The four-month countdown to obsolescence

The interesting question about a model is not how fast it types. It is how big a job you can hand over before it comes back wrong. METR measures precisely that, and publishes the numbers: not tokens per second, not a benchmark percentage, but the size of task a model finishes, priced in the time a skilled human would need for the same work.

One thing to settle first. Every date in here past April 2026 is an estimate, and this is a study of principles rather than a forecast of when. The doubling rate is a slider for exactly that reason: set it to three months or to nine and every milestone slides along with it, while the order they arrive in does not move. The claim is about shape, not schedule.

What the number means

Every task in the suite carries a price tag in human time. Skilled professionals, averaging about five years of relevant experience, were contracted to attempt them, and the geometric mean of their completion times is the task's size. Then the model attempts the same task. Its time horizon is the job size at which it succeeds a given fraction of the time.

• It is job size, not runtime. A model that clears a four-hour job in nine minutes still scores four hours. The unit is the human's clock, not the machine's.
• 50% and 80% are different questions. The first asks what a model can do. The second asks what it can be trusted to do.
• The baseline human was handed a well-specified task and little context. Most real work is neither.

What the data says

The most recent measurement METR publishes (Claude Mythos Preview, early, 7 April 2026) puts the 50% horizon at 17.4 hours of skilled human work, finished half the time. The same model's 80% horizon is 3.1 hours. Hours are the only unit used here on purpose: a "day" of work is eight hours to one reader and twelve to another, and the measurement should not move because of it.

Fit a line through the releases since 2023 and the 50% horizon doubles every 128 days, near enough four months, with an interval of 104 to 158 days. Fit the whole history back to GPT-2 and the doubling is 188 days. Which of those two you quote decides whether the next order of magnitude lands in under two years or nearer three, and both fits come from the same file.

Reliability is the lagging column

Read the table down the 80% column instead. The best 80% figure available, 3.1 hours, sits roughly where the 50% figure stood in July 2025. Around eight months separates what the frontier can do from what it can be trusted to do, and that gap has been remarkably stable while both columns climbed: the last column sits near five for almost every model measured, with Opus 4.6 the outlier at ten.

The 50% number is what you demo. The 80% number is what you staff.

This is why the slider marked "success rate I require" moves the chart so violently. Asking for four-in-five instead of one-in-two does not shave a little off the horizon, it divides it by five or more, and by ten for some models. Most of the disagreement about how capable these systems are is really a disagreement about which column someone is quoting.

What the curve does not license

The extrapolation is the part everyone wants and the part the data supports least. A few things are worth holding onto before drawing the dashed line out to 2029:

• METR states that measurements above 16 hours are unreliable on its current task suite. The newest 50% point, 17.4 hours, is already past that line.
• That point's 95% interval runs from 8.5 to 55 hours. The headline is a factor of six wide before anyone extrapolates it.
• The suite is well-specified tasks with clean success criteria. Ambiguity, stakeholders and half-written requirements are not in it, and they are most of the job.
• A doubling fitted over twenty-six releases is a description of the past, not a mechanism. It contains no reason it must continue.

What an hour of your time buys

Now assume the person never writes the code. Their hours go into briefing a job, reading what comes back, and deciding. "I could have done it faster myself" stops being the relevant comparison, because they were never going to do it. The question is leverage: how many hours of delivered work one hour of their own time is worth.

METR says the same thing in the negative. Their note on the metric's limits is blunt about it: a 50% time horizon of X hours does not mean you can delegate X-hour tasks, and turning horizons into productivity needs a measurement nobody has, which is how much time a human spends prompting, waiting and checking output.

When that was measured, the answer was ugly. In METR's randomised trial, sixteen experienced developers working in repositories they knew well were 19% slower with AI assistance. They had predicted a 24% speedup, and afterwards still believed they had been sped up by 20%. Self-reported estimates of this overhead are not merely noisy, they are wrong in a known direction.

So the chart prices the human in. Briefing a job costs an hour and a half whatever its size. Checking costs whatever share of the job you choose to inspect, and inspecting more catches more with diminishing returns. You read attempts until one passes, not all of them. Tries are then chosen for you, at whatever number makes your hour go furthest.

What that model says

The first mode is going through the work itself. Not reading it line by line, which nobody does with AI output and the model does not assume: running it, skimming the diff, trying a few cases, reading the parts that worry you. Call it attention spent on the artifact.

It fails twice over as jobs grow. The cost rises with the job, and what you catch shrinks as the job outgrows the eight hours or so of it you can hold at once. Looking at 15% of a four hour job is a check. Looking at 15% of a four hundred hour job is a ritual.

The second mode replaces it. Say once what "right" means (acceptance criteria, tests, an oracle) and read the report instead of the work. The cost barely moves with size, which is why it takes over. So the human does not fall out of the system when the work outgrows them. Their job changes.

Criteria decay too, though, and this is where the piece stops being comfortable. They catch the failures somebody thought to name, and every doubling of job size brings more ways to be wrong. How far your coverage stretches per doubling is exactly what expertise buys: someone who knows where things break names the right dozen checks, someone who does not names the obvious three.

The work outgrows what you can read long before it outgrows what you can specify.

When the obvious mistakes ran out

The layperson's row is the one that has been changing under me. Their contribution is 1.04x, which is to say almost exactly nothing, and it is falling. But it was not always nothing. Against GPT-4 the same person was worth 5x, because most of what GPT-4 got wrong was wrong in ways anybody could see: it did not run, it ignored half the request, it invented an interface that was not there. Catching that needs attention, not expertise.

What changed is not the person, it is what a failure looks like. Every doubling of capability clears out another layer of obvious mistakes and leaves the ones that need someone who knows the domain: three quarters of GPT-4's failures were catchable by anyone, against 8% of today's frontier's.

On this projection they cross into actively harmful in early 2027, and the mechanism is worth naming, because it is not that they stop catching things. A reviewer has two ways to be wrong. Once you cannot tell good work from bad you start throwing out the good, and a discarded good attempt costs exactly what a real failure costs while buying nothing.

A check that cannot tell good from bad is not a weak control. It is a negative one.

Where the value went

Everyone else has already switched, and today the switch is worth more than the skill. Going from looking to specifying takes a competent reviewer from three hours of job to twelve, a jump no amount of looking harder would buy. Expert over competent, both specifying, is a further factor of four, because false alarms punish the weaker specifier twice: once for the good work they discard, and again for the extra attempt they read in its place.

What the human is actually worth

The contribution column is the answer to the question the doubling cannot answer on its own. An expert writing criteria makes a job 16 times the size handleable at the same reliability, and turns each hour of their own time into 7.6 hours of delivered work. That is what they are worth to the system, and none of it is implementation.

It also has a ceiling, and the ceiling is honest rather than optimistic. Nothing here leans on a model succeeding less than 30% of the time per attempt, because METR measured the curve at 50% and 80% and below that there is no measurement to stand on. That alone caps the contribution near fifteen times, whatever the criteria are. Anyone quoting a bigger multiple is extrapolating a fitted curve into a region nobody has been.

Run the chart out to 2030 and the ranking stops being a ranking of people and becomes a ranking of how fast their coverage rots.

The competent specifier falls from 3.8x toward 1.2x, the layperson from 1.04x to below 1, and only the expert holds anything worth having. Each line runs until it meets the person-year mark and the middle two stop on it, in May and September 2029, because past that point the only two answers worth drawing are handing it over unwatched or having your best person specify it.

What if they stop needing watching

Everything above holds one thing fixed that has no right to be fixed. The projection doubles the 50% horizon and keeps the gap to the 80% number exactly where it is today, which quietly asserts that models get bigger without ever getting steadier. Nothing supports that. It is just the only thing the measured data can be extrapolated to say.

The chart has a control for the other case. Close a share of that gap with every doubling and the human's contribution falls off a cliff, without the models needing to be any more capable at all. Closing 5% per doubling takes an expert's contribution in 2030 from nine times down to 2.2. Closing 15% takes it to 1.19, which is nothing.

Checking is worth something only while the work is unreliable. Deciding what to build is worth something either way.

And that is the honest answer to whether the human becomes unnecessary: their supervision does, on any assumption where reliability improves along with capability. Their brief does not. Someone still has to say what should exist, and in that regime the model returns something like twenty thousand hours of work for each hour spent saying it. The leverage does not fall when checking becomes pointless. It goes vertical, and moves entirely into the sentence at the start.

Which reframes what a good engineer is for. On the frozen-reliability path they are worth fifteen times the unwatched system because they catch what it gets wrong. On the improving-reliability path they are worth almost nothing for catching, and everything for choosing: what to build, what must never happen, which of a thousand possible correct systems is the one you wanted. That second skill has no benchmark and is not doubling every four months.

Where the bandwidth runs out

All of that measures one job. Switch the chart to a person's week and a different question appears, because the human is not only a checker, they are a fixed quantity of hours. The same forty hours can brief roughly twenty-seven small jobs or brief and supervise about eight big ones, and which delivers more depends entirely on how big the multiplier still is.

Today it is not close: an expert's week delivers 4.6 times what the same week spent only briefing does. But the multiplier decays while the cost of their time does not, and the two converge. On the default assumptions they cross in late 2033, marked on the chart, after which one person delivers more work by briefing many jobs and checking none of them than by supervising a few carefully.

Which is why the two views disagree, and both are right. Per job, supervision cannot lose: a bigger job is a bigger job, and the expert line stays on top forever.

Per week it loses easily, because the expert's job is bigger but eats more of a fixed forty hours. Their advantage is the multiplier divided by that cost, so the crossing says nothing about judgement. It is the moment a multiplier of three and a half stops covering a bill of the same size. The bottleneck moves from insight to bandwidth, and bandwidth is the one thing about a person that has never doubled.

The human stops being the thing that makes the work good and becomes the thing that cannot issue instructions fast enough.

Seeing it at all

The third view puts that on one axis. Plot every line as a multiple of handing the work over unwatched and the unwatched line goes flat at 1, which is the only way to see the crossing: at the point it happens the two are 12% apart on a chart spanning ten orders of magnitude, a gap thinner than the line itself.

Against a flat baseline the expert's decline from 13x in 2019 to below 1 in late 2033 is the whole story in one shape. The layperson passes under it in 2025 and the competent specifier in late 2026.

The loop, not the gate

There is a reading of that crossing which the chart, as described so far, cannot make. A supervisor is not only a filter on output, they are a feedback loop on a process, and a feedback loop is never judged on throughput. It costs throughput by definition, and it pays in something else: the thing does not wander off.

Note what the crossing does not say. At the moment the expert "loses" in 2033, every job they touch is still 3.2 times what it would have been unwatched, and in 2038 it is still 2.2 times. Their judgement does not decay much at all.

What decays is their reach: one person can gate about eight jobs a week against the twenty-seven a briefing-only week starts, so they cover 31% of the stream, and that share is fixed while the stream is not. It is a coverage result, not a competence one.

A feedback loop starved of bandwidth is not a feedback loop that stopped working.

So the chart carries a drift control, defaulting to zero. Turn it on and it asserts that an unwatched stream slides: specs rot, regressions ship quietly, the work drifts off the goal, and none of it shows in any individual job's success. Gating half sees that at best, since a stream can pass every check it is given and still be going the wrong way.

With any drift at all the answer past 2033 is not to leave, it is to change job a third time: stop gating, start monitoring. At 5% per doubling an expert who switches recovers from 0.98x to 2.1x and climbs, while the unwatched stream is down to a third of its nominal output by 2034 with nobody whose job it is to notice. At zero, none of that is claimed.

The red dot marks a milestone rather than a prophecy: the first point where a single job handed over is one person-year of human work. Under these assumptions an expert-supervised system reaches it in early 2028, and the same models unwatched reach it seventeen months later. That gap is the clearest statement of the human's worth this model can make. Supervision does not slow the curve down or speed it up. It moves you about a year and a half along it.

So the shape is not a human handing over more and more until they are redundant. It is a human doing a different job every couple of years: going through the work, then writing the tests, then specifying what the system must never do. The doubling decides how big a unit arrives. What you are worth to it is how well you can say what right looks like, and that skill is not doubling every four months.