The Future Last Week: AI Gets Cheaper, Closer, and More Physical

The most useful AI story this week was not that one model topped another chart. It was that capable AI became cheaper to run, easier to place inside ordinary products, and harder to separate from the physical systems beneath it. The same week also supplied a reality check: power grids, scientific validation, employee trust, and the cost of a finished task do not improve just because a demo does.

This is the first issue of The Future Last Week, our Saturday read for people who have to make decisions rather than collect announcements. The mood here is optimistic. The evidence for that optimism is not a promise that every model will change your company. It is the growing number of practical choices: better models at several price points, more ways to reach them, and more examples of work completed with human review. We will say plainly when the week’s most exciting result does not yet apply to you.

The week in one minute

• OpenAI introduced GPT-6 Sol and Luna; Anthropic introduced Claude Opus 5.5. The business question is cost per accepted task, not who won a launch-day chart.
• Meta brought its Muse agent toward AI glasses, while Google introduced a live conversational avatar for enterprises. The interface is becoming less like a box you type into.
• Anthropic reported an AI-assisted finding of a previously uncharacterized enzyme system. It is an early research result, not a new medicine or a ready-to-buy business tool.
• Akamai announced an $11.6 billion, seven-year Anthropic infrastructure commitment. The International Energy Agency’s new grid report explains why electricity and network capacity belong in the AI story.
• Microsoft described a two-part AI bill: access per employee and metered agent work. That is a useful warning for anyone writing an AI budget.

1. The model race is becoming a value race

OpenAI added GPT-6 Sol and Luna to the Astra family on September 22. Its model guide positions Astra for the hardest work, Sol as a balance of capability and cost, and Luna for focused work at volume. Anthropic released Opus 5.5 the same day, saying it reaches the level of its earlier Fable 5.1 on most work and costs about 40% less to run than Opus 5 in typical token-billed use. These are provider descriptions; they are not independent proof of performance in your workflow.

The real development is choice. A year ago, teams often asked, “Can AI do this at all?” Now a more useful question is, “Which model completes this task accurately enough, at what total cost, and with how much repair?” A smaller model can be the better buy if it handles a repeatable classification or first draft. A stronger model may be worth its cost when a failed result creates expensive rework. Model price alone misses the reviewer’s time and any downstream correction.

Does this apply to you?

Yes if you already run recurring AI work or are about to buy a large seat or usage commitment. Re-run 20 real tasks on your current choice and one alternative. Count accepted outputs, reviewer minutes, corrections, and the full bill. Probably not yet if your team is still trying to define its first useful task; a new model name will not supply a missing process.

Hype versus real

Real: strong models are being offered at more than one price and speed. Hype: a vendor benchmark or a viral one-shot demo settles which model should run your finance, sales, or coding workflow. In one Claude user discussion, a developer called Opus 5.5 “crazy fast” at finding interface bugs, then admitted it might be a “day-one honeymoon phase.” That is useful firsthand feedback, not a representative test.

2. Zuckerberg’s glasses push is about access to an agent, not just eyewear

At Meta Connect, Meta said its Muse personal agent is coming to AI glasses in the coming months. The company also announced more connectors, including shopping and work services, and an email address for Muse. The examples matter more than the stagecraft: a person might point at a product or a notice and ask the agent to take the next step without describing the scene in text. Meta also announced new audio glasses and virtual-reality glasses. Several of the Muse-on-glasses experiences are future availability, not something every buyer can use today.

Google took a different route toward a similar question with Gemini 3.8 Live with Live Avatar, which it says is available in Gemini Enterprise for live voice-and-video conversations. Google’s examples include service and walkthroughs. A visually present assistant may help in a guided task, but presence is not resolution. If it cannot authenticate the right person, pull the right record, and complete the action, the moving face is decoration.

Does this apply to you?

It may if your work happens away from a desk: field service, inspections, retail assistance, accessibility, or hands-busy training. Start with one repeatable interaction where seeing the environment changes the answer. For most office teams, buying glasses this quarter is not the lesson. The lesson is to map where speaking, seeing, and acting would remove an actual handoff.

Hype versus real

Real: the interface to AI is broadening beyond the keyboard. Hype: a successful product demo proves people want a camera-facing agent throughout the day. Adoption will depend on comfort, privacy, clear permission for each action, and whether the service finishes the job. Meta’s Connect announcement is the source for what it plans to ship; customer behavior is a separate question.

3. The science result is exciting precisely because its limits are clear

Anthropic’s new life-sciences group reported that Claude agents searched large DNA databases and surfaced a previously uncharacterized enzyme system with repeat patterns reminiscent of CRISPR. Human scientists then examined the candidate in the lab. The company says the system’s function is still unknown and has released a preprint for further scrutiny. That is a meaningful early result: AI helped direct experts toward an anomaly worth testing. It is not a gene-editing treatment, a validated clinical tool, or a reason for a non-science business to buy a specialist model.

Feng Zhang of MIT and the Broad Institute, quoted in Anthropic’s research account, called the finding “genuinely intriguing and merits further investigation.” That is the right temperature. The hope is not that AI replaces a laboratory. It is that it can search a vast space of possible leads, document why one stands out, and let scarce lab time go to better candidates. The bottleneck then moves to experiments and validation.

Does this apply to you?

Directly, probably only if you run research in biology or another data-rich science. Indirectly, the pattern travels: let AI generate and rank hypotheses, but budget the human and physical capacity to test them. Our earlier analysis of the science-validation bottleneck explains why faster idea generation does not automatically mean faster discoveries.

4. The AI boom has a concrete floor plan and an electricity bill

Akamai said Anthropic committed $11.6 billion over seven years for distributed cloud infrastructure, with a possible additional expansion. The announcement concerns central processing unit workloads and software as well as infrastructure; it is not simply a purchase of the newest graphics chips. The dollar figure is a contractual commitment described by the companies, not cash already spent or revenue recognized this week.

The International Energy Agency’s report released September 21 describes a two-way problem: data centers add demand and can worsen grid congestion, while better digital tools and AI can help operators use existing lines more safely and efficiently. The U.S. Energy Information Administration separately expects record electricity generation in 2026 and 2027, citing data centers and manufacturing among the sources of stronger demand. Neither report says that every proposed data center will be built on schedule, nor that AI alone explains national power demand.

The bullish case is that infrastructure investment can turn model progress into widely available services and that smarter grids can recover capacity from existing systems. The sober case is that siting, financing, local power prices, water, construction, and community consent are not software bugs. A promising application still needs reliable compute somewhere. A resident asking who pays for a new substation is raising an operating question, not failing to appreciate innovation.

Does this apply to you?

If you buy AI services, ask about uptime, data location, and price stability before obsessing over the exact chip. If you build facilities or sell power-intensive services, ask the utility about interconnection timelines and who funds upgrades. For everyone else, the practical conclusion is modest: model availability and price are partly physical-infrastructure questions. Treat aggressive delivery promises accordingly.

5. Work software is changing its bill before every company has changed its work

Microsoft’s Jared Spataro framed the new enterprise buying problem unusually clearly this week: a flat subscription gives people access, while longer-running agent work may be billed by use. His line that “no model is best at everything” is also a sensible antidote to vendor loyalty. Microsoft’s own answer is a platform that can use several models and meter agent operations through Copilot Credits. This is a vendor position, but the budget distinction is real regardless of whose product you choose.

An employee seat can be lightly used or extremely valuable. A metered agent can generate thousands of actions and still fail to produce one accepted outcome. Keep the two bills separate. The useful measure is cost per approved report, resolved case, shipped change, or other finished unit of work. Our guide to Copilot seat and usage charges shows how to make that comparison without counting released hours as automatic cash savings.

Does this apply to you?

Yes if you are setting a 2027 software budget or piloting agents that can take several actions. Set spending limits, define the accepted output, and keep the human repair work in the measurement. If you cannot say what successful work looks like, you cannot yet judge whether a lower model price or a bigger seat bundle is a bargain.

6. Fear is a real adoption signal, not a reason to stop learning

Pew Research Center published a 37-country survey shortly before this issue’s news window. In 34 of the 37 countries, respondents were more likely to expect AI to reduce jobs than to create them over the next 20 years. Pew surveyed 42,151 adults across 36 countries from February to May 2026 and used separate U.S. surveys; the result measures expectations, not observed job losses or a forecast of how many positions will disappear. In the U.S., concern rose most among adults ages 18 to 34 compared with Pew’s earlier measure.

I remain bullish because the work is not a fixed pile of tasks. Better tools can make more analysis, support, design, and experimentation affordable, while creating new work around review, implementation, and customer service. But optimism that ignores the person asked to train the system that may change their role is thin. Leaders need to say what work changes, who gets training, who reviews AI output, and how employees share in the gains. Those choices—not a reassuring slogan—will determine whether fear gives way to useful adoption.

Does this apply to you?

Yes if you manage people. Ask a team to name one draining task AI might improve and one concern they would want answered before trying it. Give them time to learn, publish a review rule, and report the result in terms of workload and quality, not only headcount. Pew’s full survey and method is worth reading before telling staff that anxiety is merely a communications problem.

One smaller signal that could matter later

Amazon Web Services published a September 21 guide to xAI’s Grok 4.6 on Bedrock. AWS says the model launched there August 18, so this was an integration explanation, not a new model launch this week. For teams already committed to Bedrock, the significance is another model to test without moving the entire application stack. It does not mean Grok is the best choice for your workload. Elon Musk’s company matters here as a supplier in a growing distribution market, not because every launch needs another personality contest.

There was also a policy signal. On September 23, California named the expert group that will advise on independent oversight and a possible emergency stop mechanism for advanced models. This is a recommendation process following an earlier order, not a new requirement for every company using AI. If you buy or deploy powerful agents, it is worth watching how independent evaluation becomes operational; most teams do not need to rewrite their controls this week.

What people are saying—and what it tells us

The week’s reactions are more useful when read side by side. A researcher called the enzyme result intriguing but asked for more investigation. A Microsoft executive argued that no single model wins every job. A Claude user praised the speed of a new model while worrying about a launch-day honeymoon. The International Energy Agency urged “making better use of existing assets” as well as building new infrastructure. These are different standards of proof: a scientific finding needs replication, an enterprise platform needs completed work at a known cost, a user impression needs a repeatable test, and an energy plan needs physical capacity. The common thread is encouraging: people are arguing about how to use AI well, not only whether it can produce a convincing paragraph.

Five moves for Monday

1. Choose one recurring task and replay the same 20 examples on your current model and one alternative. Keep the review bar fixed; compare accepted work, repair time, and total cost.
2. If someone proposes a voice, video, or glasses experience, write down the exact action it must complete. Test a plain-text version first so visual novelty does not hide a weak process.
3. For research-heavy work, separate ideas generated from ideas validated. Identify the person, lab, customer, or test that can actually confirm the result.
4. Put your AI bills into two columns: access for people and metered work done by software. Tie both to a finished business outcome.
5. Ask employees where AI would relieve a burden and where it would create one. Commit to a training and review plan before announcing productivity targets.

The signal to watch next week: whether the new models and interfaces produce independent, repeatable gains in ordinary work—and whether providers can deliver them at a price and reliability that survives broad use. That is a more bullish question than “Did the demo look magical?” because it can be answered.

Source notes

This issue covers announcements made mainly September 20–25, 2026, with the September 17 Pew survey and the September 9 U.S. electricity forecast included as clearly dated context. Product claims are attributed to their providers; the enzyme result is early research; the forum comment is anecdotal. We linked original announcements and research at the claim. The illustrations are interpretive artwork, not documentary images.