---
title: "Cognition's SWE-2: A Fresh $48B Valuation, Still Running on Someone Else's Model"
date: 2026-09-14
tags: ["cognition","devin","agentic-coding","benchmarks","kimi-k3"]
categories: ["AI Tools","Industry"]
summary: "Cognition shipped SWE-2 on September 10, Devin's new engine, post-trained on Moonshot's 2.8T-parameter Kimi K3 base with three reasoning-effort tiers. Cognition's own numbers claim near-Fable-5.1 coding quality at a fraction of the cost — but the model underneath is still someone else's, four months after Cognition's $48B Series E."
---


![Cognition's SWE-2: A Fresh $48B Valuation, Still Running on Someone Else's Model](/images/cognition-swe-2-devin-kimi-k3.png)

Four days after closing a [$48 billion Series E](/2026/09/cognition-series-e-48-billion-close/) on nearly double its prior valuation, Cognition shipped the model that's supposed to justify it. **SWE-2**, announced September 10, 2026, is Devin's new engine — and it settles, for now, the question this blog has been asking since the Series E coverage: is Cognition building a frontier model of its own, or getting very good at renting one?

The answer is still the second one. SWE-2 is post-trained on **Kimi K3**, Moonshot AI's 2.8-trillion-parameter base model, which had already been through extensive reinforcement learning for agentic coding before Cognition touched it. Cognition's own [announcement post](https://cognition.com/blog/swe-2) doesn't hide this — it frames SWE-2 as what happens when Devin's product and evaluation infrastructure gets pointed at the best available open base, rather than claiming an in-house frontier model that doesn't exist.

## What's actually new

The interesting engineering claim isn't the base model — it's how Cognition post-trained it. SWE-2 is the first Devin model trained with **three reasoning-effort tiers (medium, high, max) in a single RL run**, using what Cognition describes as a cost-penalized objective — reward calculated as score minus a cost penalty term, tuned separately against each tier's own cost-performance curve rather than trained once and then just throttled at inference time. Medium optimizes for speed and cost; high and max trade latency for thoroughness on harder tasks. That's a genuinely different training-time approach than most labs' effort-level implementations, which tend to be inference-time knobs on a single trained model (Claude's adaptive thinking, Grok's low/medium/high split).

## The numbers, with the usual caveat

All of the following are Cognition's own self-reported figures — treat them with the same skepticism this blog has applied to every vendor benchmark table since Muse Spark 1.1's reproducibility gap:

| Benchmark | SWE-2 | Kimi K3 (base) | Grok 4.6 | Fable 5.1 |
|---|---|---|---|---|
| FrontierCode 1.1 Main | 50.0% | 44.2% | 48.0% | 50.9% |
| DeepSWE 1.1 | 73.0% | 68.5% | 67.5% | 67.4% |
| Terminal-Bench 2.1 | 92.8% | 88.3% | 88.4% | 91.4% |

Cognition's headline claim, quoted directly: SWE-2 "beats SWE-1.7 and Grok 4.6 on both score and cost, matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of their price, and comes within a few points of GPT-6 Astra at a quarter of the cost." The FrontierCode number backs that up on its face — 50.0% against Fable 5.1's 50.9% is a real one-point gap, not a rounding trick, and post-training a 2.8T base to land within a point of a $10/$50-per-Mtok frontier model is a legitimately good result if it holds up independently. None of these three benchmarks has an independent leaderboard entry for SWE-2 yet — worth rechecking once one appears, the same standard this blog applied to Fable 5.1 and GPT-6 Astra's own launch-day numbers before Terminal-Bench 4.0's independent leaderboard settled the question between them.

Pricing isn't disclosed in the launch post at all — SWE-2 isn't sold as a standalone API model with a published rate card. It's available today through Devin Desktop, CLI, Web, and the Fusion interface; the "fraction of the cost" and "quarter of the cost" comparisons are Cognition's internal framing of Devin's own usage-based pricing against competitors' list prices, not an apples-to-apples per-token comparison you can verify yourself.

## The moat question, still open

This is the same tension this blog flagged when Cognition's Series E closed: SWE-1.7 ran on Kimi K2.7 Code, and now SWE-2 runs on Kimi K3. Two model generations in a row, Cognition's competitive edge has come from what it builds *around* a Moonshot base — RL post-training, effort-tier engineering, Devin's async agent product, Auto-Triage, Security Swarm, event-triggered Automations — rather than from a proprietary foundation model the way Anthropic, OpenAI, and Google compete.

That's not automatically a weakness. Plenty of successful software companies build durable products on top of models or infrastructure they don't own — it's a legitimate strategy, and Devin's actual product surface (persistent async agents, an append-only replay-safe event log, Slack/GitHub/Linear-triggered Automations) is real differentiated engineering, not just a wrapper. But it does mean Cognition's $48B valuation is a bet on product and orchestration mattering more than model ownership — the opposite bet Anthropic is making with Claude Code, where the harness and the frontier model are built by the same company and co-evolve on the same roadmap. Devin's architecture is structurally closer to Claude Code's delegate-and-review, terminal/async-native model than to Cursor's IDE-anchored one — that part of this blog's read on Cognition hasn't changed. Whether a rented-base-model strategy can keep pace as Moonshot, Anthropic, and OpenAI ship new generations every few months, each requiring a fresh multi-week post-training cycle just to stay current, is the open question a $48B valuation is now pricing in.

---

**Sources**: [Cognition — SWE-2 announcement](https://cognition.com/blog/swe-2)