DeepMind's own index puts AI progress at 0.1 points a month even with AI writing over 75% of its code
Google DeepMind's Gemini 4 Argon report shows no significant acceleration in AI progress since April, though AI writes over 75% of the lab's code. The post-April rate is 30% above the earlier one, yet DeepMind's analysis finds no substantial difference.
The Investor · Invest desk
Before April 2026, public models gained 0.091 capability index points a month; after April, 0.118 points a month, per Crypto Briefing. The report found no substantial difference between the two rates.
| Measure | Value | As of | Claim |
|---|---|---|---|
| Pre-April, per Crypto Briefing | 0.091 points a month | 26 | |
| Post-April, per Crypto Briefing | 0.118 points a month | 26 |
What happened
- DeepMind published the assessment on October 14, 2026, as its first comprehensive internal look at whether AI is speeding up machine learning research.
- The yardstick is GCI v1, an internal index that blends scores from 123 models across 92 benchmarks into one number per model.
- A survey of 12 senior DeepMind researchers found automation highest in clearly defined tasks with checkable answers and lowest in open-ended scientific ideation.
- On RE-Bench, an internal set of 74 machine learning engineering tasks, the Argon agent solved roughly half, against 27% for Gemini 3.7 Flash.
- Decrypt reported that the findings landed a month after Anthropic CEO Dario Amodei argued recursive self-improvement is already accelerating, a warning that sent chip stocks down almost 6%.
Why it matters
- What it costs cost Argon's lead over the trend is credited to compute and organizational investment, so the next jump of that size has to be funded again.
- Who's exposed exposure Decrypt reports that more automation means broader agent access to company systems, so each rise in delegation widens the room a misaligned model would have to cause trouble.
- What doesn't add up contradiction The write-ups differ on what Argon's roughly 50% covers: Decrypt says half of the tasks it finished, Crypto Briefing half of all 74. The denominator changes the distance to the 90% red-flag line.
- What it limits constraint Slowdown arguments that rest on acceleration already under way now face a lab's own measurement, so the case that survives is about future capability and the 90% line.
Next story · Invest · Oct 1
OpenAI and DeepMind researchers go on camera to warn AI labs are rushing self-improving systemsConfirmed2 publishers
The 30% is the figure most likely to be misread. Crypto Briefing, summarising the report, gives the monthly gain for public Gemini models as 0.091 index points before April and 0.118 after [26]. That is 0.027 points, or about 30% faster [27]. On a chart it looks like acceleration. DeepMind's analysis found no substantial difference between the two [26], working from about six months of post-April data [28]. Neither write-up reproduces the confidence interval around either slope, so a reader cannot tell a flat line from a test with too little data to see a 30% change.
Argon is the second puzzle. At roughly 3.3 it sits about 0.4 points above Gemini 3.8 Flash, which scored just above 2.9 in September (Argon was unveiled on September 30) [5][17][21]. Nearly every release hugs the trend line and Argon's error bar stretches low enough to touch it [6], so if Flash is on the line, 0.4 points at 0.1 a month is four months of progress [4][21]. Haydn Belfield, a DeepMind researcher who worked on the assessment, wrote that Gemini 4 Argon sits above the trend but that the team traces this to increased compute and organizational investment, not to AI-driven acceleration [7][8]. On that account the four months were paid for, and the next four will have to be paid for the same way.
The 75% and the 88% count different things. More than 75% of production code at DeepMind is AI-written [9], yet only 11.9% of pull requests are fully delegated [10], and 100 minus 11.9 is 88.1 [22], close to the roughly 88% that Belfield says still need multi-turn human steering or human code authorship [11]. The delegated share is rising, by 4.4 points between April and September [23]. A review of 1,389 Argon agent logs found that the bigger the task, the less often the model finished it alone [13].
The tripwire sits further out. DeepMind treats a 90% score on its research engineering tasks as a red flag for full automation [16]. The step from Gemini 3.7 Flash to Argon was 23 points [24], and the remaining 40 points to that line is about 1.7 such steps [25]. Two steps of that size would cross it. The two scores come from different model lines, so the yardstick is crude.
Three outcomes would change the picture. The 30% could be real and turn significant as the sample grows. The index could hold near 0.1 while a larger compute budget lifts the next model above the line again; that would look like acceleration and be spending. Or the research-engineering score could cross 90% while the index stays flat.
We think the report earns a narrow conclusion: inside DeepMind, through September, no acceleration beyond what compute and investment explain [1]. "We therefore conclude that the risk from AI-driven ML R&D is currently low," Belfield wrote [18]. The counter-thesis is a short record, an internal index [3] with about six months of post-April data [28], and one lab's models cannot settle a claim about the whole industry. Crypto Briefing put the proof this way: a sustained break above roughly 0.1 points a month without a matching jump in compute [29]. We would add a score above 90% on the research-engineering tasks.
What to watch
- Whether later DeepMind reports show the monthly index gain holding near 0.1 points, or staying above it without a matching rise in compute.
- Whether the fully delegated share of pull requests keeps climbing from 11.9% and whether any model's research-engineering score approaches the 90% red-flag line.
- Whether other labs publish comparable measurements that test Amodei's claim that recursive self-improvement is already accelerating.
Information, not investment advice. This story is general information about companies and markets, drawn from cited reporting. It is not a recommendation to buy, sell or hold any security; check the cited sources before acting.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+15
- Incentives50
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Google DeepMind's Gemini 4 Argon safety report found no statistically significant acceleration in the pace of its AI progress after April 1, 2026.
- [2]
Google DeepMind published its first comprehensive internal assessment of AI-driven machine learning research on October 14, 2026, inside the Gemini 4 Argon Safety & Security Report.
- [3]
GCI v1 is an internal DeepMind index that blends the scores of 123 AI models across 92 benchmarks into a single number per model.
- [4]
Frontier Gemini capabilities have improved at a steady pace of about 0.1 GCI points per month.
- [5]
Gemini 4 Argon scores roughly 3.3 on the capability index, while Gemini 3.8 Flash sits just above 2.9 in September 2026.
- [6]
In the report's chart nearly every release hugs a straight line; Argon is the one dot above it, though its error bar stretches low enough to touch the trend.
- [7]
"While Gemini 4 Argon sits above [the] trend, we trace this to increased compute and organizational investment rather than reflecting AI-driven acceleration."
- [8]
Haydn Belfield is a Google DeepMind researcher who worked on the assessment.
- [9]
Most production code at DeepMind is now more than 75% AI-written.
- [10]
The share of fully delegated pull requests at DeepMind rose from 7.5% to 11.9% between April and September.
- [11]
About 88% of pull requests still require multi-turn human steering or human code authorship, and code changes remain subject to human review before submission.
- [12]
A survey of 12 senior DeepMind leaders and researchers found automation highest in clearly defined, verifiable execution tasks and lowest in open-ended scientific ideation, novel experimental design, diagnosing subtle training instabilities and long-horizon research taste.
- [13]
DeepMind reviewed 1,389 logs of Argon working as an agent inside the company; the bigger the task, the less likely the model finished it alone and the more often a human stepped in.
- [14]
On RE-Bench, an internal benchmark of 74 realistic machine learning engineering tasks, the Argon agent solved roughly half; Gemini 3.7 Flash managed 27%.
- [15]
Gemini 4 Argon solved roughly half of the internal research engineering tasks it finished.
- [16]
DeepMind treats a 90% score on its internal research engineering tasks as a red flag for full automation.
- [17]
Google unveiled Gemini 4 Argon on September 30.
- [18]
"We therefore conclude that the risk from AI-driven ML R&D is currently low."
- [19]
The report landed a month after Anthropic CEO Dario Amodei argued that recursive self-improvement is already accelerating, a warning Sam Altman and Elon Musk endorsed and one that sent chip stocks down almost 6%.
- [20]
More automation means giving AI agents broader access to company systems, and more room for a misaligned model to cause trouble.
- [21]
Argon's score is about 0.4 points above Gemini 3.8 Flash, equal to about four months of trend at 0.1 points a month.
- [22]
The complement of the 11.9% fully delegated share is 88.1%, close to the roughly 88% needing human steering or authorship.
- [23]
The fully delegated share rose 4.4 percentage points between April and September.
- [24]
Argon's roughly 50% RE-Bench score is about 23 points above Gemini 3.7 Flash's 27%.
- [25]
The gap from roughly 50% to the 90% red-flag line is 40 points, about 1.7 times the 23-point step between the two models.
- [26]
For public models, the monthly change in the capability index was 0.091 points before April 2026 and 0.118 points after; the report found no substantial difference between the two.
- [27]
The post-April monthly gain is 0.027 index points higher than the pre-April gain, about 30% faster.
- [28]
The post-April test window is about six months, from April 1 to Argon's September 30 unveiling.
- [29]
A sustained break above roughly 0.1 GCI points per month, without a matching jump in compute, would be the signal the self-improvement camp has been waiting for.
Sources
2 independent publishers whose own reporting we read for this story.
- cryptobriefing.comGoogle DeepMind finds no acceleration in AI research progress
1 article · October 15, 2026
- decrypt.coGoogle Says AI Isn't Building Itself Yet—Even Though It Writes 75% of the Code
1 article · October 15, 2026
Topics and entities
Topics
- AI-Assisted Coding
- AI safety reports
- Recursive Self-Improvement