Reading GPT-6 Astra Through a Class Lens

Selling the Work Itself: Reading GPT-6 Astra Through a Class Lens

On 3 September 2026, OpenAI began rolling out GPT-6 Astra. On the company's own announcement, the model can use a computer end to end: it fills forms, updates CRM records, manages calendars, installs and tests software; produces documents, presentations, and spreadsheets to template; does legal analysis and financial modeling. On desktop tasks it reaches a result in roughly 47 percent less time than the previous model. And it is the first model to cross the "Critical" cybersecurity threshold in OpenAI's own "Preparedness Framework": it scored 100 percent on an exploitation benchmark and found two zero-day vulnerabilities during testing. Price: $10 per million input tokens, $50 for output. Off by default on enterprise accounts; an administrator turns it on.

Dear Young Comrades,

This essay was not written to debate how "smart" an artificial-intelligence model is. Anyone can read the benchmark tables; the industry has been arguing over them for days. Our question is different: what is being sold with this product, from whom was it taken, to whom is it sold, and who will pay the cost? These four questions turn every technology launch into a class relation; and Astra is one of the rare launches in which the answers to those four questions are written so plainly.

We built the essay in seven parts. First the change in the nature of the product, then the layer of labor it targets, the class content of the price, the property meaning of the "Critical" cyber threshold, the problem of control in the labor process, monopoly concentration, and Turkey. Then a comparative table, a list of common errors, and concrete proposals.

I. What is sold is no longer an answer, but work

In earlier versions what was sold was a text: you asked a question and received an answer. However good the answer, you were still the one doing the work; the model was an input to the work. With Astra, what is sold has changed. OpenAI's own marketing language says it "handles the boring tasks for you"; industry analyses are clearer: the metric is no longer "cost per token" but "cost per completed job." The product's output is not a chat transcript; it is a filled form, an updated customer record, a presentation to template, a placed printed circuit board, a completed code change.

The Marxist name for this shift is simple. What was sold before was a tool that raised the productivity of labor-power; what is sold now is the product of labor-power itself. And where does that product come from? From the densification of texts, code, documents, drawings, and discussions produced by the labor — mostly unpaid, partly very cheap — of hundreds of millions of people. In our Karaburun paper we called this the expropriation of the general intellect — starting from Marx's concept of the general intellect in the Grundrisse: society's common production of knowledge crystallizes inside a model, and the owner of the crystal sells it as their own "invention." Astra is the most developed form of this expropriation; because it now densifies not only knowledge but the form of turning knowledge into work — how a paralegal reads a document, how an accountant builds a spreadsheet, how a programmer hunts for a bug.

Dead labor is substituting for living labor. And to the extent that it substitutes, it erodes living labor's bargaining power. The sentence "it frees engineers to invent" — which appears in the announcement with a circuit-board example — is the classic formula capital uses in every wave of machinery. The machine does not relieve the worker of a burden; it either throws the worker out of work or transfers the time saved by the machine to capital in the form of unpaid time. The power loom did not free the weaver; it either left them unemployed or bound them to the loom's speed. This law holds in information work as well; only the appearance of the loom has changed.

II. The target: the "routine" layer of the white-collar proletariat

The shortest way to see whom a product targets is to look at whom it puts in its window. The partners named in Astra's launch are these: Harvey (legal tech), Jane Street (finance), Cognition's Devin and Lovable (software production), Higgsfield (visual production). So the target field is the wage-dependent knowledge worker who, at the end of the twentieth century, was called "middle class" but is separated from the means of production: the paralegal, the financial analyst, the junior developer, office operations, data entry, content production.

To see that this targeting is no accident, it is enough to look at how the model's "capabilities" are described: better judgment under incomplete instructions, holding context as a task evolves, asking a question when needed and then continuing without waiting. These are not technical features; they are a description of coordination labor inside a team. That invisible work circulating between a manager and a team — "that's what I meant," "add this too," "use last week's template" — is being taken into the model. So it is not only executing labor that is swallowed, but the intermediate labor that keeps executing labor upright.

The result is not a homogeneous "upskilling" of the labor market, but its bifurcation. Two figures relayed by an industry newsletter summarize this: 205,000 layoffs in the first half of 2026 explicitly citing "AI"; against that, a 62 percent wage premium for workers with "AI skills." Inside the working class, a new labor aristocracy and a growing reserve army of labor are produced at once. The mechanism we described in our Unemployment Dossier — technology producing not unemployment but the threat of unemployment — finds its textbook example here: Astra need not be used to fire people; the knowledge that it exists is enough pressure in wage bargaining.

There is also a door closing at the entrance. One enters a profession from the "junior" rung; the junior covers their salary with the work they do while learning. Exactly what Astra does best is the junior's work: routine documents, standard queries, first-draft code. When that rung closes, the profession loses not only today's workers but tomorrow's. We described this in our Computing Labor Dossier as "pulling out the bottom rung of the ladder"; now the tool that pulls out the ladder has a name.

III. Price is the formula for sharing surplus-value

Look at the price list: a million input tokens for $10, a million output tokens for $50; "fast mode" at twice the price. That is roughly a two-and-a-half-fold rise per token against the previous model. But industry analyses report that cost per job — because the model finishes the work with fewer tokens — has fallen by around 43 percent on professional benchmarks.

Put those two figures side by side and what appears is not a pricing policy but a sharing formula. When an employer has Astra do a worker's job, they obtain a saving on labor cost. OpenAI cuts a portion of that saving for itself as rent — by raising the token price — and leaves enough for the employer to be persuaded. The worker's share is, by definition, zero; because the worker is not a party to this bargain but its object.

The product portfolio is class-structured too. The triple structure relayed by analysts — one model for cheap high-volume work, another for routine production, Astra for the "hard 10 percent" — builds a hierarchy inside capital itself. Astra is off by default on enterprise accounts; consumer subscriptions have a weekly quota, and when the quota runs out one drops to a cheaper model. One analyst's summary is bare: "keep the consumer number sacred, harvest the API and the overage." So the strongest automation tool is a "luxury SKU" for the small firm and the individual, and infrastructure for big capital. Access to automation is itself stratified by class: those who use the machine most efficiently are already the largest.

IV. The "Critical" threshold: ownership of a dangerous means of production

The most discussed side of the launch is the model's capacity for cyber attack. In OpenAI's own words, Astra can find and exploit previously unknown security vulnerabilities; it is the first model to cross the "Critical" level in the company's own risk classification. This capacity will be opened, through a program called "Daybreak," to "approved defenders" — enterprise security firms, actors linked to the state — and closed to the ordinary user.

Here one must see the dual function of the discourse of "alignment" and "safety." The public-safety concern is real and must not be minimized; free distribution of this capacity is genuinely dangerous. But the same discourse is at the same time an access-control technology; that is, a gatekeeping that consolidates property. One industry analyst put it in a single sentence: "We own the dangerous model and we will not give it to your intern." Who counts as a defender and who as an attacker is decided by a private company and its relation to the state. The capacity itself is social — distilled from millions of security studies, open-source code, forum debates — but its ownership sits with a single company, and that company has also appropriated the authority to decide who may use the capacity.

This is the software form of what Lenin described for the monopoly stage: the scissors between the socialization of production and the privatization of property. The scissors open as capacity grows. And a second point: critics note that Astra "behaves better but is worse to audit." The model's reasoning is less transparent than earlier ones; enterprise auditability is limited. So both worker and employer are forced to trust a black box; transparency gathers upward, toward the company's own monitoring systems. That is how the architecture of the digital panopticon works: the watcher is not seen; the watched cannot watch.

V. Control in the labor process: no one to hold to account, someone who works

The industry press's warning to corporate managers says something else for us. Agents appear in enterprise systems as a "service account"; there is no record of which model version initiated which action. For managers this is a "governance" problem. For labor it means this: responsibility blurs, but responsibility never really blurs; it flows downward.

When an agent wrongly updates a customer record, misreads a contract, or errs in a financial model, who pays the cost? Not the shareholder; the worker the company has tasked with "supervising the agent." That worker will be obliged to catch the error the agent did not make, without being given time enough to redo the agent's work from scratch. This is a further stage of the deskilling process Braverman described in Labor and Monopoly Capital. The worker first lost their craft, then their decision; now they are losing their record as well. On the register of the work done, their own name is not written but a service account's; yet they are still the one held to account for the work.

The next step of what we call mental Taylorism is here too. Taylor broke the worker's job into parts and timed each part. Mental Taylorism broke the knowledge worker's mental work into parts and measured it by the platform. At the stage Astra represents, the parts are taken from the worker and given to the machine; what remains for the worker is to check whether the parts have joined correctly — that is, the most boring, most responsibility-laden, and least instructive part of the work. The announcement's sentence "it handles the boring tasks for you" turns into its opposite: the machine takes the creative part; the boring supervisory part stays with the worker.

VI. Monopoly concentration: competition exists, choice does not

The week Astra appeared is instructive for the map of monopoly capital. In the same days Anthropic, Google, and Meta also announced new models; on independent benchmarks Astra leads in some domains and trails in others. The same week Nvidia bought the open-source model repository Hugging Face for $12.9 billion. Astra's distribution runs through Microsoft Azure and Amazon Bedrock; so outside three or four hyperscale clouds there is no access to this tool.

In this picture "competition" is real, but it is competition inside the oligopoly. One can choose among models; one cannot step outside the monopoly of cloud, chip, and distribution channel. In a week when "open source" too has entered capital's gravitational field, the option for workers, small producers, and Southern economies is only which monopoly's tenant to become. Let us repeat here the answer we gave earlier to the techno-feudalism debate: this is not a "new feudalism" but monopoly capitalism equipped with digital means of production. Accumulation, competition, the compulsion to grow, entanglement with the state — all of it is in place; only the name of the means of production has changed.

VII. Turkey: a dollar-priced machine, lira-priced labor

Astra's target list — document production, customer-record management, legal and financial analysis, software — touches directly one of the few growing employment fields in Turkey: export-oriented computing and office services. There is a specific asymmetry here: the machine is priced in dollars, the worker is paid in lira. On the one hand this makes the tool expensive for small firms in Turkey; on the other it puts a knowledge worker paid in TL into competition with a dollar-priced machine. Every time the exchange rate moves, the terms of that competition are rewritten against the worker.

The most likely short-term scenario is not mass unemployment but intensification: the same work done by fewer people, for longer hours, in "agent-supervising" positions; the closing of junior rungs at entry; further erosion of bargaining power in the non-union white-collar segment. In our Computing Labor Dossier we spoke of the "Branch No. 10" barrier: the difficulty of organizing for the computing worker who has no branch of their own. Astra is a tool that raises the cost of that lack of organization. The unorganized worker bargains with the machine alone; and bargaining with the machine alone is not bargaining.

Comparative table: what the launch says, what we see

The launch's claim	Historical-materialist assessment
"The smartest and most aligned model."	Intelligence is distilled from social labor; alignment is aligned with the property relation. Neither is a property of the company, but of the society that produced them and the property that constrains them.
"It handles the boring tasks for you."	Creative and instructive parts go to the machine; the responsibility-laden supervisory part stays with the worker. The boring does not leave; it changes hands.
"It frees engineers to invent."	The formula of every wave of machinery. Time gained is transferred to capital as unpaid time, or the worker is made unemployed.
"Cost per job fell 43 percent."	The saving from labor cost is shared between company rent and employer profit. The worker is not a party to the bargain but its object.
"Off by default in enterprise; weekly quota."	Access to automation is stratified by class: the strongest tool is big capital's infrastructure, the small one's luxury.
"Critical cyber capacity only for approved defenders."	A socially produced capacity is privately appropriated; the owner also decides who may use it. Socialization of production, privatization of property.
"Better aligned, less drift."	"Behaves better, worse to audit." Transparency gathers upward; the watcher is not seen.
"Agents integrate seamlessly into the workflow."	On the register, the service account's name; on the bill, the worker's. Responsibility does not blur; it flows down.
"A competitive market; many models."	Competition inside the oligopoly, no option outside it. The cloud, chip, and distribution monopoly stays fixed.

Reduce the table to one sentence: Astra's "intelligence" is society's, its "alignment" is the owner's; and the distance between the two is the present field of class struggle.

Three common errors

First: sliding into hostility to technology. Some of the work Astra does is genuinely boring, genuinely mechanical, and genuinely unfit for the human mind. Refusing to hand those tasks to the machine is to fall into the Luddite trap. The problem is not that the machine does the work; it is that the machine's owner alone takes the gain born of handing the work to the machine. The left positions itself not against technology but on the ownership of technology and the sharing of its gains.

Second: entering the "safety" debate inside the company's frame. Debating Astra's cyber capacity on the axis "dangerous or not" skips the question of who decides. It is dangerous; but that a private company decides to whom a dangerous capacity is opened is a separate danger. Our question is not "how safe" but "who decides."

Third: mistaking the advice "acquire AI skills" for a solution. The 62 percent wage premium is real; but that premium is the price of producing a labor aristocracy inside the working class. Individual skill is individual salvation; and individual salvation is paid for with the erosion of collective bargaining power. Skills must be acquired; but separating skill from organization is to choose to bargain with the machine alone.

Concrete proposals

None of these alone changes the order. But they leave you better equipped for the next launch, the next "productivity" presentation, the next wave of layoffs.

1. Ask every launch four questions. What is being sold? From whom was it taken? To whom is it sold? Who pays the cost? These four questions turn the brightest announcement into a class map. For Astra the answers are in this essay; for the next model you will write them.

2. Make "agent" rollout at the workplace a collective-bargaining issue. An employer's inserting an autonomous agent into the workflow is no different from a machine investment, and the effect of machine investment on labor is the oldest bargaining topic in the union tradition. Which jobs are being handed over, who holds supervisory responsibility, who is held to account when there is an error, how is the time gained to be shared — these are not technical questions but contract questions.

3. Demand transparency in "AI-justified" layoffs. The figure of 205,000 is the total for companies that explicitly name this reason; those that do not name it are far more. Mandatory disclosure of automation as a ground for dismissal, its taxation, and a re-employment obligation are the first defensive line of this wave.

4. Bring the shortening of working time back onto the agenda. If Astra cuts time per job by 47 percent, whose that time goes to is a political question. Eight hours was once a utopia. Thirty hours is today's concrete utopia; and its technical justification is written in capital's own promotional brochure.

5. Demand an audit trail. Without a record of which model version, under whose instruction, and at which moment every action an agent took was done, responsibility cannot be loaded onto the worker. Making visible the responsibility hidden behind the name "service account" is not only a governance demand but a labor right.

6. Keep the computing branch and its union on the agenda. The workers Astra targets are workers in Turkey who have no branch of their own. The tasks in our Computing Labor Dossier have become not less but more urgent with this launch. The unorganized worker bargains with the machine alone.

7. State common ownership of model and data as a demand. This demand will not be realized today; but if it is not spoken, it will not be realized tomorrow either. Astra's "intelligence" is society's; the demand that what society produces be society's property is not a utopia but a claim of ownership. The concept of the knowledge commons is exactly the name of that.

Dear Young Comrades,

On 3 September OpenAI released its "smartest and most aligned" model. The source of the intelligence is the unpaid labor of the millions who trained that model. The source of the alignment is the property relation of the model's owner. The distance between the two is the subject of this essay and the field of this period's class struggle.

Astra is not a chatbot; it is a machine that sells the work itself. Its target is white-collar wage-earners who long took themselves for the middle class. Its price is the formula for sharing the saving from labor cost between company and employer. The "Critical" threshold is the announcement of private ownership of a socially produced capacity. The architecture of control moves responsibility downward and transparency upward. And none of this is a conspiracy; it is the ordinary working of monopoly capitalism at its present stage.

The problem is not technology itself. The problem is the form of property that sets the product of living labor against living labor. That is why the answer, too, is not technical but political: contract, transparency, hours, record, organization, common ownership. And the precondition of all of them is that the white-collar segment this wave targets finally recognize itself as a class.

Young comrade, when at the next launch they tell you "it can do this job now too," the question to ask is clear: whose job was that, and to whom does the gain go? If the answer is "not to you," that launch is not your launch; but that job is still your job — and it can be reclaimed.

Comradely, Knowledge Commons

Sources

Launch and technical data

Cybersecurity threshold and the audit problem

Business model, labor market, and monopoly concentration

Marxist framework

Karl Marx, Capital vol. I, ch. 15: "Machinery and Large-Scale Industry" — the machine's effect on the worker and the ideology of "freeing the worker"
Karl Marx, Grundrisse, "Fragment on Machines" — the concept of the general intellect
V. I. Lenin, Imperialism, the Highest Stage of Capitalism (1916) — socialization of production and privatization of property
Harry Braverman, Labor and Monopoly Capital (1974) — deskilling and the conception–execution split
Knowledge Commons, "The Revolt of Crystallized Labor: The Expropriation of the General Intellect in AI Models, Mental Taylorism, and the Digital Panopticon", 20th Karaburun Science Congress, September 2026
Knowledge Commons, "The Computing Worker's Handbook", September 2026 — the computing branch and organizing
Knowledge Commons, "Unemployment Dossier", August–September 2026 — the reserve army of labor and technology
Knowledge Commons, "Technology in Place of Poetry: From Futurism to Techno-Fascism", 6 September 2026 — reply to the techno-feudalism debate