Forecast Industry

What AI Will Do by 2030

The heads of Anthropic, OpenAI and Google are publicly naming dates for human-level AI — and these are no longer futurism but corporate plans with budgets attached. Here is what the companies promise on the record, which promises the data supports, and the signals that will tell you which scenario is unfolding.

Author: Alexey VoroninPublished: Aug 16, 2026Reading: 11 min

TL;DR

  • Dario Amodei (Anthropic) put a date on "powerful AI" in an official White House filing (March 2025): late 2026 or early 2027.
  • Sam Altman declared in "The Gentle Singularity" (June 2025) that "we are past the event horizon": 2026 — systems that find novel scientific insights, 2027 — robots doing real-world tasks.
  • The largest survey of AI researchers (2,700+) is far more sober: a 10% collective probability that AI outperforms humans on most tasks by 2027.
  • The one measured trend: the length of tasks models can complete autonomously doubles roughly every seven months (METR) — at that pace, working days rather than hours by 2028.
01 —

What the Companies Promise on the Record

Forecasts have become checkable: every lab has public documents, founder essays and a measurable release cadence. The plans of the four key players — from primary sources, not retellings.

Anthropic

"Powerful AI" by 2027 and the Mythos-class pipeline

The most specific position on the market. In "Machines of Loving Grace" (October 2024) Amodei described the "compressed 21st century" — a hundred years of scientific progress in five to ten — and Nobel-caliber AI by 2027. In March 2025 the company repeated the date in an official White House filing: "we expect powerful AI systems will emerge in late 2026 or early 2027." Summer 2026 showed what that looks like in practice: four releases in two months, the opening of the Mythos class (Fable 5), and a mechanism for shipping above-Opus models behind safety classifiers. The cadence: a new model every four to eight weeks.

OpenAI

"Past the event horizon"

Altman staked the position in two essays. "Reflections" (January 2025): "we are now confident we know how to build AGI", with the first agents "joining the workforce." "The Gentle Singularity" (June 10, 2025) opens with "we are past the event horizon; the takeoff has started" and lays out a calendar: 2026 — systems capable of novel scientific insights, 2027 — robots doing tasks in the physical world, the 2030s — intelligence and energy becoming "wildly abundant." By mid-2026 Altman had softened the phrasing to "AI surpasses human intelligence by 2030."

Google

Quieter words, louder infrastructure

DeepMind rarely names dates — Hassabis has historically been the most cautious of the three. The product strategy speaks instead: a split between a reasoning flagship (Gemini 3.1 Pro, Deep Think) and a cheap fast layer (Flash 3.5/3.6) for mass agent loops. The bet is not a record-setting single model but AI embedded in Search, Workspace and Android — that is, in billions of devices.

Open

Catching up in months, not generations

Meta assembled a dedicated unit with superintelligence as its stated goal. Chinese labs (Z.ai with GLM-5.2, DeepSeek with V4) have cut the open-weight gap to the closed frontier to months — by their own developers' estimates. For forecasting this means: any flagship capability becomes commodity within six to twelve months, including on your own hardware.

Counterpoint: in the largest survey of AI researchers (2,700+ participants), the collective probability that AI outperforms humans on most tasks by 2027 came out at 10%. And the authors of the famous "AI 2027" scenario began walking back their own timelines by late 2025. "AGI in two years" is the position of executives selling AI — worth keeping in mind.
02 —

Five Capabilities: Measured vs Promised

Trend ✓

Long autonomy

The only capability with a measured curve: per METR, the length of tasks models complete autonomously at acceptable reliability doubles roughly every seven months. Today's flagships hold for hours of work; at this pace, working days by 2028. This is exactly where all three labs are aiming — Frontier-Bench and GDPval-AA measure precisely this.

Trend ✓

The agent as a team, not a tool

Already in products: Fable 5 spins up parallel sub-agents for large code migrations; GPT-5.6 Sol Ultra is effectively a multi-agent mode. By 2027–2028 the expected norm is "orchestrator plus a dozen workers" instead of one chat: you set the goal, the system handles decomposition and verification.

Promise

Scientific discovery

Altman's calendar puts "systems finding novel insights" in 2026; Amodei speaks of Nobel caliber by 2027. There is no confirmed case yet of a major discovery made autonomously — there is acceleration of human scientists. Honest status: plausible, unmeasured.

Promise

The physical world

"Robots in 2027" is the boldest part of the forecasts. The software half (planning, vision, instructions) is improving fast; the hardware half is not — the price and reliability of manipulators follow hardware laws, not LLM curves. Betting on mass robotics before 2030 is faith, not arithmetic.

Trend ✓

Memory and price

Boring but most consequential: flagship context grew to a million tokens, GPT-5.5-level pricing halved within one generation (Terra at $2.50/$15), caching discounts reach 90%. By 2028 "an assistant that remembers the company's entire history" stops being a project and becomes a pricing tier.

03 —

What the Forecasts Leave Out

Three factors have already bent the beautiful curves. First, government: 2026 saw the first state-adjusted release schedules in industry history — the Fable/Mythos 5 pause and GPT-5.6's restricted start at the White House's request (full breakdown in our regulation analysis). Any "model X ships in quarter Y" now contains a political variable. Second, inference economics: the promises assume compute keeps getting cheaper; the $690 billion the industry committed to infrastructure for 2026 alone is a bet, not a guarantee, and power grids are already the bottleneck. Third, measurability: classic benchmarks are saturated, agentic ones are young and disagree with each other (we covered the Endor Labs vs Fable 5 case in our rankings methodology). "Progress" is ever harder to reduce to one number — and therefore to verify.

03b —

How the Products Themselves Will Change

Superintelligence forecasts are abstract; interface changes are concrete and already visible in betas. First, chat stops being the main window: instead of dialogue — delegation. You describe the outcome, set a token budget and a deadline; the agent returns with finished work and a decision log. Opus 5's effort dial is an early version of that future — quality and cost become a slider, not a model property. Second, memory becomes the product battleground: million-token context technically lets an assistant hold a project's entire history; the constraint shifts from capability to trust and data-retention regimes. Expect tiers split not by "smartness" but by memory depth and privacy guarantees. Third, price stratification accelerates: Sol/Terra/Luna and Fable/Opus/Sonnet are the same logic — routing tasks by cost, with the expensive model firing on a small share of requests. By 2028 a typical product will transparently push 90% of traffic through the cheap layer — the main reason "AI in every button" becomes economically possible. For users it adds up to one feeling: AI stops being a place you go and becomes a property of all software, the way search and sync once did.

04 —

Three Scenarios to 2028 — and Their Signals

Base
~60%

Fast saturation of work tasks, no "jump"

Models keep gaining capability and shedding cost; agents close ever-longer tasks; AGI rhetoric fades (the shift is visible since late 2025). Signals: releases every 1–2 months, autonomy tracking the METR curve, government interventions staying targeted.

Fast
~20%

The promises land on schedule

By late 2027 — systems running multi-day projects alone and visible scientific results. Watch for: a public discovery made by a model with no human in the loop; autonomy jumping faster than the seven-month doubling; new Mythos releases shipping without pauses.

Slow
~20%

A wall: data, energy or the regulator

The curve breaks on quality-data scarcity, compute costs, or hard regulation after an incident. Signals: longer gaps between flagships, postponed announced releases, API prices rising instead of falling.

Editorial position: the probabilities above are our subjective estimate, not a model output. We will update this page as the observable signals move — the freshness date sits at the top. What all this means for jobs and wages is covered separately: "AI and the Labor Market".
05 —

Questions We Get

So is AGI coming by 2027 or not?
Depends on the definition — which is half the problem. "A system that beats humans at most office tasks" — probably close: the autonomy trend is on its side. "A universal intelligence replacing any specialist" — the research consensus is far humbler than the CEO statements. Our advice: track the measurable, not the word AGI — autonomous task length and the share of work agents actually close.
Whose forecasts should I trust?
Nobody's entirely. The rule of thumb: lab CEOs mark the upper bound (they need to sell and hire), researcher surveys the lower bound (science is conservative), measured trends like METR's are the most reliable thing available. Historically the truth landed in between: faster than the skeptics, slower than the visionaries.
What does this mean for an ordinary company right now?
Do not wait for "real AGI" — build processes so every new model tier slots in without a rebuild: prompts, tests and data kept separate from the vendor. The difference between companies in 2028 will come not from model access — that is equal for everyone — but from process readiness. Where to start: the adoption plan.

Sources: "Machines of Loving Grace" (D. Amodei, Oct 2024), Anthropic's White House filing (Mar 2025), "Reflections" and "The Gentle Singularity" (S. Altman, Jan and Jun 2025), METR task-horizon research, the AI Impacts survey (2,700+ researchers), Forbes "10 AI Predictions for 2026", public 2026 announcements by Anthropic, OpenAI and Google. Scenario probabilities are the editors' estimate.