METR’s Anthropic Report Turns AI-Led Acceleration Into a 1.9x Slope Ratio That Reproduces From Its Charts, but the Ratio Runs From 1.1x to 2.9x With the Fit Function, and a Second METR Report Says Anthropic’s Public Risk Report Understated the Pace

On November 5, METR published an independent assessment of whether AI is speeding up Anthropic’s own AI research, the product of an 11-week engagement with access to internal data. The Verge’s Robert Hart covered it under the headline “Why a new report on Anthropic automating AI development is alarming researchers,” and led with a number that is easy to remember: Anthropic’s models are “improving roughly twice as quickly as a year ago.” In METR’s report that number is 1.9x, and it is the ratio of two slopes drawn through a few dozen dots on a chart that METR published.

I read the Verge story beside both METR reports from that day, Anthropic’s own pages and its Responsible Scaling Policy, and then re-derived the headline from the charts. The Verge’s account is accurate on most points, including the redaction fight and the cancelled compute check. Three things are missing from it. First, the 1.9x reproduces from the chart and survives dropping any one model, but it is a ratio whose fit and denominator are choices: it runs from 1.1x to 2.9x across the fit functions METR tried, is 1.4x when other labs’ models are included, and is 1.5x against the fastest earlier stretch of three consecutive snapshots. Second, “mostly AI” is a guess of about 80% that rests on evidence METR says Anthropic redacted. Third, METR’s companion report concludes that Anthropic’s public August Risk Report understated the acceleration, which the Verge piece does not mention.

One limit up front: I read the plotted points off METR’s chart images, to within about 0.2 index points, so what follows tests METR’s fit, not its underlying benchmark scores, which I could not find published by METR or Anthropic.

What the Verge reported, against what METR wrote

The Verge’s quoted spans from METR and from Anthropic co-founder Sam McCandlish, 20 of them, all match their sources; the story’s other quotations come from other people and I did not check them. The differences are in what the paraphrases leave out. METR’s two reports are its Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 and its review of the “Risks from automated R&D” section of Anthropic’s August 2026 Risk Report.

The 1.9x is a ratio of two slopes, and it reproduces

How the ratio is built

AECI is Anthropic’s fork of Epoch AI’s Epoch Capabilities Index: one number per model built from many benchmark scores, with Claude 3.5 Sonnet anchored at 130. METR fits a line with one break through the frontier snapshots by ordinary least squares (on its chart the two segments join at the break), estimates the break at October 7, 2025, and reports the ratio of the two slopes. Each snapshot is dated by when most Anthropic employees could use it, which puts unreleased models on the same axis as public ones. The chart prints +15 a year before the break and +29 a year after. Twenty-nine over fifteen is the 1.9x, and a year at +29 buys about what two years bought at +15, which is the “roughly two years of historical progress over the last year” sentence. A second chart that adds other developers’ public models prints +17 and +24, which METR labels 1.4x.

Re-deriving it from the charts

I downloaded METR’s two chart images, calibrated both axes from the tick labels (the fit error was 0.02 index points vertically and 1 to 2 days horizontally), and located the centre of every solid dot by colour. The script below carries those readings, refits the one-break line, and runs the checks METR did not print. It uses only the Python standard library.

"""refit_aeci.py: refit METR's one-break AECI trend on points read off its published charts.
Standard library only (Python 3.10 or later). Run: python refit_aeci.py"""
from datetime import date, timedelta

# (name, date the snapshot was available to most Anthropic employees, AECI), read off METR's chart; solid dots only
ANTHROPIC = [
    ("3 Opus", "2024-02-29", 124.78), ("3.5 Sonnet", "2024-06-16", 130.05), ("3.5 Sonnet New", "2024-10-10", 133.26),
    ("3.7 Sonnet", "2025-02-19", 141.76), ("Opus 4", "2025-05-19", 144.85), ("Sonnet 4.5", "2025-09-22", 147.82),
    ("Opus 4.5", "2025-11-19", 152.38), ("Opus 4.6", "2026-01-24", 157.32), ("Mythos Pre. Early", "2026-02-24", 162.39),
    ("Mythos Pre.", "2026-04-01", 165.72), ("Mythos 5", "2026-06-01", 168.67), ("Model 2 Early", "2026-07-10", 171.82),
    ("Model 2", "2026-08-07", 174.12), ("Model 3 Early", "2026-09-09", 176.43), ("Model 3", "2026-09-22", 177.04),
    ("Model 4", "2026-09-25", 178.70),
]
# public Anthropic models that are not on the frontier because an internal model is ahead of them (hollow dots)
HOLLOW = [("Opus 4.7", "2026-04-11", 160.34), ("Opus 4.8", "2026-05-19", 163.46), ("Opus 5", "2026-07-17", 169.39),
          ("Mythos 5.1", "2026-08-07", 172.56), ("Opus 5.5", "2026-09-09", 174.72)]
# the running record across developers (second METR chart)
ALL_DEV = [
    ("GPT-4o", "2024-05-14", 127.28), ("3.5 Sonnet", "2024-06-21", 130.20), ("o1-mini (high)", "2024-09-13", 138.13),
    ("o1 (high)", "2024-12-18", 144.01), ("Gemini 2.5 Pro Exp", "2025-03-25", 146.76), ("o3 (high)", "2025-04-16", 149.27),
    ("grok-4-0709", "2025-07-09", 149.46), ("GPT-5 (high)", "2025-08-07", 152.05), ("GPT-5 Pro", "2025-10-06", 153.45),
    ("Gemini 3 Pro Preview", "2025-11-18", 154.70), ("GPT-5.2 Pro", "2025-12-11", 158.47),
    ("Gemini 3 Flash Preview", "2025-12-17", 159.15), ("Gemini 3 Deep Think", "2026-02-12", 161.28),
    ("Mythos Pre. Early", "2026-02-24", 162.53), ("Mythos Pre.", "2026-04-01", 165.86), ("Mythos 5", "2026-06-09", 168.80),
    ("Model 2 Early", "2026-07-11", 171.97), ("Model 2", "2026-08-08", 174.26), ("Model 3 Early", "2026-09-10", 176.56),
    ("Model 3", "2026-09-23", 177.17), ("Model 4", "2026-09-26", 178.83),
]
BREAK = date(2025, 10, 7)  # the breakpoint METR reports for its main specification


def yrs(d):
    return (date.fromisoformat(d) - date(2024, 1, 1)).days / 365.25


def solve(A, b):
    n = len(b)
    M = [row[:] + [b[i]] for i, row in enumerate(A)]
    for c in range(n):
        p = max(range(c, n), key=lambda r: abs(M[r][c]))
        M[c], M[p] = M[p], M[c]
        for r in range(c + 1, n):
            f = M[r][c] / M[c][c]
            M[r] = [x - f * y for x, y in zip(M[r], M[c])]
    x = [0.0] * n
    for r in range(n - 1, -1, -1):
        x[r] = (M[r][n] - sum(M[r][k] * x[k] for k in range(r + 1, n))) / M[r][r]
    return x


def ols(rows, ys):
    k = len(rows[0])
    A = [[sum(r[i] * r[j] for r in rows) for j in range(k)] for i in range(k)]
    b = [sum(r[i] * y for r, y in zip(rows, ys)) for i in range(k)]
    beta = solve(A, b)
    sse = sum((y - sum(c * v for c, v in zip(beta, r))) ** 2 for r, y in zip(rows, ys))
    return beta, sse


def hinge(points, brk):
    """Continuous piecewise-linear fit with one break: returns (slope before, slope after, sum of squared errors)."""
    t0 = yrs(brk.isoformat())
    rows = [[1.0, min(yrs(d) - t0, 0.0), max(yrs(d) - t0, 0.0)] for _, d, _ in points]
    beta, sse = ols(rows, [a for _, _, a in points])
    return beta[1], beta[2], sse


def learned_break(points, step=1):
    best, d = None, date(2025, 1, 1)
    while d <= date(2026, 3, 1):
        s1, s2, sse = hinge(points, d)
        if best is None or sse < best[0]:
            best = (sse, d, s2 / s1)
        d += timedelta(days=step)
    return best


def slope(points):
    return ols([[1.0, yrs(d)] for _, d, _ in points], [a for _, _, a in points])[0][1]


b1, b2, sse = hinge(ANTHROPIC, BREAK)
print("1. Refit with the break fixed at", BREAK)
print(f"   Anthropic frontier, {len(ANTHROPIC)} snapshots: {b1:.2f}/yr -> {b2:.2f}/yr, ratio {b2 / b1:.2f}   (METR prints +15, +29, 1.9x)")
a1, a2, _ = hinge(ALL_DEV, BREAK)
print(f"   All developers, {len(ALL_DEV)} records:      {a1:.2f}/yr -> {a2:.2f}/yr, ratio {a2 / a1:.2f}   (METR prints +17, +24, 1.4x)")

best = learned_break(ANTHROPIC)
print(f"2. Break learned by a daily scan from 2025-01-01 to 2026-03-01: {best[1]}, ratio {best[2]:.2f}   (METR reports 2025-10-07)")
print("   ratio at other break dates: " + " | ".join(
    f"{d} {hinge(ANTHROPIC, date.fromisoformat(d))[1] / hinge(ANTHROPIC, date.fromisoformat(d))[0]:.2f}"
    for d in ("2025-05-01", "2025-07-01", "2025-09-01", "2025-11-01", "2025-12-01", "2026-01-15")))

fixed, relearned = [], []
for i, (name, _, _) in enumerate(ANTHROPIC):
    rest = ANTHROPIC[:i] + ANTHROPIC[i + 1:]
    s1, s2, _ = hinge(rest, BREAK)
    fixed.append((s2 / s1, name))
    relearned.append((learned_break(rest, step=2)[2], name))
lo, hi = min(fixed), max(fixed)
print(f"3. Drop one snapshot at a time (16 refits), break fixed: ratio {lo[0]:.2f} (without {lo[1]}) to {hi[0]:.2f} (without {hi[1]})")
print(f"   break re-learned each time:                        ratio {min(relearned)[0]:.2f} (without {min(relearned)[1]}) to {max(relearned)[0]:.2f} (without {max(relearned)[1]})")

pool = sorted(ANTHROPIC + HOLLOW, key=lambda p: p[1])
record, top = [], 0.0
for p in pool:  # keep only public snapshots, and of those only the ones that set a new public record
    if p[0] in {h[0] for h in HOLLOW} or (p[0] in {a[0] for a in ANTHROPIC} and not p[0].startswith(("Mythos Pre", "Model"))):
        if p[2] > top:
            record.append(p)
            top = p[2]
p1, p2, _ = hinge(record, BREAK)
print(f"4. Public Anthropic models only ({len(record)} record-setters, no internal snapshots): {p1:.2f}/yr -> {p2:.2f}/yr, ratio {p2 / p1:.2f}")

pre = [p for p in ANTHROPIC if date.fromisoformat(p[1]) < BREAK]
print("5. Fastest stretch before the break, any run of consecutive frontier snapshots (the RSP compares with this too):")
windows = []
for k in (3, 4, 5, 6):
    for i in range(len(pre) - k + 1):
        w = pre[i:i + k]
        windows.append((slope(w), k, w[0][0], w[-1][0]))
for s, k, first, last in sorted(windows, reverse=True)[:4]:
    print(f"   {k} snapshots, {first} to {last}: {s:.2f}/yr, so the new slope is {b2 / s:.2f}x of it")

t_end = yrs(ANTHROPIC[-1][1]) - yrs(BREAK.isoformat())
print(f"6. From the break to Model 4's date ({t_end:.2f} yr) the fit gains {b2 * t_end:.1f} points, which is {b2 * t_end / b1:.2f} years at the old slope")

print("7. Amdahl's law with a 12x uplift on a fraction f of the pipeline and nothing on the rest (an illustration, not METR's model):")
print("   " + " | ".join(f"f={f}: {1 / ((1 - f) + f / 12):.2f}x" for f in (0.3, 0.5, 0.6, 0.8))
      + f" | f needed for 1.9x: {(1 - 1 / 1.9) * 12 / 11:.2f}")

Output on Python 3.13.14:

1. Refit with the break fixed at 2025-10-07
   Anthropic frontier, 16 snapshots: 15.37/yr -> 29.32/yr, ratio 1.91   (METR prints +15, +29, 1.9x)
   All developers, 21 records:      17.10/yr -> 23.60/yr, ratio 1.38   (METR prints +17, +24, 1.4x)
2. Break learned by a daily scan from 2025-01-01 to 2026-03-01: 2025-10-07, ratio 1.91   (METR reports 2025-10-07)
   ratio at other break dates: 2025-05-01 1.92 | 2025-07-01 1.95 | 2025-09-01 1.94 | 2025-11-01 1.89 | 2025-12-01 1.87 | 2026-01-15 1.84
3. Drop one snapshot at a time (16 refits), break fixed: ratio 1.81 (without Sonnet 4.5) to 1.94 (without 3 Opus)
   break re-learned each time:                        ratio 1.76 (without Sonnet 4.5) to 1.95 (without Opus 4)
4. Public Anthropic models only (14 record-setters, no internal snapshots): 14.86/yr -> 26.80/yr, ratio 1.80
5. Fastest stretch before the break, any run of consecutive frontier snapshots (the RSP compares with this too):
   3 snapshots, 3.5 Sonnet New to Opus 4: 19.49/yr, so the new slope is 1.50x of it
   3 snapshots, 3.5 Sonnet to 3.7 Sonnet: 17.39/yr, so the new slope is 1.69x of it
   4 snapshots, 3.5 Sonnet to Opus 4: 16.99/yr, so the new slope is 1.73x of it
   4 snapshots, 3 Opus to 3.7 Sonnet: 16.79/yr, so the new slope is 1.75x of it
6. From the break to Model 4's date (0.97 yr) the fit gains 28.3 points, which is 1.84 years at the old slope
7. Amdahl's law with a 12x uplift on a fraction f of the pipeline and nothing on the rest (an illustration, not METR's model):
   f=0.3: 1.38x | f=0.5: 1.85x | f=0.6: 2.22x | f=0.8: 3.75x | f needed for 1.9x: 0.52

Three published numbers check the readings. Claude 3.5 Sonnet reads 130.05 against its anchor of 130. Model 4 reads 178.70 against the 178.6 in METR’s text. Opus 5.5 reads 174.72 against 174.56 in Anthropic’s Haiku 5.5 system card, which uses a different fit, so a gap of 0.2 is expected.

The refit lands on METR’s printed slopes (15.37 to 29.32 a year, ratio 1.91) and on its all-developers slopes (17.10 to 23.60, ratio 1.38). The break is the better test. I did not set it by hand in step 2, and a daily scan picks October 7, 2025, the date METR reports. Dropping any one of the 16 snapshots moves the ratio only between 1.81 and 1.94, and moving the break anywhere from May 2025 to mid-January 2026 gives 1.84 to 1.95. So the headline is not hanging on one model or on the exact break date.

Where the ratio moves

Read the table by columns. The numerator, a post-break slope near 29 points a year, barely moves. The denominator and the fit family are choices. The link function says how a benchmark score relates to latent capability: the logistic default treats easy and hard tasks symmetrically, a log-log link assumes a long tail of very hard tasks and gives 2.9x, and a complementary log-log link assumes a long tail of easy or guessable tasks and gives 1.1x. METR says the fits are of similar quality (out-of-sample R² from 86.1% to 87.2%), that it has no strong prior reason to prefer one, and that it keeps the logistic because Epoch and Anthropic use it.

Two other rows are explained by METR itself. METR says the Haiku 5.5 system card reports 1.7x, and that the main difference from its own 1.9x is that it includes Anthropic’s recent internal-only models; my public-only refit gives 1.80. And the all-developers chart shows other labs holding the running record almost throughout the period before October 2025, so Anthropic’s own earlier slope (15 a year) is lower than the frontier’s (17). METR offers two readings of the 1.4x. Either Anthropic’s models were subfrontier in 2025 and the 1.9x is overstated, or Anthropic’s earlier points were more overfit to benchmarks and the 1.4x is understated.

What the interval leaves out

METR’s 95% interval comes from a bootstrap over benchmark scores. METR says it “only accounts for uncertainty in the AECI scores for each model, and does not include sampling uncertainty over which models are included in the fit.” My drop-one refits address the second part only at the level of the point estimate, and they say the estimate is stable. They cannot say how large the score noise is, and by eye the error bars on METR’s chart for the newest snapshots span roughly 20 index points, against a few for the early ones.

Three further limits are in METR’s own text. Model 4 “achieves greater than 50% on all the AECI benchmarks on which it was tested,” which METR says makes it “somewhat nervous about our ability to measure AECI now and in the future.” A typical benchmark attempt used about 2.5M input and 72k output tokens on Anthropic’s internal benchmarks, against roughly 0.5M and 10k for Epoch’s public ones, and METR guesses extra tokens could add “a handful of points” to every model’s score, probably without changing the ratio. And METR says the 1.9x “could be explained by overfitting to the benchmark distribution or otherwise be an artifact of the AECI methodology.”

“Mostly AI” is a guess, and the evidence for it is the redacted part

The Verge wrote that the researchers said the majority of the acceleration as of October “is due to AI,” with its contribution growing over time. METR’s sentence is: “We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI.” It calls the judgment one given “with substantial uncertainty.” METR also says its “more detailed estimates for the cumulative and instantaneous effects of AI” on Anthropic’s rate of progress were redacted. Its own table shows what the public can and cannot see.

One of the seven classes, the benchmark scores, is fully described, and that is the class METR says might be an artifact of its method. The evidence METR says “made us take this apparent ≈2x acceleration more seriously” sits in classes that were redacted or only partly described. METR’s own summary is that “the public has to rely on METR’s judgment much more than would be ideal.”

Interviews coded by Claude

METR held about 15 interviews with Anthropic staff, Anthropic ran its own, and METR reproduces quotes from 11 people in an appendix, sorted by whether each supports 2x or more acceleration from AI in their own work. The sorting was done by Claude, which METR asked to read its interview notes. The tally is 2 yes, 3 leans yes, 4 borderline or unclear, 0 leans no and 2 no, which METR reads as “mildly in favor of 2x or greater acceleration.” The quotes span the range. One researcher says, “I would definitely endorse >2 years of progress in 1 year and I don’t think 4 is crazy.” Another says, “The slope change is 50-80% Claude.” A third “Guesses 15-20% is due to AI. Don’t believe in self-reported productivity increase.” METR notes that each interviewee “often interpreted the question in somewhat different ways,” for example cumulative against instantaneous speedup, or personal against organization-wide.

Claude inside the evidence chain

METR lists two weaknesses in its adversarial-robustness section. Most of its data came through two Anthropic technical staff, and “Most analysis relied on Claude,” which METR says was deeply involved in “the majority of data collection and analysis that we received,” in “writing a significant share of the documents we received,” and in “the majority of our own analysis.” METR reports that it did “not detect any significant discrepancies consistent with a systematic bias even when checking against human interviews.” Anthropic’s own pace-of-development page has the same dependency: a Claude agent lists each person’s tasks, a Claude judge assigns the automation level, and the judge matched human raters exactly 59% of the time, against 35% between two humans. The same page says Claude “leads” 26% of Anthropic’s AI R&D work as of August and is not fully autonomous for any measured subset, which fits the Verge’s line that much of the acceleration appears to come from AI assisting human researchers.

I am not suggesting that any of those numbers is wrong. The point is narrower: the evidence about whether Claude speeds up Anthropic passed through Claude at several steps, which is why METR’s caveat that “greater access would allow us to increase the degree of adversarial robustness” matters. It is also the territory of my earlier post on Anthropic’s S-1, which found the filing saying that its evaluations could overstate safety.

Twelve times is a different unit from 1.9x

The Verge reports that employees “on average self-reported being around 12 times as productive,” “around double what staff had reported just a few months earlier while using Mythos 5.” Both numbers hold. Reading METR’s chart on its log axis, the broad Slack poll gives about 2.8x for Opus 4.6, 3.9x for Mythos Preview, 5.2x for Mythos 5 and about 12x for Model 4, so the last step is a factor of about 2.4. METR’s text says staff “use reactions to vote on their estimated speedup from the most recent model at the time,” and calls the result “only very loose evidence: we expect the responses come from an unrepresentative subset of staff and the responses reflect fairly low-effort consideration.” It adds that “there is some evidence that self-report survey estimates of uplift can overestimate true productivity impact.” The last Slack point sits above both curves METR fitted to the series.

Uplift and acceleration are different quantities, and METR says so: “individual uplift/speedup estimates do not correspond to acceleration rates,” because “sustaining a particular growth rate in outputs like model capabilities generally requires a similarly sustained growth in inputs like labor and/or labor productivity to overcome diminishing returns.” Context from METR’s earlier work: its May survey of 349 technical workers found a median self-reported change in the value of work of 1.4x to 2x and a median speed change of 3x, which METR expects to be higher than the value change. Anthropic’s poll for Mythos Preview, about 3.9x on my reading, sits near that 3x, though the statistic (an average against a median), the population and the question all differ.

A toy calculation shows how a 12x uplift and a 1.9x acceleration can both be true. If AI speeds up a fraction f of the research pipeline by 12x and leaves the rest alone (compute-bound experiments, for instance), Amdahl’s law gives 1.38x overall at f = 0.3, 1.85x at 0.5, 2.22x at 0.6 and 3.75x at 0.8, and about 0.52 would produce 1.9x (step 7 in the output above). That is my illustration, not METR’s model, and it treats a speedup of work as if it were a change in a growth rate, which METR says it is not. One interviewee’s remark in the same appendix describes a similar split: “Now only bottlenecked on ideas & compute, no longer bottlenecked on code.”

The second METR report: “less than a factor of 2” was 1.97

Anthropic’s Responsible Scaling Policy (version 3.4, July 8, 2026) has two triggers for its automated R&D threat model. One is a model that can substitute for its research staff. The other is “dramatic acceleration” of AI progress for reasons that likely relate to automating AI R&D. METR’s review compares what Anthropic’s public August Risk Report said with the private version METR was shown.

METR’s summary: “We agree with Anthropic that catastrophic risks from July 15th models were low, but we think it is plausible that they had crossed their early warning threshold at the time.” It adds that both its own researchers and Anthropic’s relevant staff “lean toward it not having been crossed,” and that “the public report overall understates the level of acceleration that was occurring.” It also notes that the public report says Model 2’s score has large error bars but omits the private report’s qualification that later snapshots of that base model were expected to score higher, and that further snapshots did improve by about as much as Mythos 5 had over Mythos Preview.

The baseline the policy names is stricter than the one the headline uses

The RSP’s test for that second trigger has two parts. Part (a) reads: “(a) we observe or expect double the rate of progress in AI aggregate capabilities compared to both the rate we’d expect and the fastest rate of extended progress we’ve observed in the absence of significant AI contributions to AI R&D,” with a footnote that “extended” means “over at least three model generations.” Part (b) asks whether it is plausible that the doubling is substantially attributable to the automation of research or engineering. METR’s 1.9x divides by the average slope before the break. In my readings the fastest pre-break stretch of three consecutive frontier snapshots (3.5 Sonnet New to Opus 4) runs 19.49 points a year, and the post-break slope is 1.50 times that; over four snapshots the ratio is 1.7. METR says its holistic report does “not pay special attention to the thresholds set in the RSP,” and I am not claiming the line was or was not crossed. The point is that “double” is relative to a baseline, and the baseline the policy names is a higher bar than the one the headline number uses.

Redactions, contract terms and METR’s own checklist

METR’s summary of the contract says Anthropic “may require redaction of confidential or non-public information,” and that the parties discuss redactions in good faith, “with Anthropic’s redaction requests controlling.” METR can always disclose five things: the existence and general terms of the review and redaction process, whether Anthropic exercised redaction rights, whether any findings depended significantly on redacted information, the level of access, and the confidentiality framework. Whether the redactions were justified is not on that list. My earlier post found that on October 16 Anthropic’s page dropped a sentence saying METR was “free to characterize our redactions publicly, and say whether it believes the redactions were justified.” METR’s first footnote now reads: “We do not make a comment on whether these redactions were justified overall, since we do not think METR is well positioned to weigh a company’s business interest in preserving this confidential information against the public’s interest in better understanding the risks from AI development.”

METR also scores itself against the AI Evaluator Forum’s AEF-1 checklist. By my count of its 26 top-level items, 15 are requirements and 4 of them are marked No. Item 1.1 (technical access) is one: METR corrected it from Yes to No on November 6, and says its staff never requested query access to unreleased models. Item 4.1, sharing enough methodological detail for independent review, is marked No with the note “This was not possible due to redaction requirements.” Item 4.6, that the provider “did not have authority to redact results to conceal concerning findings,” is marked No with the note “Anthropic had unilateral redaction rights.” Item 5.4, a responsible disclosure policy, is marked No. Two of 11 recommendations are also marked No, one is “Not requested” and one is N/A.

The checklist and the disclosures also record the conflicts. One team member had been in discussions with Anthropic about a role (the offer was declined) and held equity in a competitor as an ex-employee, so their work was checked by an unconflicted colleague; one incoming METR staff member holds Anthropic equity. METR notes it has “other independent evaluation arrangements with Anthropic” and that its work with nonpublic models and free tokens “incentivize a cordial relationship with AI companies.” It says Anthropic’s feedback on drafts of the Risk Report Review was in some cases “overly forceful,” that Anthropic was receptive to its pushback, and that it does not think either report is “significantly distorted by this.”

Anthropic’s response came from co-founder Sam McCandlish, who wrote on X: “They had complete independence to form their own conclusions.” He also wrote: “We redacted a lot, and I accept that this makes it harder for others to check METR’s work. But we determined that sharing exactly what is driving the speedup isn’t the right thing to do, both because it’s commercially sensitive and because it risks handing out a roadmap for others to go faster.” METR’s own post says that “the public must rely on our judgment more than we consider ideal.” Both statements can be true. Independence over conclusions is what the contract gives METR; checkability is what the redactions take away.

The compute check that was cancelled

METR’s contract originally covered “a verification of Anthropic’s compute allocation, especially the allocation between capabilities and safety.” Anthropic changed the scope midway, and METR reports that Anthropic “stated this was due to wanting to keep the engagement focused on risk assessment for the time being, and that it would revisit compute allocation verification later on.” That matters because Anthropic’s pace-of-development page publishes the only compute figure I know of: about 6% of the compute that went to AI R&D in one week in July was allocated to safety, and about 12% of the compute that went to AI-driven AI R&D. Those figures come from a Claude-based classifier: for research training and evaluation runs it covered about 14% of the week’s almost 10,000 runs, weighted toward the biggest, and a variant read agent transcripts for inference. The page itself says: “A measurement meant to be trusted by outsiders will need to be complete, accurate, and technically enforced.” It also plans a pipeline so that “an independent third-party could re-run the classifier on a random subsample of jobs and transcripts and check both the sorting and the totals.” METR had reached no conclusions on compute before the component was cancelled, so nothing suggests the figure is wrong. It simply has no outside check yet.

What would settle it

• Anthropic’s own report. McCandlish says it will come “in the coming weeks,” including “where we stand against our Responsible Scaling Policy.” I found no such report on Anthropic’s news or research pages on November 7. Worth checking when it lands: which link function, which baseline (the average slope or the fastest stretch) and which benchmarks it uses.
• A table of scores. METR’s AECI section was “very minimally redacted.” A table of snapshot dates, scores and intervals would let anyone repeat the checks above without reading pixels off a chart.
• The leading indicators. Both reports mention figures that were redacted. METR gives only a range of 1.3 to 2.8.
• Compute verification. Anthropic says it will revisit it; the original contract scope shows it was meant to be checked from outside.
• A human-coded interview table. The current sorting was done by Claude on notes METR describes as varying in depth.

Limits of this check

The points are read off raster charts, with no error bars, so I tested METR’s fit rather than its data, and I could not reproduce its bootstrap interval or its link-function comparison. The fastest-stretch check reads “model generations” as consecutive frontier snapshots, which is my interpretation. The Amdahl calculation is an illustration. I did not see METR’s roughly 100-page full report, only the public version, and I checked the Verge story against METR’s text, not against Robert Hart’s reporting. The Haiku 5.5 card’s 1.7x sits in a chart image I did not decode, so I attribute it to METR.

Sources

• METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 and its Risk Report review, November 5, 2026; METR’s post on X; METR’s May 2026 survey of technical workers.
• The Verge, Why a new report on Anthropic automating AI development is alarming researchers.
• Sam McCandlish, post on X, November 5, 2026.
• Anthropic, Measurements for understanding the pace of AI development inside frontier labs, the Haiku 5.5 system card and the Responsible Scaling Policy v3.4.
• Epoch AI’s public ECI code, which METR says it uses to estimate AECI, and the AI Evaluator Forum’s AEF-1 checklist.
• Earlier sxz.io posts: Anthropic’s METR arrangement and its changed publication terms, Anthropic’s S-1 and the “closing the loop” cycle, and METR’s pilot red team of OpenAI’s agent monitor.