The task horizon: how fast the models are moving, why, and what a token costs now

The models a business can buy in February 2026 are not the same kind of thing as the ones it could buy two years ago, and the difference is not simply that they are bigger. This briefing takes the single most useful measure of the change, METR's task horizon, as its starting point, and explains the four things behind it: models that reason before answering, a training method that teaches that reasoning without human examples, architectures and chips that make a large model cheap to run, and a body of engineering for making the result reliable. It ends with the numbers a manager should plan around.

How long a task can a model hold?

≈5.3 hours

Length of software task Claude Opus 4.5 completed unaided at 50% reliability, measured by the time a skilled human takes [1]

131 days

Doubling time of the task horizon since 2023, against 197 days across the full history since 2019 [1]

≈1 hour

The same measure for the best model in March 2025, when METR first published it [2]

METR, a research organisation that evaluates frontier models, measures the length of task a model can complete on its own half the time, with length defined by how long a skilled human takes. When it first published the measure in March 2025 the best model managed tasks of about an hour and the horizon had been doubling roughly every seven months since 2019.^[2] Its update on 29 January 2026 puts Claude Opus 4.5 at about 320 minutes, with the doubling time since 2023 at 130.8 days, a fifth faster than the long-run average; the task suite grew to 228 tasks and METR warns that its measurements above about sixteen hours are unreliable because it is running out of long enough tasks.^[1] OpenAI's GDPval, from October, said the same thing about quality rather than length: on 1,320 real deliverables from 44 occupations, frontier models were approaching expert grade at about a hundredth of the time and cost.^[3] The unit of work a model can be handed has moved from a paragraph to an afternoon.

Thinking before answering

Until September 2024 every mainstream model answered in a single pass. OpenAI's o1 introduced a model that first writes out a chain of reasoning, checks it, and only then answers, trained by reinforcement learning to make that reasoning useful.^[4] DeepSeek's R1, released in January 2025 and published in Nature in September after peer review, showed the recipe could be reproduced cheaply: reinforcement learning against problems whose answers can be checked automatically, with no human-written examples of reasoning, produced a model that re-checks its own work because that behaviour was rewarded.^[5] The method, now called reinforcement learning with verifiable rewards, is behind every frontier release since, from GPT-5 in August through Gemini 3 in November to Opus 4.5 and GPT-5.2 in December.^[6][7][8] Epoch AI's capabilities index has reasoning models improving at roughly fourteen points a year since they appeared, against six for non-reasoning models.^[9] The trade-off is that a reasoning model may spend thousands of tokens thinking before it says anything, which is why the labs now expose a budget or effort setting so that a simple question does not pay for a hard one's deliberation.

Cheaper by the month

Capability is not the only thing on a curve. Andreessen Horowitz's analysis puts the price of a fixed level of model quality falling about tenfold a year, with GPT-3-level ability now roughly a thousand times cheaper than its 2021 price.^[10] Epoch AI's more careful version finds the rate depends heavily on the task, from nine times a year to several hundred, and cautions that the fastest declines may not persist.^[11] The flagship prices show it directly: Anthropic's Opus 4.5 launched in November at $5 per million input tokens, a third of the previous flagship, and GPT-5 in August at $1.25.^[7][6] Behind the curve are three architectural changes. Mixture-of-experts layouts mean only a fraction of a model runs for any token: DeepSeek V3 activates 37 billion of 671 billion parameters.^[12] Compressed attention cuts the memory needed to hold a long conversation by around ten times.^[12] And the arithmetic itself has shrunk from 16-bit to 4-bit, with Nvidia's Rubin generation, unveiled at CES in January and due to ship later in 2026, built around it; Google's Ironwood TPU and Amazon's Trainium 3 tell the same story on different silicon.^[13][14][15]

Context: long, and not as long as advertised

A model's context is how much it can hold in view at once, and it has grown from a few thousand tokens to a million on the current Gemini and Claude tiers.^[16] Independent testing finds the advertised number optimistic: a study by Chroma in July 2025 showed performance degrading steadily as input grew, well before the stated limit, a phenomenon now called context rot.^[17] The practical response is a discipline Anthropic's engineering team wrote up in September as context engineering: treat the context as scarce, give the model the smallest set of high-signal tokens, keep state in external notes it can re-read, compact long sessions, and hand sub-tasks to fresh sub-agents that report back a summary.^[18] Most of the difference between an agent that works and one that wanders is in that discipline rather than in the model, and the agents themselves are a briefing of their own.

Reliability

Two problems deserve plain statement. Hallucination is not a mystery: OpenAI's September 2025 paper shows it is the predictable result of training and scoring that reward a confident guess over an admission of uncertainty, and that the fix is in the incentives; GPT-5 hallucinates markedly less than its predecessor and still does.^[19] The April 2025 GPT-4o episode, in which a thumbs-up reward signal made the model flatter its users and was rolled back in four days, is the same lesson from the other side.^[20] Prompt injection, where content a model reads carries instructions it follows, is unsolved at the architectural level because a model sees its instructions, the user's request and an untrusted document as one stream of text; the June 2025 zero-click attack on Microsoft 365 Copilot showed it against a shipped product.^[21] Any agent with tool access and exposure to email, web pages or documents is a live attack surface and should be built as one.

The cost controls, by contrast, are mature. Prompt caching, which Anthropic introduced in August 2024, cuts the cost of repeated context by up to 90% and latency by up to 85%; batch processing halves the price of anything that can wait a day; structured outputs reduce both token volume and retries.^[22] A surprising number of production systems use none of them.