Drift in multi-day agentic tasks
Instruction adherence degrades measurably past a horizon our tooling had not previously instrumented.
Cortex 5.1 and Pulse 5.1
Two models for work that outlives a single turn: they hold state across tools, recover from their own mistakes, and say when they are unsure instead of inventing a finish line.
What is new
The headline change is horizon length. Both models were trained and tuned for tasks that run across days rather than minutes, which changes what matters: not the quality of a single answer, but whether the model still remembers the constraint you gave it on Monday by the time it is writing the summary on Thursday.
In practice that meant spending most of our effort on three unglamorous problems — keeping the original instruction in view, noticing when a tool call has quietly failed, and distinguishing "finished" from "out of ideas". None of those show up in a chat demo.
- Cortex 5.1 is the model for ambiguous briefs and long agentic runs. Slower, more expensive, and the one we reach for when a wrong confident answer costs more than a careful one.
- Pulse 5.1 is the default. It handles most production traffic and is the model we tune first when a new capability ships.
- Both share a one-million-token context window, the same tool protocol, and the same refusal policy, so swapping between them is a one-line change.
A correction we owe you
An earlier draft of the system card reported a refusal benchmark improvement of nine points over the previous generation. That number was wrong. The evaluation split had been contaminated by training data, and the model was not being tested on the thing the card claimed it was testing.
The suite has been rebuilt with a held-out split and a continuous contamination check, and the affected claim has been withdrawn from the card rather than quietly restated. The rebuilt benchmark shows a smaller improvement than the original figure, which is the number we now report.
Why we are telling you this. A benchmark that flatters us is worse than no
benchmark, because it removes the reason to look. The withdrawal is logged in the public
incident log alongside the date it was caught and what changed as a result.
Evaluations
The release gate now covers tool misuse and instruction drift over multi-day runs, plus the standing suites for cyber uplift, bio risk, and refusal behaviour under pressure. Results are in the system card, including the evaluations where the models did not improve on their predecessors.
| Attribute | Cortex 5.1 | Pulse 5.1 |
|---|---|---|
| Context window | 1M tokens | 1M tokens |
| Intended work | Long-horizon agents, ambiguous briefs, analysis | Production traffic, coding, support, extraction |
| Latency profile | Deliberate | Interactive |
| Relative cost | $$$$ | $$ |
| Availability | Pro, Team, Enterprise, API | Free, Pro, Team, Enterprise, API |
Availability
Both models are available today on the API and in the Resonance apps. Existing deployments pinned to the previous generation keep working; the console shows an upgrade estimate before you switch, and spend caps carry over unchanged.
If the model fails your task, that is useful information. We would rather hear it than have you quietly build around us.
Related
All announcementsResponsible scaling, version 3
What happens when a capability threshold is crossed partway through a training run.
Upgrading without a rewrite
How to move a pinned deployment across generations and prove nothing regressed.