AI, Work and the Economy

Latest/Story

Anthropic's models

Five items, from 28 Sep 2026 to 11 Nov 2026, oldest first.

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

AI and technologyNew data

Anthropic releases Claude Sonnet 5.5 (28 Sep): GDPval-AA 1844 Elo vs 1449 for Sonnet 5, same $2/$10 token price

Anthropic released Claude Sonnet 5.5 on 28 Sep. On GDPval-AA v2.1 (economically valuable professional tasks) it scores 1844 Elo against Sonnet 5's 1449, at unchanged token prices ($2 in, $10 out per million) and, Anthropic says, up to 30% lower cost per task.

Why it matters. A cheaper, mid-priced AI model now scores far higher on a benchmark built from real professional work, which lowers the cost of handing office tasks to machines.

Next. Independent benchmark results and customer pricing over the coming weeks.

Sources: anthropic.com, news.ycombinator.com

AI and technologyNew dataUpdate

Anthropic's Claude Haiku 5.5 (7 Oct): 1620 on GDPval-AA vs 735 for Haiku 4.5, at a tenth of the price

Anthropic released Claude Haiku 5.5 on 7 October at $0.10 input and $0.50 output per million tokens (up to 100k tokens), against $1/$5 for Haiku 4.5. It scores 1620 Elo on GDPval-AA, a test of professional work across 44 occupations, against 735 for Haiku 4.5, 1437 for OpenAI's GPT-6 Luna and 1840 for Sonnet 5.5; on OSWorld computer use it scores 72.4% against 15.7%.

Why it matters. Anthropic's new small model scores more than twice its predecessor on a benchmark of professional work across 44 occupations, at a tenth of the price. If scores of this kind translate into practice, the cost of automating routine office tasks might be falling faster than firms can reorganise around it.

Next. Independent benchmark results and any price responses from OpenAI and Google in the coming weeks.

Sources: anthropic.com, news.ycombinator.com

AI and technologyNew dataUpdate

Anthropic's Claude Fable 5.2 (9 Nov): 41.3% on the Remote Labor Index vs 17.9% for Fable 5.1, at $8/$40

Anthropic released Claude Fable 5.2 on 9 November, with a restricted twin, Mythos 5.2, for vetted cyber and life-sciences users, at $8 input and $40 output per million tokens. On the Remote Labor Index, paid freelance projects judged against human work, it completes 41.3% against 17.9% for Fable 5.1 and about 21% for Opus 5.5 and GPT-6 Astra; on AutomationBench business workflows 51.4% against 31.4%, and on a robotics simulation 64.9% against 27.6%.

Why it matters. Anthropic's new model more than doubles its predecessor's share of paid freelance projects completed to a client's standard, to about two in five. Scores on whole projects rather than single tasks are closer to what employers buy; if they hold up in independent tests, the share of remote office work that software could take on might be rising faster than firms and labour markets can adjust.

Next. Independent test results and measures of how long a task the model can finish unaided, expected in the coming weeks.

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

AI and technologyNew dataUpdate

Anthropic (9 Nov): Fable 5.2 designs three working circuit boards; PCB placement score 69% vs 25% for Fable 5

With Diode Computers, Claude Fable 5.2 designed a speakerphone, a voice hub and a handheld device from short prompts; all three were manufactured from its files without design changes and worked for weeks. It scores 69% on Diode's PCB Gym (Fable 5: 25%) and 83.5% on BenchCAD 2.1, about double Fable 5.1.

Why it matters. Anthropic reports that its newest model designed three working electronic devices whose circuit boards were built straight from its files, more than doubling its predecessor's score on a board-layout test. If this holds at production scale, skilled hardware design could become another field where AI does much of the work.

Next. Independent tests of the model on engineering tasks, and whether electronics firms adopt it for production designs.

Sources: anthropic.com · Point 1 of our view

AI and technologyNew dataUpdate

Epoch (11 Nov): Fable 5.2 scores a record 174 on its capability index, above trend; open model builds Firefox exploits for ~$1,000

Epoch's capability index puts Anthropic's Claude Fable 5.2 (released 9 Nov) at 174, above the previous high of 169 for GPT-6 Astra and above the trend for reasoning models. In the same brief, the open-weight GLM 5.3 built working exploits for six recent Firefox versions from unreleased patches, at about $1,000 per attack.

Why it matters. An independent measure ranks Anthropic's newest model above every earlier system and ahead of the recent pace of improvement, and a freely available model produced working browser attacks for about $1,000 each. If such gains carry over to ordinary office tasks, the share of work machines can do cheaply might grow faster than current forecasts assume.

Next. Further independent measurements of the model's ability on long and economically useful tasks over the coming weeks.

Sources: epochai.substack.com · Point 1 of our view