CHOOSELYIndependent capability measurement
Choosely AI Progress Index
How close AI is to doing complex digital work on its own.
A conservative, evidence-backed measure of what today’s best AI can actually do — not what companies say it can do.
Reviewed 14 October 2026. The Index places frontier AI at 50.8 out of 100 across reasoning and adaptation, real-world work, and autonomy and agency.
Verified capability—not marketing claims.
The capability balance
The shape of AI capability
Three scores. One evolving form.
Reasoning develops the head. Real-world work develops the arms and hands. Autonomy develops the torso, spine and legs. Each system’s surface reveal and lights respond automatically to its score. Compare approved snapshots, replay the change, then read the precise balance below.
The form is a visual metaphor for the three measured capabilities. Its surface reveal and lights use exact scores; labels round to one decimal.
Approved
What moved
| Pillar | Previous | Current | Change |
|---|---|---|---|
| Reasoning & Adaptation | 78.9 | 78.9 | Unchanged |
| Real-World Work | 25.7 | 28.9 | +3.2 pts |
| Autonomy & Agency | 46.2 | 46.2 | Unchanged |
Lowest pillarReal-World Work28.9 out of 100
Each point moves outward as its pillar score rises. A lower score pulls the profile toward the center.
As the Index changes, the shape changes with it.
What is driving the score?
Three ways to understand where AI is now.
Reasoning & Adaptation
Can AI figure out difficult problems and adapt when things change?
Real-World Work
Can AI reliably finish useful work from start to finish?
Autonomy & Agency
Can AI keep working on its own without constant human help?
What has actually been counted?
The Live Edge
Choosely separates proven capability from promising claims and unanswered questions.
ARC-AGI-3 Semi-Private rises to 62.71% under the provider-neutral Standard harness.
ARC-AGI-3 · Verified evidenceARC-AGI-2 reaches 95.0% on the frozen Semi-Private surface.
ARC-AGI-2 · Verified evidenceAI can attempt work that would take human experts almost 12 hours, but uncertainty rises as tasks get longer.
METR · Verified evidenceChecked before they can change the Index
CHECKINGNot counted without independent evidence
EXCLUDEDThis first snapshot is where the series begins
OPENNot published until defensible evidence exists
OPENWhat is holding AI back?
What AI still can’t do.
Impressive demos are not the same as dependable autonomous work. These are the clearest gaps in the frozen evidence.
Make interactive adaptation dependable
ARC-AGI-3 has advanced sharply, but 62.71% still leaves substantial failure on hidden-rule environments.
ARC-AGI-3 · 62.71%Deliver client-acceptable work
Most professional projects still need human work before independent reviewers call the result client-ready.
RLI · 21.25%Finish computer workflows cleanly
AI often makes useful progress on computer tasks but fails to complete every required step.
OSWorld 2.0 · 20.6% binary accuracyPass broad agent tasks in full
Broad agent tasks only count when the whole job is finished correctly; partial progress does not pass.
ALE · 38.2% full pass rateHow did capability reach this point?
How AI got here.
A few capability shifts changed what AI systems could plausibly do next. This is context—not fabricated Index history.
- GPT-4 launches
Broad multimodal reasoning advances
- Function calling
Models begin using external tools
- Claude 3 Opus
Long context and robust analysis
- GPT-4o
Real-time multimodal interaction
- Agentic systems
Autonomous step-by-step workflows
- AI Progress Index V1.0
First canonical measurement
- Reliable autonomy
The work ahead
Why trust the measurement?
What the Index is based on.
Six independent tests. Exact results are frozen before they can influence the headline score.
95.0%
AI can solve most difficult unfamiliar static puzzles on the frozen semi-private surface.
ARC-AGI-2 ↗62.71%
AI can now solve a majority of the semi-private interactive adaptation surface under ARC Prize’s provider-neutral Standard harness.
ARC-AGI-3 ↗21.25%
Most professional projects still do not reach a client-acceptable finish without human help.
RLI ↗38.2% full pass rate
AI completes more than one in three broad professional agent tasks in full on the frozen ALE-V1 surface.
ALE ↗~11h 59m at 50% success
AI can now tackle tasks that take human experts hours, but reliability falls as work gets longer.
METR ↗20.6% binary accuracy
AI often makes partial progress across computer workflows but rarely finishes them cleanly.
OSWorld 2.0 ↗The Change Brief
Follow the frontier as it evolves.
Weekly intelligence on the AI changes that actually matter.
By submitting, you’re asking to receive The Change Brief by email. Unsubscribe anytime.