Cognition ships SWE-2 Flash: a coding agent 3x faster, 1/5 the price
Free in Devin through Nov 30, though it trades accuracy for speed. Plus: MiniMax M2 and Mercury 2.x near retirement; a DeepSeek route doubles.

Copy markdown
SWE-2 Flash: speed over the leaderboard
Cognition's new 276B-A12B agent, post-trained from Thinking Machines' Inkling-Small, runs ~3x faster than SWE-2 (about 200 tokens/sec) at one-fifth the price, and it's free in Devin Desktop and CLI through Nov 30. The catch: it edges SWE-2 on FrontierCode 2.0 (24.0 vs 23.8) but slips on Terminal-Bench 4.0 (20.0 vs 27.3) and DeepSWE 1.1 (60.1 vs 73.0) — it buys speed with accuracy. Weights stay private; a partner-sharing program is weeks out.
MiniMax M2 and Mercury 2.x are on the clock
If either is in your stack, migrate now. MiniMax has already retired M2 and M2.1; M2.5 and M2.7 stop serving pay-as-you-go on Nov 18 and reroute to M3.1 Flash. Inception closed Mercury 2 and 2.5 to new users, and existing access ends Nov 15 — the vendor points you to Mercury 3.
Check your OpenRouter bill
In the week to Nov 5, DeepSeek V4 Pro's OpenRouter route roughly doubled, to $1.60/M in and $3.20/M out. It's not all one way — Kimi K3 output dropped from $20 to $16/M and GLM-5.2 output eased to $6.80/M. Route prices aren't vendor list prices, but they're what you pay on OpenRouter, so it's worth a spend check.