Leading US AI lab maintains a 3-9-month capability lead

Last updated

Tracker summary: We imagine the others to be 3-9 months behind OpenBrain (page 3, Late 2025). By Early 2026, several competing AIs match or exceed Agent-0 (page 6). The 3-9 month gap is the Late 2025 state; near-parity emerges by Early 2026.

Late-2025 public competition was closer than the predicted sustained three-to-nine-month lead. Undisclosed internal advantages remain unresolved.

At a glance

  • Assessment: Behind
  • Confidence in assessment: 85%
  • Outcome: unresolved
  • Timing: Original window passed; outcome unresolved
  • Evidence: proxy
  • Predicted timing: Late 2025
  • Primary source: AI 2027 archived scenario: Late 2025: The World’s Most Expensive AI, p. 3; Early 2026: Coding Automation, p. 6

Assessment scope: Behind (85%) concerns the fixed late-2025 three-to-nine-month lead, not a milestone that can be fulfilled whenever a lab later pulls ahead. Public competitive convergence is assessed separately from that lead and from a literal identification of the fictional Agent-0.

What AI 2027 Predicted

Tracker summary: The leading US lab maintains a three-to-nine-month advantage over competitors in late 2025. Evidence of a smaller lead therefore contradicts this particular prediction, even if it indicates faster competitive progress overall. AI 2027

Scenario source: archived version and page-number reference: Late 2025: The World’s Most Expensive AI, p. 3; Early 2026: Coding Automation, p. 6.

How We Track This

Compare contemporaneous capabilities on common evaluations, rather than equating release frequency with capability. Distinguish observable public competition from hidden internal work.

Current Evidence

Two comparisons: assessment on November 9, 2026

Retained Behind (85%) for the late-2025 lead. The dated ECI comparison below supports close public competition in late 2025, rather than the substantial sustained advantage forecast for that period. A later three-to-nine-month lead would be a later state, not late fulfillment of this fixed historical claim. Separately, public near-parity was already evident by late 2025, before the scenario’s early-2026 competitive-convergence stage. That is directional support for convergence, not proof that particular real models match or exceed the fictional Agent-0. Public point scores, specialized capabilities and undisclosed internal advantages remain different comparisons; the 23-day change of public leader is not a direct universal capability-lag estimate.

Epoch’s ECI table shows Gemini 3 Pro taking the public frontier on November 18, 2025, then GPT-5.2 Pro surpassing its point score 23 days later. Their year-end scores, 153.06 and 155.07 respectively, have overlapping 90% intervals. This supports close public competition rather than a clearly demonstrated sustained lead of several months. Epoch ECI table

The scenario’s authors independently judged the 2025 lead closer to zero to two months in their self-grading. That is the authors’ assessment, not an independent controlled measurement. AI Futures self-grading

Counterevidence & Limitations

An aggregate public index can obscure specialized strengths. Overlapping intervals do not prove exact equality, and public releases do not fully reveal internal capability. Those limits prevent a definitive claim about every lab’s private advantage.

Behind refers to the forecast advantage failing to appear clearly in public evidence. The previous interpretation treated narrower competition as confirmation, reversing the direction of the original claim. This correction preserves the evidence of convergence while assessing it against the stated prediction.

What Would Change Our Assessment

  • Strengthen the original claim: Dated, credible internal evidence establishes a sustained three-to-nine-month lead in late 2025.
  • Strengthen the shortfall assessment: Multiple comparable contemporary evaluations show persistent close competition across capability dimensions.
  • Clarify: Document which capabilities and public or internal frontier the claimed advantage describes.

Update History

DateUpdate
2026-11-09Assessment reviewed. Status: Behind. Confidence: 85%. Retained Behind (85%) for the fixed late-2025 lead. The dated ECI record supports public near-parity already in late 2025, separately from fictional Agent-0 equivalence. A later lead cannot fulfill the past state; public competitive convergence is directional support for the early-2026 stage, not a universal capability-lag measurement.
2026-11-07Assessment reviewed. Status: Behind; confidence: 85%. The historical ECI table shows Gemini 3 Pro taking the public frontier in November 2025 and GPT-5.2 Pro surpassing its point score 23 days later, with overlapping intervals. This supports close public competition rather than the scenario’s sustained three-to-nine-month lead. Public releases and aggregate scores do not exclude an undisclosed internal or specialized advantage. The September 7 assessment correction remains Emerging (65%) to Behind (85%). Earlier confirmation of a zero-to-two-month comparison did not confirm the scenario’s three-to-nine-month lead. Revenue is business context, not capability evidence; the unverified $30B/$24B annualized-revenue comparison and four-times-less training-spend claim are withdrawn from this assessment.
2026-09-07Reconciled close public competition with the direction of the original prediction. The dated ECI comparison supports a smaller public lead; this contradicts the forecast advantage, while internal gaps remain unresolved. 2026-11-07 · Correction: status Emerging (65%) → Behind (85%). The dated ECI comparison shows a smaller public lead than the forecast 3 to 9 months, which contradicts the predicted advantage.
2026-09-06Assessment revised from confirmed (0.90) to emerging (0.65). Release clustering and later competition do not resolve the original late-2025 three-to-nine-month capability lead.
2026-06-22xAI made Grok 4.3 available on Amazon Bedrock and Anthropic launched Fable 5. These releases reinforce the confirmed assessment that top US labs remain clustered near the frontier.
2026-04-13Anthropic surpassed OpenAI in revenue ($30B vs $24B ARR, April 2026). Anthropic achieved this while spending roughly 4x less on training (SaaStr, The AI Corner). The competitive gap between top US labs has fully closed — arguably inverted — with Anthropic, OpenAI, and Google trading frontier position every few weeks. Prediction firmly reinforced. 2026-11-07 · Correction: Revenue is business context only, not a measure of capability lead. The $30B/$24B annualized-revenue comparison and four-times-less training-spend claim are unverified and are withdrawn from the capability assessment. The capability claim must be tested with comparable dated evaluations, not revenue rankings. 2026-11-07 · Correction: this row assessed the later zero-to-two-month comparison, not confirmation of the scenario’s three-to-nine-month prediction.
2026-03Frontier race between OpenAI, Anthropic, and Google is extremely tight. The corrected 0-2 month gap is clearly borne out by alternating benchmark leads. 2026-11-07 · Correction: this row assessed the later zero-to-two-month comparison, not confirmation of the scenario’s three-to-nine-month prediction.
2026-02AI Futures grading explicitly conceded: “The race appears to be closer than we predicted, more like a 0–2 month lead between the top US AGI companies.” The scenario’s 3–9 month OpenBrain lead did not materialize.
2026-01AI Futures Project authors grade the inter-lab gap as 0-2 months, tighter than the 3-9 months predicted in AI 2027. AI Futures Project’s clarification post confirmed: “The race appears to be closer than we predicted, more like a 0-2 month lead between the top US AGI companies.”
2025-12GPT-5.2 released three weeks after Gemini 3 triggered an internal “Code Red” at OpenAI. The reactive cadence — Gemini 3 (Nov 18) → GPT-5.2 (Dec 11) — illustrates a gap of weeks, not months, between top US labs.
2025-11Gemini 3 and GPT-5.1-Codex-Max release on November 18; Claude Opus 4.5 releases on November 24 — six days later. Labs are trading benchmark leadership on a weekly basis. The AI 2027 prediction of 0-2 month gap between top labs appears confirmed. 2026-11-07 · Correction: this row assessed the later zero-to-two-month comparison, not confirmation of the scenario’s three-to-nine-month prediction.
2025-08GPT-5 (August 7), Gemini 2.5 Deep Think availability expansion (August 1), and Claude Opus 4.1 (August 5) all release within days of each other. Multiple frontier-class releases in the same week from competing labs.
2025-05Claude Opus 4 (May 22), Google Gemini 2.5 Flash GA (May 20), and OpenAI’s continued iteration: three major labs releasing competitive models within days of each other. The gap between top US labs is compressing.
2025-04Meta Llama 4 (April 5, Apache 2.0) and Alibaba Qwen3 (April 28) both produce open-weight models competitive with some closed frontier models. The competitive surface for frontier capabilities is widening beyond the top three US labs.