Frontier-model vendors are now cutting prices materially in the tier where teams make recurring choices about coding agents, long-context assistants, and tool-using workflows.
Observed facts: Simon Willison lists GPT-6 Luna at $0.10/M input tokens and $0.50/M output tokens, compared with $0.20/M and $1.20/M for GPT-5.6 Luna. He also lists GPT-6 Sol at $2/M input and $10/M output, versus $4/M and $20/M for GPT-5.6 Sol, while Claude Opus 5.5 is priced at $4/M input and $20/M output after the prior Opus price of $5/M and $25/M. The source further reports that Opus 5.5 cache reads fell 60%.
The mechanism is a lower marginal cost for the repeated model calls that make agent systems viable: planning, tool use, review, retries, and continued context. Why now: the price reductions arrived with a concentrated set of releases, making cost rather than only benchmark rank an immediate selection variable. The implication is that teams can reconsider agent loops previously constrained by per-task spend, but they should evaluate completed-task cost rather than token price alone. The source reports two Opus 5.5 max-reasoning runs that hit a 128,000-token output limit without an answer, which shows why cheap tokens do not by themselves establish reliable execution.
Evidence boundary: this is one practitioner’s release-day comparison and anecdotal testing, not an independent benchmark or a production reliability study. Watch condition: independent evaluations should show whether the lower-priced models preserve completion reliability, latency, and tool-use quality under long-running workloads. This begins a distinct record because the evidence captures a synchronized release-and-pricing shift across competing model tiers, rather than a single-model capability claim.