AI capability and capital brief
A weekly view of how capable the leading AI models are, how quickly that capability is getting cheaper, and where investment in AI companies is going, with a section on Singapore. It brings independent test results, funding reports and official announcements into one place, and every figure links to its source.
At a glance
Headline measures. Funding figures are amounts announced and counted by Crunchbase.
Top score, Artificial Analysis Intelligence Index
Frontier score, Epoch Capabilities Index (ECI)
Gap between the best closed and best open-weights model, ECI points
Largest single round of 2026 so far
Companies raising rounds of $1 billion or more, latest quarter
Model capability
Two independent testers. Artificial Analysis combines its own test runs into one index. Epoch AI combines many benchmarks into one capability scale and keeps a history, so it can show the trend.
Artificial Analysis Intelligence Index, top 15
Index v5.0.1. Scores are not comparable with earlier versions of the index. Open-weights models are shown in green.
Highest ECI score by quarter
Best score among models released by the end of each quarter. 2026 Q4 covers releases up to 9 November (hollow points).
Leading models on the Epoch Capabilities Index
Top ten closed models and top six open-weights models. The scale does not start at zero, so the gaps are easier to see.
| Model | Developer | Country | ECI | Released | Weights |
|---|
Capability against cost, Artificial Analysis tests
Each dot is a model. Higher is more capable. Further right costs more per test task. The cost scale is logarithmic: each gridline is ten times the one before.
| Rank | Model | Developer | Index | Cost per task | Minutes per task | Weights |
|---|
Is AI getting cheaper? Price of the same capability over time
Source: Epoch AI, The plunging price of thought, 22 September 2026. The chart shows the longer price-per-token series behind this finding.
Cheapest price per million tokens of any model that scores at least as well as the original GPT-4 (March 2023). The price scale is logarithmic. Epoch AI published this series in March 2025, so it stops at February 2025. It follows the cheapest model at a fixed level, not the price of the newest top models.
| Test and level | Date | Cheapest model at that level | US$ per million tokens |
|---|
Model releases
Selected releases from the last seven weeks, newest first. Dates are the week the release was reported. Figures are as announced and many are reported by the developer.
Investment in AI companies
Venture funding counted by Crunchbase. Crunchbase revises totals as late deals are added, so a figure can change between reports.
Global venture funding by quarter, 2026
US$ billions. Crunchbase gave no AI dollar figure for Q2, only a share above 70%.
Global venture funding by month, 2026
US$ billions, as first reported. Months with no monthly report in the sources read are left out. The darker part of each bar is AI, where Crunchbase gave a dollar figure.
Largest AI rounds of 2026
US$ billions. Rounds of $1 billion or more named in Crunchbase News and other reports. Foundation model companies are shown in the darker colour.
Where the largest rounds went
Total of the rounds in the chart, by part of the AI stack, with the number of rounds in brackets. This covers only the rounds listed, not all AI funding.
Round details
| When | Company | Amount | Valuation | Area | Country | Note | Source |
|---|
Singapore
Government funding for AI and notable AI startup rounds in Singapore this year, newest first.
Data and method
A weekly check looks for new scores, releases and funding reports. A figure is only replaced when the new value comes from a source page that was read. A failed check leaves the last figures in place and is recorded in the log.
Source freshness
| Source | Used for | Data as of | Last checked | Last data update | Next expected | Status |
|---|
Refresh log
| Time (SGT) | Task | Result | Note |
|---|
Definitions and what is left out
- Artificial Analysis Intelligence Index
- A single score that Artificial Analysis builds from its own runs of several tests, including reasoning, knowledge, coding and agent tasks. Each model is run at a stated effort setting, such as Max or High.
- Epoch Capabilities Index (ECI)
- Epoch AI's single capability scale, built from many benchmarks. Its history allows a fair comparison of the best score over time.
- Open weights
- The model's trained parameters are published, so others can run or adapt it. Some open-weights licences restrict use.
- AI-related funding
- Crunchbase's count of venture funding to companies it tags as AI, across the stack. Crunchbase does not publish a fuller definition.
- Left out for now
- LMArena human preference rankings and the Artificial Analysis image and video arenas could not be read automatically, so they are not shown. Price per million tokens is given only where a release note states it. Company revenue and total funding raised are not shown because the sources read gave them without dates.
An independent research project by Shamim Akhtar. It brings together capability scores and price trends from Epoch AI, model releases from the ThursdAI release list, funding figures from Crunchbase News and announcements from Singapore government newsrooms. It is checked for new releases every week. It shows data only and does not make recommendations.