How Capable Is AI Today?

<p class="kicker">MST0052 · Lecture 2 · Extra</p>

## How capable<br>is AI today?

### Language models and AI agents, September 2026

Vegard H. Larsen — BI Norwegian Business School

---

## From ChatGPT to GPT-6 and AI agents

<div style="display: grid; grid-template-columns: repeat(4, 1fr); gap: 0.9em; margin: 0.6em 0 1em;">
<div class="callout">
<h3>2022</h3>
<p>ChatGPT</p>
<p class="muted" style="font-size: 0.8em;">a conversation with a model</p>
</div>
<div class="callout">
<h3>2024</h3>
<p>o1 and reasoning</p>
<p class="muted" style="font-size: 0.8em;">more computation before answering</p>
</div>
<div class="callout">
<h3>2025</h3>
<p>Coding agents</p>
<p class="muted" style="font-size: 0.8em;">Claude Code and Codex</p>
</div>
<div class="callout">
<h3>2026</h3>
<p>Fable 5.1 and GPT-6 Astra</p>
<p class="muted" style="font-size: 0.8em;">work across many steps</p>
</div>
</div>

- **With tools:** the model can be part of agents that search, run code and use the computer. Access and features depend on the product and setup. Can carry out tasks that take hours or days.
- **Several providers:** including OpenAI, Anthropic, Google and DeepSeek.

---

## Approaching human level on many tasks

<p class="muted">The figure shows results relative to a human baseline. Source: Stanford HAI, <a href="https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_2_technical.pdf">AI Index 2026, ch. 2</a> (<a href="https://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND 4.0</a>).</p>

---

## How long are the tasks AI can complete on its own?

![METR 50% time horizon by model release date, 2023–2026](../figures/metr-time-horizon.svg)

</div>
<div style="font-size: 0.72em;">

- **50% horizon:** the task length at which the model is estimated to succeed half the time.
- Time is measured in *human expert work*, not how long the agent runs.
- Well-specified tasks, mainly in software, machine learning and cybersecurity.
- Mythos Preview (early): about **17 hours** at 50%, against **3 hours** at 80%. The estimate above 16 hours is especially uncertain.

</div>
</div>

<p class="muted">Source: <a href="https://metr.org/time-horizons/">METR</a> and <a href="https://metr.org/blog/2026-1-29-time-horizon-1-1/">TH1.1</a>. Own figure of METR's published estimates; bars show 95% intervals.</p>

---

## An AI agent is a model in a work loop

![The agent loop: observe, plan, act, check – then continue or stop](../figures/agent-loop.svg)

<p style="font-size: 0.62em;">A fixed workflow follows predetermined steps. An agent can choose the next step based on intermediate results.</p>

</div>
<div style="font-size: 0.74em;">

- **Building blocks:** model, data, tools, memory and stopping rules. MCP is one open protocol for connecting data and tools.
- **Illustration for an economist:** read the accounts → compute deviations → investigate explanations → draft a memo.
- **Checkpoint:** reconcile figures against the sources, document assumptions and have a responsible person approve the deliverable.
- **Risk:** errors, prompt injection and overly broad permissions can propagate between steps.

</div>
</div>

<p class="muted">Source: Anthropic, <a href="https://www.anthropic.com/engineering/building-effective-agents">Building effective agents</a>; <a href="https://modelcontextprotocol.io/specification/2025-06-18/server/tools">MCP: Tools</a>. The economist example is an illustration.</p>

---

## The price of a given level of intelligence is falling fast

![Price per million tokens for the cheapest model at GPT-4 and GPT-3 level, 2021–2025](../figures/ai-price-decline.svg)

<p class="muted">Cheapest model reaching a fixed test level. Selected series from Epoch AI (through February 2025).</p>

</div>
<div style="font-size: 0.72em;">

- **GPT-4 level:** from USD 37.50 to 0.12 per million tokens in under two years – more than 300 times cheaper.
- Across test levels: price declines of 9 to 900 times per year, median around 50. History, not a guarantee for the future.
- **Locally:** a single gaming GPU costing under NOK 25,000 runs models that match the frontier from about 12 months earlier.

</div>
</div>

<div class="callout fragment">
<p><strong>What is frontier level today soon becomes cheap – and possible to run on your own machine.</strong></p>
</div>

<p class="muted">Source: Epoch AI, <a href="https://epoch.ai/data-insights/llm-inference-price-trends">LLM inference price trends</a> (March 2025) and <a href="https://epoch.ai/data-insights/consumer-gpu-model-gap">Frontier AI capabilities can be run at home</a> (Aug. 2025).</p>

---

## How far behind are Chinese open models?

<p class="kicker">Artificial Analysis Intelligence Index v4.3</p>

![Index points: Claude Fable 5.1 and GPT-6 Astra 53, GLM-5.3 45, Kimi K3 44](../figures/us-china-models.svg)

<p class="muted">Max setting.</p>

</div>
<div>
<div class="stat" style="text-align: left; padding-top: 0;">
<span class="stat-value" style="font-size: 2.4em;">About 15%</span>
<span class="stat-label" style="text-transform: none; letter-spacing: 0; font-family: var(--r-main-font); font-size: 0.62em;">lower score for the best Chinese open model than for the US frontier</span>
</div>
<div class="stat stat--blue" style="text-align: left;">
<span class="stat-value" style="font-size: 1.7em;">About 4 months</span>
<span class="stat-label" style="text-transform: none; letter-spacing: 0; font-family: var(--r-main-font); font-size: 0.62em;">average lag, open vs closed models. Epoch AI, January–May 2026.</span>
</div>
<p style="font-size: 0.62em;"><strong>Open weights:</strong> can be downloaded and self-hosted, but require substantial hardware.</p>
</div>
</div>

<p class="muted">Source: <a href="https://artificialanalysis.ai/leaderboards/models">Artificial Analysis</a>; Edwards and Emberson, Epoch AI, <a href="https://epoch.ai/data-insights/open-closed-eci-gap">Open models lag state-of-the-art closed models by 4 months</a>.</p>

---

## Young workers in AI-exposed occupations

<p class="muted">The figure shows 2021–2025 for two occupational groups. The number on the right comes from a broader analysis through June 2026.</p>

</div>
<div style="font-size: 0.72em;">

- **Stanford (Aug. 2026):** employment among 22–25-year-olds in highly AI-exposed occupations was about 19% below a path in line with less-exposed peers.
- The pattern appears to come mainly through fewer new hires. Experienced workers show no corresponding gap.
- **Not a causal estimate:** education, other market trends and the sample all affect the interpretation.

</div>
</div>

<p class="muted">Source: Stanford HAI, <a href="https://hai.stanford.edu/ai-index/2026-ai-index-report">AI Index 2026</a> (<a href="https://creativecommons.org/licenses/by-nd/4.0/">CC BY-ND 4.0</a>); Brynjolfsson, Chandar & Chen, <a href="https://digitaleconomy.stanford.edu/news/canariesaug26/">Canaries update</a>. US observational data.</p>

---

## “Math is solved?”

<div class="cols">
<div class="callout" style="font-size: 1em;">
<h3>The New York Times · 8 September 2026</h3>
<p>“OpenAI Says It Has Cracked One of Math’s ‘Millennium Problems’”</p>
<p><a href="https://www.nytimes.com/2026/09/08/science/openai-proof-millennium-problem.html">Read the article ↗</a></p>
</div>
<div class="callout" style="font-size: 1em;">
<h3>OpenAI · 8 September 2026</h3>
<p>“On the Navier–Stokes Millennium Prize Problem”: the company’s own account of the result</p>
<p><a href="https://openai.com/index/navier-stokes-solution/">Read OpenAI’s write-up ↗</a></p>
</div>
</div>

---

## In ten years?

1. **More tasks can be delegated to agents.** The value will depend on workflows, data and the cost of checking.
2. **Problem formulation and verification become more important.** Domain knowledge is needed to spot plausible errors and assess consequences.
3. **Entry into the professions may change significantly.** How do new economists, marketers and technologists learn when AI writes the first draft?