Signal #78: OpenAI sold the Analyst, Anthropic Ran the wage maths

I built FRAKTON in Prishtina on one piece of arithmetic. A strong engineer here cost a fraction of a strong engineer in Munich or Stockholm, and that gap paid for everything else, right up to the exit. Most of the software firms I know in this region still run on the same sum, and many still report growth as headcount.

This week that sum got hit from three sides. OpenAI put a model on sale against a job description. Anthropic published what happens to the people holding that job if the models keep improving. And Romania's mid-tier outsourcers showed, in their own filings, what a services business looks like when output holds and the payroll shrinks.

Below: what Astra is being sold to do, the Romanian filings, the labour share number every services founder should plan against, and the labelling duty now sitting on anyone who publishes an AI image in the EU.

OpenAI released GPT-6 Astra on 3 September at $10 per million input tokens and $50 per million output, two and a half times GPT-5.6 Sol. On 10 September it launched ChatGPT for Financial Services on it, with Morgan Stanley and Evercore as design partners, more than 50 data connectors from PitchBook to FactSet, and a task list that reads like a first-year analyst's job spec: LBO models, buyer screens, earnings analysis, pitchbooks, research notes. OpenAI's enterprise follow-up claims Astra finishes Financial Modeling World Cup challenges about four times faster than the winning human.

The benchmark story is messier than the launch. ARC Prize scored Astra at 99.9% on ARC-AGI-3 with OpenAI's own harness and 62.7% on the provider-neutral one, and Fortune found OpenAI had revised five published figures on its launch page after it went live. Artificial Analysis puts Astra level with Claude Fable 5.1 at 53 on its Intelligence Index. Strip out the harness arguments and what is left is a model sold against a job description, at a price a finance director can hold up next to a salary.

If you employ analysts, associates or junior engineers, run that comparison before a client runs it for you. Take three tasks your juniors did last month, put them through Astra on your own data with your own reviewer, and log the cost per finished task, not per token. Then decide what those people should be doing in 2027, because no vendor benchmark will answer that for you.

Found the right person, but they’re in a country where you don’t have an entity? Setting one up can take months and significant cost.

Remote removes that barrier by becoming the legal employer through our own entities — handling compliant contracts, local benefits, tax setup, and onboarding for you. In fact, an employee is onboarded to Remote every 7 minutes.

Once they’re hired, the same in-house teams that support employment locally also run payroll — so you’re not bouncing between disconnected providers. Less setup, less complexity, and less time between finding the right person and getting them started.

See how it works

Ziarul Financiar spent this week going through the 2025 accounts of Romania's mid-tier software firms, and the pattern repeats. R Systems Computaris Europe booked 111.4m lei, level with the 112.8m it made in 2022, with 176 staff against 235, while profit almost quadrupled. Intel's Timișoara software unit lost 12.9% of revenue and 78 of its 301 people, and grew profit 88.6%. Oradea's Qubiz posted its first revenue fall in five years, down 7.6% to 110.2m lei, with 300 people against 321.

None of these filings names AI as the cause, and nobody should pretend they do. What they show is the shape Anthropic's economists describe in the scenario paper below: output roughly holding while the headcount drops, inside an economy built on selling engineering hours. Revenue per employee at Computaris rose about 32% in three years. The outlier points the other way: GlobalLogic Romania grew revenue 5.1% to 360.4m lei with about 1,000 staff, and is building a platform for clients to control humanoid robots.

If you run a services firm in Prishtina, Skopje, Belgrade or Cluj, stop presenting headcount as your growth number to your board and your buyers. Track revenue per engineer every quarter, and move at least one client from time-and-materials to a fixed price per deliverable before renewal, so the productivity gain lands in your margin instead of in the discount your client will ask for.

Anthropic's economists published three scenarios for the US economy in 2030 on 9 September, with a working paper led by Anton Korinek and Charles Jones. In the modest case GDP is 1.6% higher than without AI and unemployment barely moves. In the substantial case, where the median of 10,980 surveyed Americans landed, GDP is 8.3% higher, cognitive employment falls about 4% and knowledge-worker wages go roughly flat. In the extreme case GDP is 32.4% higher, labour's share of income drops from 60% to 45%, cognitive wages sit 11.5% below the no-AI path and nearly one in five cognitive workers is out of work.

The number to hold onto is the labour share, because it tells you who gets paid when the economy grows. Capital is already pricing past the modest case: Cognition raised $2bn at a $48bn valuation on 8 September against $900m of annualised revenue, roughly 53 times, for software that does engineering work. Households are nervous too: the New York Fed's August survey put the mean expected chance of higher US unemployment a year out at 44.4%, the highest since April 2020.

The model covers only the US and has no European equivalent, so founders here are planning on someone else's labour market. Build your own two-column version this month: your revenue if clients expect the same work for less, and your cost base if you hire as you do today. If the plan only works in the modest case, you are betting on a future the median respondent in Anthropic's own survey has already ruled out.

OpenAI shipped ChatGPT Images 2.5 on 8 September to every ChatGPT tier, up to 50% faster than the April release, with better likeness from reference photos and steadier multi-turn edits. Users now make more than 3 billion images a week across ChatGPT and the API, each tagged with C2PA metadata and an invisible watermark. The system card is blunt about the trade: the new realism could, "absent safeguards, allow more convincing deepfakes, including political, sexual, or otherwise sensitive imagery of real people, places or events."

In the EU, OpenAI marks the file but the disclosure falls on whoever publishes it. Article 50 of the AI Act has applied since 2 August: deployers must label AI images that resemble real people, objects, places or events and would pass as authentic, with a lighter touch for evidently artistic or satirical work. The only grace period, to December, covers machine-readable marking by providers of systems already on the market. Fines reach €15m or 3% of worldwide turnover.

Marketing teams should add one line to the content checklist this week: any AI image of a real person, product or place that could pass as a photo gets a visible label before it ships. Keep the C2PA data intact through your export pipeline, since re-exports and platform uploads can strip it. Agencies working for EU brands should write the labelling duty into contracts now, before a client's legal team writes it onto you.

Dev: Cursor Projects runs the long jobs. Cursor released Projects in beta on 10 September: a coordinator agent that holds context across months, delegates to subagents, runs on schedules and watches Slack channels and pull requests. Cursor says its heaviest internal users merge six times as many PRs. Start with a chore nobody wants, like a dependency upgrade or design-system cleanup, and review only the high-impact calls.

Founders: Mistral raised €3bn at €21bn. The round, led by Samsung, EQT's Scaleup Europe Fund and PSG Equity on 8 September, is the largest equity raise by a European tech company, with BlackRock and Luxembourg's government joining. Mistral plans 1GW of European compute by 2030. If a public-sector or regulated buyer asks for a European model, this is now the vendor with the balance sheet to still be around at renewal.

Legal: a five-person Bulgarian legal AI sold to Milan. Lexroom acquired Praven Intelekt on 8 September: 1.2 million Bulgarian legal documents, 19,000 registered users, 900 clients, founded in 2025. Lexroom has €20m ARR and $73m raised. The lesson for vertical AI founders in small markets: local-law data and a user base can be the exit, and the buyer is a Western European platform that cannot translate its way in.

Mobility: Zagreb has robotaxis with nobody at the wheel. Pony.ai and Verne started fully driverless passenger test rides on 10 September on a 22km route to the airport, which they call a European first. Verne has logged over 200,000km since its April launch, and Uber handles bookings. If you run logistics or fleet software in the region, Zagreb is now the reference city for your pitch.

Happy Friday,

Çelik