We can think of GAIA as a stress test for "general assistants" that must do what humans casually do all day: find the right source, read it correctly, combine a few steps, and give a precise answer. Not just "reason" in the abstract, but execute: browse, extract, verify, and finish cleanly.
When we plot agent effectiveness on GAIA against time and capability, it naturally forms an S-curve.
At the beginning of the curve, plain LLM behavior doesn't help much. We can write plausible text, but GAIA tasks punish plausibility. Without disciplined tool use, the system either can't reach the needed information or can't assemble it reliably. Improvements in prompting and base reasoning move the needle, but not dramatically, because the failure mode is execution, not eloquence.
Then the middle of the curve arrives and the slope gets steep. This is where tool use and orchestration show up: search, browsing, structured extraction, multi-step planning, retries, and basic self-checks. Once an agent can consistently do "find → read → compute → answer" instead of guessing, accuracy jumps fast. This is the part of the S-curve that feels like progress is suddenly compounding.
Finally we hit the plateau. Not because we stopped improving, but because the remaining errors live in the long tail. The last few percentage points aren't about doing the common case better. They're about not breaking on messy pages, not selecting the wrong source when multiple are plausible, not misreading a table in a PDF, not dropping a constraint halfway through, and recovering when an early step goes wrong.
On GAIA specifically, the human baseline is roughly in the low 90s. The best agent systems on the public leaderboard are now essentially there as well: the overall average is in the ~91–92% range. In other words, in GAIA terms, we're already in the top-right of the chart: phase three, near the asymptote.
Here's the mental picture:
Effectiveness (GAIA %)
100 | ________ human ~92%
95 | _____/
90 | _____/ we are here (SOTA ~91–92%)
85 | _____/
80 | _____/ steep gains: tools + orchestration
70 | ___/
60 | __/
50 | _/ early: LLM-only struggles on execution
40 | _/
30 |/
0 +---------------------------------------------------------
Phase 1 Phase 2 Phase 3
(no tools) (tools+agentic) (robust autonomy)
What changes once we're on that plateau is the nature of work that matters. Benchmarks become less about average score and more about reliability characteristics: variance, tail risk, and "cost to correct." The question stops being "can we solve this kind of task?" and becomes "how often do we fail in annoying ways, and how expensive is it for a human to catch and fix it?"
This also explains why Level 1 tasks look "basically solved" while harder levels still leak errors. The limiting factor isn't raw intelligence; it's robustness under ambiguity and messy real-world inputs. A system can be brilliant and still pick the wrong page. It can reason correctly and still extract the wrong number from a table. It can follow a plan and still silently drop a constraint.
So where are we in the chart? We're already at the part where gains come from boring, high-leverage engineering: verification loops, provenance discipline, better fallback strategies, and tight recovery when the first attempt goes off the rails. Once the mean is near-human, the differentiator is not "more IQ." It's fewer unforced errors.