. They provide an understandable base cost and can suit work that varies from day to day. The thesis is narrower: supplement the seat budget with what actually gets done. Pick one recurring assignment and track how many results you can use, how much of your time they require, and what value they provide. That is useful material for Mindset Forge if you want to discuss which AI work is worth investing in.

The podcast is an AI-generated masterclass created with NotebookLM from Hammer's in-depth, AI-generated research into AI providers' updates and features. The material covers OpenAI, Anthropic Claude, Google Gemini/DeepMind, SpaceX/xAI Grok, and Mistral. Research reports are dated within the completed week; some recover earlier announcements under their original dates. This article draws on the reports, NotebookLM synthesis, and checked primary sources, not an audio transcript.