FutureFast

The Speed of Change Index

Is technology getting faster? Each Monday this index measures how quickly computing, AI, energy, genomics and space launch are improving, and whether that pace is speeding up or slowing down. Every week is kept permanently, so the record can be read across months or years.

Week of 2026-09-28 · method v1 · how it is measured

This week at a glance

2
trend lines accelerating
3
trend lines slowing
2
steady or too new to call
12
verified case studies this week

The weekly record

One point per Monday since the index began on 2026-09-28. The line shows how many tracked trend lines were accelerating that week. It never resets and nothing is ever revised; a change of method starts a new, labeled version.

The record starts with this week. Each Monday adds a point here.

Measured trend lines

Each line is measured data from a named source, not news coverage. Doubling time is how many months it takes a measure to double (or a price to halve). A shorter doubling time than in the window before means the pace is accelerating.

AI capability (Epoch Capabilities Index)

Accelerating

Epoch AI's Capabilities Index combines results from dozens of benchmarks into one score per model. Higher means more capable. The index tracks the best score reached so far (the frontier).

The frontier is rising 14.7 ECI points a year, compared with 10.7 a year in the window before.

1101291481672024202520262023-02-24: 1102023-03-14: 1262024-01-25: 1262024-02-29: 1272024-04-09: 1272024-05-13: 1292024-06-20: 1302024-09-12: 1352024-09-12: 1362024-12-17: 1422025-03-31: 1442025-04-16: 1462025-04-16: 1472025-06-10: 1472025-08-07: 1502025-10-07: 1502025-11-18: 1532025-12-11: 1532025-12-11: 1552026-02-05: 1572026-03-05: 1592026-04-23: 1592026-04-23: 1622026-07-24: 1632026-09-01: 1652026-09-03: 1672026-09-22: 167
1 month
+1% better
6 months
+5% better
12 months
+12% better

Latest: 167 ECI points (Claude Opus 5.5), data through 2026-09-22. Source: Epoch AI, Capabilities & benchmarking. License: CC BY 4.0 (Epoch AI).

Length of task AI can complete (METR time horizon)

Accelerating (clear)

How long a software task (measured in the time a skilled human needs) an AI model completes with 50% success. Measured by METR, a nonprofit AI evaluation lab.

At the current pace it doubles every 3.5 months, compared with every 7.1 months in the window before.

Raw values are not shown: the source has not yet granted permission to republish its data. The rates above are our own calculation from its published results.

1 month
no new data
6 months
+45% better
12 months
+415% better

data through 2026-04-07. Source: METR, via Epoch AI benchmarking hub. License: No published license (METR). Internal computation only; cited derived figures; permission requested..

Compute used to train notable AI models

Slowing

Total computing operations (FLOP, floating-point operations) used to train each notable AI model. The index tracks the largest training run so far.

At the current pace it doubles every 7.7 months, compared with every 4.1 months in the window before.

1001.0 × 10⁹1.0 × 10¹⁶1.0 × 10²³19601980200020201950-07-02: 401957-01-01: 694,8951959-02-01: 6.0 × 10⁸1960-03-30: 7.2 × 10⁸1987-06-06: 2.8 × 10¹⁰1989-11-27: 1.8 × 10¹¹1989-12-01: 1.5 × 10¹²1992-05-01: 1.8 × 10¹³1994-12-02: 1.9 × 10¹³1997-11-15: 3.2 × 10¹³2000-11-28: 6.3 × 10¹⁵2007-06-22: 7.7 × 10¹⁷2007-06-22: 1.4 × 10¹⁸2013-01-16: 2.6 × 10¹⁸2014-06-18: 3.4 × 10¹⁸2014-09-04: 1.2 × 10¹⁹2014-09-10: 5.6 × 10¹⁹2014-12-03: 3.0 × 10²⁰2015-10-01: 3.8 × 10²⁰2016-01-27: 1.9 × 10²¹2016-09-26: 6.6 × 10²¹2018-05-02: 8.7 × 10²¹2019-09-17: 9.1 × 10²¹2019-09-17: 2.2 × 10²²2019-10-23: 3.3 × 10²²2019-10-30: 1.1 × 10²³2020-01-28: 1.1 × 10²³2020-05-28: 3.1 × 10²³2021-08-11: 3.7 × 10²³2021-09-03: 2.0 × 10²⁴2022-03-15: 2.6 × 10²⁴2022-06-29: 2.7 × 10²⁴2023-03-15: 2.1 × 10²⁵2023-12-06: 5.0 × 10²⁵2025-02-17: 3.5 × 10²⁶2025-02-27: 3.8 × 10²⁶2025-07-09: 5.0 × 10²⁶2026-09-03: 1.0 × 10²⁷
1 month
no new data
6 months
+100% better
12 months
+100% better

Latest: 1.0 × 10²⁷ FLOP (GPT-6 Astra), data through 2026-09-03. Source: Epoch AI, Notable AI Models. License: CC BY 4.0 (Epoch AI).

Price of lithium-ion battery packs

Not enough data yet

Volume-weighted average price of a lithium-ion battery pack per kilowatt-hour, from BloombergNEF's annual survey. Each figure is the headline from that year's press release.

At the current pace it halves every 5.5 years.

202420252023-07-01: 1392024-07-01: 1152025-07-01: 108
1 month
no new data
6 months
no new data
12 months
no new data

Latest: 108 US$ per kWh (as published each year), data through 2025-07-01 (this source updates slowly). Source: BloombergNEF annual Battery Price Survey (press releases). License: Cited public headline figures (BloombergNEF press releases).

Cost to sequence a human genome

Slowing (clear)

Cost of sequencing one full human genome, as tracked by the US National Human Genome Research Institute (NHGRI).

At the current pace it halves every 2.6 years, compared with every 7.4 months in the window before.

1,000100,00010.0 million201020202001-07-01: 95.3 million2002-07-01: 61.4 million2003-07-01: 40.2 million2004-07-01: 18.5 million2005-07-01: 13.8 million2006-07-01: 10.5 million2007-07-01: 7.1 million2008-07-01: 342,5022009-07-01: 70,3332010-07-01: 29,0922011-07-01: 7,7432012-07-01: 5,9012013-07-01: 5,0962014-07-01: 4,0082015-07-01: 1,2452016-07-01: 1,1762017-07-01: 1,0152018-07-01: 1,2322019-07-01: 6062020-07-01: 5122021-07-01: 4542022-07-01: 525
1 month
no new data
6 months
no new data
12 months
no new data

Latest: 525 US$ per genome (current dollars), data through 2022-07-01 (this source updates slowly). Source: NHGRI, via Our World in Data. License: CC BY 4.0 (Our World in Data); underlying data from NHGRI (US government, public domain).

Cost to launch a kilogram to orbit

Not enough data yet

Launch cost per kilogram of payload for each launch vehicle, by first launch year. The index tracks the cheapest vehicle so far.

At the current pace it halves every 20.0 years.

10,000198020001962-07-01: 14,9001965-07-01: 8,2001967-07-01: 5,4001985-07-01: 5,1002010-07-01: 2,6002018-07-01: 1,500
1 month
no new data
6 months
no new data
12 months
no new data

Latest: 1,500 US$ per kg to low Earth orbit (constant 2021 dollars) (Falcon Heavy), data through 2018-07-01 (this source updates slowly). Source: CSIS Aerospace Security Project, via Our World in Data. License: CC BY 4.0 (Our World in Data); underlying data credited to CSIS Aerospace Security Project.

Price of solar panels

Slowing (clear)

Average price of a solar photovoltaic module per watt of capacity, adjusted for inflation.

At the current pace it halves every 6.3 years, compared with every 2.9 years in the window before.

1.00101001980200020201975-07-01: 1321976-07-01: 1001977-07-01: 731978-07-01: 511979-07-01: 431980-07-01: 371981-07-01: 291982-07-01: 261983-07-01: 211984-07-01: 201985-07-01: 171986-07-01: 141987-07-01: 121988-07-01: 111989-07-01: 121990-07-01: 121991-07-01: 111992-07-01: 101993-07-01: 9.761994-07-01: 9.231995-07-01: 8.541996-07-01: 7.981997-07-01: 7.951998-07-01: 7.171999-07-01: 6.622000-07-01: 6.502001-07-01: 6.282002-07-01: 5.752003-07-01: 5.462004-07-01: 4.702005-07-01: 4.752006-07-01: 5.172007-07-01: 5.212008-07-01: 4.752009-07-01: 3.172010-07-01: 2.512011-07-01: 2.062012-07-01: 1.112013-07-01: 0.862014-07-01: 0.792015-07-01: 0.742016-07-01: 0.682017-07-01: 0.572018-07-01: 0.512019-07-01: 0.472020-07-01: 0.372021-07-01: 0.332022-07-01: 0.372023-07-01: 0.322024-07-01: 0.27
1 month
no new data
6 months
no new data
12 months
no new data

Latest: 0.27 US$ per watt (constant 2025 dollars), data through 2024-07-01 (this source updates slowly). Source: IRENA; Nemet (2009); Farmer and Lafond (2016); processed by Our World in Data. License: CC BY 4.0 (Our World in Data); underlying data credited to IRENA, Nemet, Farmer and Lafond.

The six D's: how fast technologies move through their stages

Peter Diamandis and Steven Kotler describe six stages an exponential technology passes through (from their book Bold, 2015). If newer technologies move from one stage to the next faster than older ones did, that is a second, independent sign that change is accelerating.

19801990200020102020Digital photographyDigitized: 1975-12-01Disruptive: 2003-01-01Demonetized: 2010-10-06Dematerialized: 2007-06-29Democratized: 2012-01-19Recorded musicDigitized: 1982-10-01Disruptive: 1999-06-01Dematerialized: 2001-10-23Democratized: 2008-10-07Human genome sequencingDigitized: 2003-04-14Disruptive: 2008-01-01Demonetized: 2014-01-14AI language assistantsDigitized: 2018-06-11Disruptive: 2022-11-30Demonetized: 2024-05-13
  • Digitized. The thing has become information: it can be copied, shared and improved at the speed of software.
  • Deceptive. Early exponential growth that still looks small and slow; most observers dismiss it.
  • Disruptive. It now beats the incumbent product or process on cost or capability and is taking real market share.
  • Demonetized. The cost of using it has collapsed toward zero, or money is no longer the main barrier.
  • Dematerialized. Separate physical products, devices or services have been absorbed into it.
  • Democratized. Anyone, almost anywhere, can access it cheaply.

Years from digitized to disruptive

  • Digital photography (digitized 1975): 27.1 years
  • Recorded music (digitized 1982): 16.7 years
  • Human genome sequencing (digitized 2003): 4.7 years
  • AI language assistants (digitized 2018): 4.5 years

Stage dates are documented judgments with sources and confidence levels, not measurements. This starting list was chosen because the technologies are well documented, so it illustrates the method rather than proving the thesis on its own. Six D's of exponentials: Peter H. Diamandis and Steven Kotler, Bold (2015).

This week's case studies

Before-and-after changes found in this week's news. Each one was extracted from the article, then checked: the numbers, the quote and the headline must all be supported by the source. Company claims are labeled as such.

  • The cost to achieve 75% on ARC-AGI-1 dropped from $26 to a penny in 19 months.

    cost per task to achieve 75% on ARC-AGI-1: 26 (2025-02-15) to 0.010 (2026-09-30) USD. 2600x better in 19 months.

    “Hitting 75% on ARC-AGI-1 got 99.95% cheaper in 19 months, from $26 per task to a penny.”

    AI capability · independent measurement · Welcome to September 30, 2026 (2026-09-30)

  • Shodh AI's physical AI model cut chemical process optimization time from over 500 hours to under 6 hours in a live pilot plant.

    Time to optimize chemical manufacturing process: 500 (2026-08-15) to 6.00 (2026-08-15) hours. 83.33x better.

    “lifted yield from 82.4% to 96.7% while cutting an optimization process from over 500 hours down to under six”

    AI capability · independent measurement · EXCLUSIVE: Shodh AI Built India's First Physical AI Model and Cut 500 Hours of Chemistry Work to 6 (2026-09-23)

  • Gemini 4 Argon achieved a 1M-token output limit, up from 64K tokens in the previous generation.

    maximum output tokens: 64,000 (2026-02-15) to 1.0 million (2026-10-01) tokens. 15.63x better in 7 months.

    “Google cites an industry-leading 1M-token output limit, up from 64K”

    AI capability · independent measurement · [AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output (2026-10-01)

  • The price of AI capability is falling 13 times per year

    Price of a given capability per year: 1.00 (2025-09-24) to 0.077 (2026-09-24) fold reduction per year. 12.99x better in 12 months.

    “the price of a given capability is falling 13x a year, apparently the fastest drop of any transformative technology”

    AI capability · independent measurement · Welcome to September 24, 2026 (2026-09-24)

  • Perplexity Photon reduced p99 retrieval latency from 800ms to 65ms while reducing machine count by 20%.

    p99 retrieval latency: 800 (2026-09-15) to 65 (2026-09-29) milliseconds. 12.31x better.

    “Internal p99 fell from about 800ms to about 65ms, on about 20% fewer machines with 2.5 more data per document.”

    AI capability · independent measurement · [AINews] Opus 5.5 is good at explainer videos (2026-09-29)

  • Claude Sonnet 5.5 improved from 10.3% to 70.6% on Terminal-Bench 4.0 benchmark.

    Terminal-Bench 4.0 benchmark score: 10 (2026-07-01) to 71 (2026-09-29) percent. 6.85x better in 3 months.

    “leaps from 10.3% to 70.6% on Terminal-Bench 4.0”

    AI capability · independent measurement · Welcome to September 29, 2026 (2026-09-29)

  • NVIDIA speech model compressed from 1.2GB to 178MB while maintaining accuracy and running at 113x realtime on CPU

    Model size: 1.20 (2026-07-01) to 0.18 (2026-09-22) GB. 6.74x better in 3 months.

    “@vikhyatk released Parakeet Redux, compressing NVIDIAs speech model from 1.2GB to 178MB, running at 113x realtime on CPU”

    AI capability · independent measurement · [AINews] Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M (2026-09-22)

  • Chinese AI models surged from 13% to 67% of tokens used on OpenRouter platform in seven months.

    share of tokens used: 13 (2026-02-15) to 67 (2026-09-15) percent. 5.15x better in 7 months.

    “On OpenRouter, Chinese models accounted for up to 67% of all tokens used in mid-September, up from just 13% in February.”

    AI capability · independent measurement · ALERT: China's AI Models Now Run 67% of Global Developer Workloads and US Committees Are Panicking (2026-09-28)

  • Gemini's performance on Terminal Bench 4.0 rose from 19.0% to 57.6%.

    Gemini performance on Terminal Bench 4.0: 19 (2026-03-15) to 58 (2026-10-01) percent. 3.03x better in 7 months.

    “Terminal-Bench 4.0 rose from 19.0% to 57.6%”

    AI capability · independent measurement · [AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output (2026-10-01)

  • Claude Opus 5.5 costs 40% as much as Opus 5 while matching competitor performance

    Cost relative to previous generation: 100 (2026-09-15) to 40 (2026-09-24) percent of Opus 5 cost. 2.5x better.

    “matching Fable 5.1 on most work at 40% less cost than Opus 5”

    AI capability · independent measurement · Welcome to September 24, 2026 (2026-09-24)

  • Gemini 4 Argon's output token pricing fell from $20 to $10 per million tokens with introductory discounts.

    price per million output tokens: 20 (2026-02-15) to 10 (2026-10-01) USD per 1M output tokens. 2x better in 7 months.

    “Standard pricing is $4/$20 per 1M input/output tokens. A 50% introductory discount brings it to $2/$10”

    AI cost · independent measurement · [AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output (2026-10-01)

  • Shodh AI's physical AI model improved chemical yield from 82.4% to 96.7% in a live industrial pilot.

    Chemical reaction yield in bioreactor optimization: 82 (2026-08-15) to 97 (2026-08-15) percent. 1.17x better.

    “lifted yield from 82.4% to 96.7% while cutting an optimization process from over 500 hours down to under six”

    AI capability · independent measurement · EXCLUSIVE: Shodh AI Built India's First Physical AI Model and Cut 500 Hours of Chemistry Work to 6 (2026-09-23)

How the index is measured (method v1)

Trend lines come first. The backbone is measured data from named sources: Epoch AI's Capabilities Index and its database of AI training runs, METR's measurements of how long a task an AI system can complete, solar module prices (IRENA and others, via Our World in Data), the cost of sequencing a human genome (US National Human Genome Research Institute), the cost of launching a kilogram to orbit (CSIS), and lithium-ion battery pack prices (BloombergNEF press releases). News never moves a trend line.

Doubling time. For each line we fit the trend over the most recent 24 months (10 years for sources that publish once a year) and convert it to the number of months the measure takes to double, or a price to halve. That puts AI, batteries and genomes on one scale. For record-style measures (the best AI model, the cheapest launcher) only new records count. For average prices every observation counts, so a price that rises again shows up as a slowdown instead of being hidden.

Acceleration and slowdowns. Each line's current window is compared with the window before it. Faster than before by more than 10% is marked accelerating, slower by more than 10% is marked slowing, and anything between is steady. “Clear” means the 90% confidence ranges of the two windows do not overlap. Slowdowns are published exactly as prominently as speedups.

Changes over 1, 6 and 12 months compare the latest value with the value that stood that many months ago. When a source has published nothing new in that period, the index says so rather than showing a false zero.

Case studies. Disruption Radar reads several hundred technology articles a week. A specialized judging model, Jev from TypeSafe AI, first asks whether each article reports a measurable change. For articles that do, Claude (Anthropic) extracts the before and after values. Jev then checks that the article supports the numbers, the quote, the change over time and the plain-language headline. Items that fail go to human review, and the same development reported by several outlets is counted once.

Permanence. Each week's index is frozen when it is published and kept forever. Corrections are added as new versions; earlier values are never overwritten. A change to this method creates a new version of the index alongside the old one.

Limits. Some sources update slowly, and those lines say how old their data is. METR has not yet granted permission to republish its data, so only our derived rates are shown for that line. The six D's stage dates are judgments with stated sources. This is an independent analysis and not investment advice.

Follow the index

Join the FutureFast list to receive Speed of Change Index updates and reports.

We respect your privacy. Unsubscribe at any time.