Newsletter · November 8, 2026
Weekly Digest 45
METR asks how much AI is speeding up Anthropic’s models. Plus, US productivity rises while workers’ share of output falls to a postwar low, and GPT-6 Astra reads and edits neural network weights.

Topics we are tracking
Anthropic’s models are improving faster. How much of that is down to AI?
Sources: METR
We have written before about how AI is changing work inside Anthropic. A new assessment from METR looks at whether AI is accelerating the development of the models themselves.
METR examined internal benchmark results and research inputs and interviewed Anthropic employees. It estimates that since October 2025, Anthropic’s models have been improving on well-specified tasks at roughly twice their earlier pace. Its central estimate is 1.9x, measured using Anthropic’s version of the Epoch Capabilities Index. More interestingly, METR believes that using AI in research likely accounts for most of the current acceleration. It is more confident that faster growth in training compute does not explain it.
Anthropic’s benchmark scores have risen about twice as fast since October 2025
AECI score of each Anthropic model, by the date it became available to most employees
21 Anthropic models from Claude 3 Opus in early 2024, at about 125, to Model 4, the newest internal model, at 178.6. The fitted trend rises 15 points a year before October 2025 and 29 points a year after.
Recreated from METR’s chart. Scores and dates are read off the published figure, so they are approximate; METR states 178.6 for Model 4. AECI is Anthropic’s version of the Epoch Capabilities Index. The trend lines are METR’s fits, a 1.9x acceleration with a 95% confidence interval of 1.5x to 2.8x. Confidence intervals for individual models are left out.
Source: METR · 2026-11-05
How far this recursive improvement loop can go is an open question, but Anthropic is far from the only company working on it. Recursive Superintelligence and Jeff Dean’s new company, Discovery Loop, are two other examples of companies betting that AI can shorten the time between a research idea and an experimental result. Sustained acceleration still depends on whether the labs have the resources to run those experiments. In a related paper, two of the METR researchers and their co-authors point to compute and data as potential bottlenecks.
As models improve, benchmark scores become harder to interpret. Once a model gets nearly every question right, a higher score tells us relatively little about what the next model can do. The change in slope is perhaps the most interesting part for us: one strong model could reflect a single breakthrough, but if successive models keep improving at the faster pace, it becomes harder to explain the trend that way. How much of that change is due to AI remains difficult to judge from the outside, since Anthropic redacted much of the evidence METR found most informative.
We don’t know how much compute the labs will allocate to AI-assisted research going forward, but there is a clear trade-off when capacity is tight. Using that compute to serve paying customers generates revenue today. Using it to run research agents could help the lab develop a better model sooner, but the return is less certain and comes later. If Anthropic and OpenAI go public, it will be interesting to see how quarterly reporting affects that decision. How much capacity will the labs be willing to reserve for research when they could use it to meet customer demand today?
US productivity rises while workers’ share of output falls to a postwar low
Sources: BLS · San Francisco Fed · New York Fed
US businesses produced 3.7% more output annualised in the third quarter, while hours worked grew by just 0.1% which put productivity growth at 3.6% for the quarter. These figures can move around quite a bit, but the improvement is also visible over a longer period. Since the end of 2019, output per hour in the nonfarm business sector has grown by 2.2% a year, up from 1.5% in the previous business cycle and back in line with its postwar average.
What is driving the improvement? Part of it comes from businesses investing in equipment and software that allow workers to produce more per hour. Firms have also been making more intensive use of the equipment they already have. The San Francisco Fed estimates that investment and higher utilisation together contributed around 1.3 percentage points to annual productivity growth from early 2023 through the second quarter of 2026. There is some evidence that AI is contributing as well. For example, industries with higher AI adoption generally saw faster productivity growth, but the industries that adopted AI most widely were not always the ones whose contribution to US growth increased the most. Retail, for example, saw a larger increase in its contribution than the information sector despite reporting lower AI use.
What stood out to us was that workers’ share of output fell to 53%, the lowest level in the BLS series going back to 1947. Workers still saw a gain in purchasing power but real hourly compensation, including benefits, rose only at a 1.1% annualised rate. The dollar value produced per hour grew at roughly 5.7%, while hourly compensation grew by 2.9%. Workers earned more, but their pay accounted for a smaller share of each dollar of output.
Workers’ share of US output is the lowest on record
Labor share, US nonfarm business sector, quarterly, Q1 1947 to Q3 2026 (%)
319 quarterly readings of the US nonfarm business labor share from Q1 1947 to Q3 2026. It starts at 65.6%, peaks at 66.2% in Q4 1960 and ends at 53.0%, the lowest reading in the series.
Labor share is the compensation of all workers as a share of the sector’s output. BLS publishes the history as an index, which we scaled to the 53.0% it reported for the third quarter of 2026. The share peaked at 66.2% in the fourth quarter of 1960.
Source: U.S. Bureau of Labor Statistics · 2026-11-05
It’s tempting to put this down to the growth of tech companies, which can serve millions of customers with relatively few employees. But the New York Fed found that shifts between industries explained very little of the decline. From 2019 to 2024, payroll fell as a share of the value businesses produced within industries.
How much of this improvement can we attribute to AI is still an open question. The FT’s John Burn-Murdoch pointed out on X that the pickup in US productivity was already visible in 2020, well before the widespread use of generative AI. Tech investors like Gavin Baker looked at the same charts and saw evidence that AI investment was starting to pay off, pointing to the renewed rise in productivity after ChatGPT launched. For us, the more interesting question is how much AI is adding to this trend and perhaps even accelerating it.
GPT-6 Astra identifies games from neural network weights and edits Qwen’s weights to fix a reasoning loop
Sources: Takehi et al., MIT CSAIL · Alex Krentsel on X
This is perhaps the most interesting story of the week, or even the most interesting in recent memory. Researchers at MIT gave GPT-6 Astra the weights of four small neural networks and asked it to identify which game each had been trained to play. Astra knew what the inputs and possible moves were, but it was never told the names of the games and could only see the weights.
In a separate experiment, the researchers asked Astra to make a small Qwen model better at maths. This time, Astra could run Qwen on practice problems and inspect its weights. By changing just 0.016% of those weights, Astra raised Qwen’s score on MATH-500 from 7.9% to 48.1%, largely because Qwen got stuck in reasoning loops less often and could finish problems it was already capable of solving.
Editing 0.016% of Qwen’s weights lifted its scores as the loops receded
Qwen3.5-0.8B before and after GPT-6 Astra’s weight edit (%)
MATH-500: base 7.9%, edited 48.1%; GSM8K: base 13.8%, edited 67.1%; IFEval: base 21.0%, edited 62.3%; MATH-500 answers that hit the token limit: base 93.0%, edited 33.1%
The first three groups are benchmark scores, where higher is better. The last is the share of MATH-500 answers that ran to the 32,768-token limit, mostly because the model looped. Astra changed 131,072 of 0.8 billion weights without any training, using 50 forward passes on practice problems.
Source: Takehi, Teoh, Zhang and Khattab (MIT CSAIL) · 2026-11-02
Google researcher Alex Krentsel reproduced the Breakout experiment with Astra, Claude Opus 5.5 and older models. What stood out to us was where the step change appeared. Earlier models could already write a strategy for playing Breakout in Python. Roughly one model generation later, Astra and Opus 5.5 could write that strategy directly into the weights of a small network, although Astra scored slightly higher.
It’s still unclear how well this scales to larger models, but it could open up a new path for mechanistic interpretability and for post-training. As humans, we find it difficult to look at a model’s weights and work out why it behaves a certain way. To elicit a particular behaviour, we often build elaborate datasets, environments and training regimes. This could be another example of Rich Sutton’s The Bitter Lesson. If a general-purpose model can use a few examples to find and change the weights responsible for a particular behaviour, then some fixes that would otherwise require a new round of post-training might start with a researcher simply describing what they want to change.
Seen on X
Other interesting stories
- Financial Times: Anthropic strikes $3.4bn deal to generate gas-fired power for data centres
- SemiAnalysis: We Visited China’s Stargate in Inner Mongolia
- Construction Physics: How Do Building Codes Affect the Price of Housing?
- Interconnects: How big is the open model economy today?
- Bloomberg: Sequoia Quits London After US Team Lands Buzzy European AI Deals



