The AI Futures Project moved its Superhuman AI Researcher (SAR) forecast into 2027 on November 10, 2026, citing METR's high-access Anthropic acceleration study as the single largest update. The Q3.5 timelines post is explicit that modeling may contain errors, but the authors argue high-level conclusions are unlikely to collapse if METR's headline numbers hold.
If you follow OpenAI's AI Futures blog on power concentration or Sam Altman's pacing comments, this is the Daniel Kokotajlo / AI Futures side of the same conversation—not a lab product launch. Builders should read it as scenario planning, not a shipping schedule for your next fine-tune.
TL;DR
| Item | Claimed | Verified publicly |
|---|---|---|
| METR overall acceleration | ~2x since Oct 2025 vs prior trend | Partial — AECI slope chart published |
| AI share of acceleration | ~80% per METR | Weak — redacted supporting evidence |
| AI Futures derived pace | ~1.7x AI-attributable | Their math on METR headline |
| SAR median timing | 2027 (several models) | Scenario, not measurement |
| TED-AI before 2028 | 35–65% band | Subjective forecast |
| Anthropic alignment on 2x | Disputed in system card | Public text, not full data |
What triggered the update
METR spent August 12 to October 29, 2026 inside Anthropic with unusual data access: pretraining metrics, RL pipelines, compute accounting, internal benchmark scores, staff interviews, and forecasts. Their bottom line: Anthropic is likely improving on well-specified tasks at about twice the rate seen from March 2024 through October 2025, with wide uncertainty (1.3x to 3x plausible).
They also judge that a majority of today's acceleration is downstream of AI use in R&D, not only more GPUs—while admitting other factors mattered more in late 2025.
The one chart many outsiders can actually inspect compares AECI—Anthropic's internal take on the Epoch Capabilities Index family. Post-October 2025 points follow a steeper line than the pre-break trend. METR says a one-time jump fits worse than a slope change once newer points land.
AI Futures reads that package and asks: if 1.7x AI-attributable pace is real in November 2026, where does Superhuman AI Researcher land?
Definitions that matter for builders
Superhuman AI Researcher (SAR) in AI Futures' writeup is not "GPT passes more LeetCode." It is an Automated Coder that also matches top-human research taste on an AGI project—picking directions, designing experiments, reading results.
That is stricter than AI R&D parity, where a lab would rather keep AIs than average human researchers. AI Futures uses SAR because they think it may mark a point-of-no-return for misalignment risk if recursive self-improvement kicks in—themes we tracked when Claude Fable 5.2 met METR's RSI scrutiny.
Top-Expert-Dominating AI (TED-AI) is broader: systems that beat top humans at virtually all white-collar work. AI Futures puts a 35–65% chance before 2028, with the same METR-driven update.
Why 1.7x pace is not the same as 12x "feel faster at work"
A common misread collapses three quantities:
- Labor uplift — how much faster researchers feel with copilots (METR cites informal 12x self-reports on Anthropic's best internal model).
- Software R&D uplift — effective algorithmic progress, capped by finite GPUs.
- Pace of overall progress — capabilities the world sees, including compute scaling.
AI Futures uses a car on a hill analogy: doubling road speed requires more than doubling engine effort because the hill steepens (diminishing returns) and the car still needs fuel (compute). A 1.7x pace acceleration is therefore a bigger deal than making staff 1.7x more productive.
That distinction matters for engineering managers quoting METR on Slack. Your team's Cursor throughput is not frontier capability slope.
What AI Futures' models now say about 2027
The November 10 post runs several extrapolations:
- Research taste linear extrapolation on accelerated AECI → SAR around late 2027.
- Filtered AI Futures Model conditioned on 1.7x November acceleration → SAR median near February 2027.
- Tom Davidson toy model → superintelligence roughly a year after a software-only singularity assumption.
They also explore why their August model underestimated recent speed, listing hypotheses like near-Automated-Coder skills lifting R&D more than expected, non-coding skills (data generation) mattering, or overly harsh diminishing-return assumptions.
The authors publish all-things-considered medians per forecaster for SAR and TED-AI. The post warns that median SAR minus median TED-AI is not the median gap between milestones because heavy right tails skew the arithmetic.
What is not verified — and why AI Futures is angry about redactions
METR states plainly that Anthropic redacted a majority of the evidence behind the 2x and ~80% AI figures. Public readers get narrative, interviews, and the AECI slope—not the underlying tables that convinced METR.
AI Futures calls that "very unfortunate" and asks Anthropic to release enough detail for external audit. They note Anthropic's Fable 5.2 system card language that Mythos 5.2 does not double aggregate internal R&D progress—wording that does not clearly endorse METR's company-wide 2x claim.
For explainx.ai readers, this is a claimed-vs-verified post in the governance lane:
- Verified: METR published a dated assessment, names methods, and shows at least one capabilities slope figure.
- Claimed but opaque: The precise mix of evidence behind 80% AI attribution and some acceleration subcomponents.
- Scenario output: AI Futures' 2027 SAR distribution is transparent modeling on partly secret inputs.
We are not saying the forecast is propaganda. We are saying your team's compliance spreadsheet should not treat February 2027 as a milestone with engineering sign-off.
Policy implications vs. shipping implications
AI Futures argues that if SAR is plausible in 2027, labs are not on track to align such a system, and recursive self-improvement with a misaligned SAR is an extreme failure mode. They advocate pacing or industry-wide pause ideals—overlapping Anthropic's pacing-not-pausing IPO rhetoric and the Pacing the Frontier employee letter debate on explainx.ai.
For practitioners, the nearer-term hooks are:
- Document what your agents automate today versus what METR counts as R&D acceleration.
- Treat internal capability indices like AECI as vendor metrics unless you reproduce them on your workload.
- Expect governance fights to cite this post in procurement reviews—similar to how METR remote labor estimates entered enterprise agent ROI slides.
How this differs from our prior METR coverage
explainx.ai already covered METR's October assessment headline in the context of Anthropic's system card and RSI evaluations. This November 10 AI Futures post is the timelines community's Bayesian update—a second lens on the same METR event, not a new primary measurement.
If you need a builder-first frame: assume capability releases continue on public API cadence; use AI Futures for risk registers and eval breadth, not for sprint planning dates.
Comparison to the AI 2027 scenario and public lab statements
AI Futures estimates METR's November numbers sit about five months ahead of the AI 2027 storyline on 1.7x AI-attributable pace—not a small shift when your safety case assumes gradual uplift.
They contrast that with AI 2027 predicting 2x software R&D uplift in December 2026, which would imply slower headline progress than METR's November read. The disagreement is not about whether Anthropic shipped Claude Fable 5.2; it is about how much of recent slope is AI loop versus compute step-ups versus organization learning.
Anthropic's public Mythos 5.2 wording in the system card denies that one external-facing model doubles aggregate internal R&D progress. METR's corporate-level ~2x claim is broader. Outsiders cannot adjudicate without the redacted exhibits—exactly why AI Futures asks for release.
For builders, the lesson is familiar from claimed-vs-verified Vertex Carbon posts: headline capability marketing and third-party evaluations can diverge while both cite partial charts.
A pacing checklist teams can use without waiting for SAR
You do not need to pick a side in the 2027 debate to improve posture this quarter:
Map what "R&D" means in your org. METR counts pretraining, RL, eval design, and tooling. Your platform team may count only fine-tuning customer adapters. Align vocabulary before copying acceleration multiples into board slides.
Separate copilot uplift from capability slope. Track PR throughput, incident rate, and eval regressions independently. A 12x "feel faster" Slack poll does not imply your public model jumped an ECI step.
Publish eval coverage when you cannot publish weights. If suppliers redact evidence, customers should escalate reproducible eval suites—the same ask AI Futures makes of Anthropic.
Document agent automation boundaries. Treasury agents on AgentCore and coding agents in CI both need RBAC narratives for auditors.
Watch policy forks. Paris Call shutdown language, AI Overwatch GPU access bills, and pacing letters will cite METR-style reports whether or not SAR arrives on schedule.
What developers are asking
Does SAR in 2027 mean coding agents replace staff next quarter? No. SAR is a hypothetical AGI-project role, not "Copilot writes all PRs."
Should I freeze model upgrades? The post argues for policy-level pacing, not individual repo freezes. Your security and eval gates still matter—see OpenAI's control room kill switch story for an operational analog.
Is METR's 12x self-report credible? Treat it as directional culture signal. METR itself labels interviews hard to interpret.
Related reading
- Claude Fable 5.2 and METR RSI system card coverage
- Remote labor index and automation framing
- OpenAI AI Futures on power concentration
- Sam Altman on sandboxing superhuman models
- Anthropic pacing not pausing around IPO
- Paris Call and fail-safe shutdown debate
- Anthropic homeopathic learning and training mixtures
Official sources: AI Futures Q3.5 update (Nov 10, 2026), METR Anthropic acceleration assessment (Nov 5, 2026).
Forecasts and redaction status accurate as of November 10, 2026.