⚡ Key takeaways
- Most confident (85%): every frontier lab sells its own agent platform by mid-2027.
- Restricted tiers for cyber and bio become standard (75%); more labs publish self-improvement ratings (70%).
- Prices halve again (70%) while open models stay within six months of the frontier (60%).
- Most enterprise agent pilots still stall (65%).
- Labs are less likely to certify each other or run inside customers' data centres (55%).
🗺️ The predictions
Thesis 1: Ten claims, each falsifiable and dated.medium confidence
| # | Prediction | Confidence | Check by | Wrong if |
|---|---|---|---|---|
| 1 | Every frontier lab ships its own agent platform (harness + builder + runtime) to enterprises | 85% | 2027-06 | A frontier lab exits or spins out its agent platform |
| 2 | Open-weight models reach parity with the best closed model on a major coding benchmark within 6 months of each closed release | 60% | 2027-06 | The gap on the Benchmark Tracker widens two quarters in a row |
| 3 | At least two more labs publish a self-improvement capability rating for a released model | 70% | 2027-06 | No new self-improvement ratings in system cards by mid-2027 |
| 4 | Third-party evaluation becomes a routine purchase for EU general-purpose AI providers with systemic risk | 55% | 2027-12 | Obligations delayed or evaluations kept fully in-house |
| 5 | Price per million tokens for frontier-class models falls at least 50% again by end of 2027 | 70% | 2027-12 | List prices flat or rising across three labs |
| 6 | Restricted-access models for cyber and bio become a standard product tier at three or more labs | 75% | 2027-06 | Labs fold these back into general models |
| 7 | Agent-to-agent purchasing (an agent buys a service from another company's agent) appears in a major platform's GA product | 45% | 2027-12 | No GA agent-payment product by end 2027 |
| 8 | Most enterprise agent pilots still do not reach production in 2027 surveys | 65% | 2027-12 | Major surveys report majority in production at scale |
| 9 | A frontier lab acquires an evaluation, environment or agent-infrastructure start-up | 60% | 2027-12 | No such acquisition announced |
| 10 | Labs avoid on-premises deployment inside regulated customers' own data centres | 55% | 2027-12 | A frontier lab ships its own on-prem appliance |
🔧 Reproduce this: predictions.py
predictions.py · reads data/predictions-2027.json
"""predictions.py — the Interlude's ten predictions for 2027, ready to be scored later.
Reads data/predictions-2027.json. Prints each claim with its confidence and check date,
the expected number that come true if the confidences are honest, and how they will be
scored (Brier score = mean of (confidence - outcome)^2 once outcomes are known).
Standard library only.
"""
import json, os
d = json.load(open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "..", "predictions-2027.json")))
for p in sorted(d["items"], key=lambda p: (p["check"], -p["conf"])):
print(f"#{p['id']:2d} {p['conf']:.0%} by {p['check']} {p['claim'][:78]}")
exp = sum(p["conf"] for p in d["items"])
print(f"expected to come true if calibrated: {exp:.1f} of {len(d['items'])}")
done = [p for p in d["items"] if p["outcome"] is not None]
print(f"scored so far: {len(done)}" + (f", Brier {sum((p['conf'] - p['outcome']) ** 2 for p in done) / len(done):.3f}" if done else " (first checks due 2027-06)"))$ python3 predictions.py
# 1 85% by 2027-06 Every frontier lab ships its own agent platform (harness + builder + runtime)
# 6 75% by 2027-06 Restricted-access models for cyber and bio become a standard product tier at t
# 3 70% by 2027-06 At least two more labs publish a self-improvement capability rating for a rele
# 2 60% by 2027-06 Open-weight models reach parity with the best closed model on a major coding b
# 5 70% by 2027-12 Price per million tokens for frontier-class models falls at least 50% again by
# 8 65% by 2027-12 Most enterprise agent pilots still do not reach production in 2027 surveys
# 9 60% by 2027-12 A frontier lab acquires an evaluation, environment or agent-infrastructure sta
# 4 55% by 2027-12 Third-party evaluation becomes a routine purchase for EU general-purpose AI pr
#10 55% by 2027-12 Labs avoid on-premises deployment inside regulated customers' own data centres
# 7 45% by 2027-12 Agent-to-agent purchasing (an agent buys a service from another company's agen
expected to come true if calibrated: 6.4 of 10
scored so far: 0 (first checks due 2027-06)
So what: A forecast you cannot be wrong about is not a forecast; these will be scored in public.
📊 Where the labs are less likely to go
- Our reading.
- Inside customers' own data centres.
- Certifying other labs' models.
- Customer-specific test environments built over years.
- Sovereign deployments that require local control.
📄 Method and sources
- Confidences are our judgement; evidence comes from the site's sourced maps and the surveys in Brief 4.
- Scoring: Brier score, the mean of (confidence − outcome)² once outcomes are known.
- Outcomes are recorded in
data/predictions-2027.jsonand shown on the AI Market Map.
Facts are sourced; figures marked ⚠️ come from secondary reports or company statements. Scores, weights and predictions are our judgement and are labelled. Disclosure: Claude, made by Anthropic, helps write this site. Anthropic is one of the companies analysed here, under the same rules as the others.
🧭 Foundations: the AI market, Nov 2026
Live data: AI Market Map & Investments · AI Investment Tracker.
Read
Read
Read
Predictions for 2027
You are here