AI, Work and the Economy

Latest/Story

New AI models

Seven items, from 1 Oct 2026 to 19 Nov 2026, oldest first.

See also: Point 1 of our view: AI does more office work, cheaper

AI and technologyReport

Google releases Gemini 4 Argon (30 Sep): #1 on AutomationBench at 51.3%, $2/$10 per M tokens introductory

Google DeepMind's Gemini 4 Argon scores 77.9% on DeepSWE v1.1 and leads the Vals Index across finance, legal and tax work; output limit rises to 1M tokens from 64K. Introductory price $2/$10 per million tokens, then $4/$20; first to cyber defenders, then paid API users.

Why it matters. Google's newest AI model leads a benchmark of automating office work and launches at the same low price as its rivals' latest models, a third major release in a week.

Next. Independent tests of the model's ability to do long, paid tasks are due in the coming weeks.

Sources: blog.google, news.ycombinator.com

AI and technologyNew dataUpdate

Mistral Large 4 (6 Oct): 59.9% on AutomationBench at $1.36/$4.18 per M tokens, open weights by month-end

Mistral released a 1-trillion-parameter model (49bn active) in preview, reporting 59.9% on AutomationBench against Gemini 4 Argon's 51.3% a week earlier, at about two-thirds of Argon's introductory price; weights are due by the end of October. Scores are the vendor's own.

Why it matters. A European lab says its new model, which anyone will be able to download, beats last week's leader at automating office workflows for less money.

Next. The model's weights are due to be published by the end of October.

Sources: mistral.ai, news.ycombinator.com

AI and technologyNew dataUpdate

Google's Gemini 4 Carbon (28 Oct) ranks #1 on Artificial Analysis index at 59; 88.2% on SEC-bench Pro

Google released Gemini 4 Carbon, first to cyber defenders and trusted partners, at $2/$10 per million tokens introductory ($4/$20 after). It scores 59 on the Artificial Analysis Intelligence Index (#1), 88.2% on vulnerability finding and 54.8% on ExploitGym, both stated as state of the art.

Why it matters. Google's new model ranks first on a widely used composite of AI capability and sets new marks in finding and exploiting software flaws. It is the third leading lab this month to show a large gain in skilled technical work, which could, if it reaches wide use, compress demand for some security and coding roles.

Next. Independent measurements on work tasks and Google's date for general availability.

Sources: blog.google, news.ycombinator.com · Point 1 of our view

AI and technologyNew dataUpdate

Mistral releases Large 4's weights (29 Oct) at $0.70/$2.10 per million tokens, about half its preview price

Mistral published the weights of Mistral Large 4, a 1-trillion-parameter mixture-of-experts model, under a modified MIT licence permitting commercial use; its API price is $0.70 input and $2.10 output per million tokens against $1.36/$4.18 at the 6 Oct preview. It scores 41% on Terminal-Bench 4.0 (28% in preview) and 48% on Artificial Analysis's coding-agent index, tied with the best open models.

Why it matters. A European laboratory has released, free to download and use commercially, a model close to the best open models on coding and agent tasks, and halved its hosted price within a month. If capable models keep getting cheaper at this pace, replacing routine office tasks may become affordable for many more firms.

Next. Independent test results on the released model, and any price response from the larger laboratories.

Sources: mistral.ai, news.ycombinator.com · Point 1 of our view

AI and technologyNew dataUpdate

Google's Gemini 4 Hydrogen (12 Nov) beats Argon on the AA index (51 vs 50) at $1.50/$7.50, vs Argon's $4/$20 list

Google DeepMind released Gemini 4 Hydrogen on 12 Nov, a fast model scoring 51 on the Artificial Analysis Intelligence Index against 50 for its frontier Gemini 4 Argon, at $1.50/$7.50 per million tokens. Argon launched in October at $2/$10 introductory, $4/$20 standard; Hydrogen ranks first on Vals' tax-agent benchmark (52.3%) and ties first on legal research (55.8%).

Why it matters. Google's new fast model matches its own frontier model of a month ago on a widely used index, at well under half that model's standard price, and leads a benchmark of professional tax work. If capability keeps cheapening at this pace, automating skilled office tasks could become economic in many more firms within months.

Next. Independent benchmark results over the coming week, and Google's pricing for Argon from 1 December.

Sources: blog.google, news.ycombinator.com · Point 1 of our view

AI and technologyNew dataUpdate

OpenAI's GPT-6.2 Astra (17 Nov): +6.2 points on GDPval v2 at $8/$40, 20% cheaper; first model rated High for self-improvement

OpenAI released GPT-6.2 Astra and Sol on 17 Nov. Astra gains 6.2 points over GPT-6 Astra on GDPval v2 (professional work tasks) with 22% fewer output tokens, and 5.2 points on Agents' Last Exam at under a third of the cost per task; its list price falls 20% to $8/$40 per million tokens, while Sol ($2/$10) gains 5.3 points on GDPval v2. Its system card rates Astra High for AI self-improvement, the first released OpenAI model to reach that level, and both models Critical for cybersecurity; Epoch's capability index puts Astra at 174, level with Claude Fable 5.2.

Why it matters. OpenAI's new frontier model does measurably more professional work than its predecessor at a fifth less per token, and OpenAI rates it the first of its released models able to speed up AI research materially. If such ratings hold, the cost of doing office tasks by machine might keep falling faster than firms can reorganise around it.

Next. Independent tests of the model's work-task scores and costs, expected in the coming days.

Sources: openai.com, unite.ai, implicator.ai · Point 1 of our view

AI and technologyNew dataUpdate

Google's Gemini 4 Neon (18 Nov): 61 on the AA index, #1 on Vending-Bench 2 ($16,843), at $3/$15 per m tokens

Google released Gemini 4 Neon on 18 November: third on the Artificial Analysis intelligence index (61), first on its coding-agent index (58.6), first on Andon Labs' Vending-Bench 2 with $16,843 and on Zapier's AutomationBench (56.3%). Google puts its cost per task 50-80% below the leading competing models; list price $3 in and $15 out per million tokens.

Why it matters. Google's new model ranks near the top on tests of whole tasks, including one that has an AI run a small business over months, at a cost per task Google puts at half to a fifth of the leading competitors'. If such scores hold up independently, the price of a given level of machine office work may be falling faster than firms can adopt it.

Next. Independent scores from Artificial Analysis and Andon Labs, and any measurement of how long a task the model can complete.

Sources: blog.google, news.ycombinator.com · Point 1 of our view