Cognition just dropped SWE-2, and the coding-model market will never price the same way again. The new model hits 50.0% on FrontierCode 1.1 Main — within one point of Claude Fable 5.1 — while costing 64% less. And in its first 24 hours on Hacker News, it pulled over 400 points.
The Headline Numbers
Cognition, the company behind Devin, announced SWE-2 yesterday as its "most advanced coding model yet" — and for once, the blog-post superlatives are backed by benchmarks that independent developers can verify in their own terminals:
| Benchmark | SWE-2 | Kimi K3 | Grok 4.6 | Claude Fable 5.1 | GPT-5.6 Sol | GPT-6 Astra |
|---|---|---|---|---|---|---|
| FrontierCode 1.1 Main | 50.0% | 44.2% | 48.0% | 50.9% | 47.5% | 53.3% |
| DeepSWE 1.1 | 73.0% | 68.5% | 67.5% | 67.4% | 72.7% | 74.1% |
| Terminal-Bench 2.1 | 92.8% | 88.3% | 88.4% | 91.4% | 88.8% | 89.9% |
| Terminal-Bench 4 | 27.3% | 21.5% | 20.3% | 55.8% | 37.3% | 57.9% |
The story here is not that SWE-2 wins every benchmark — it clearly doesn't. On the hardest eval, Terminal-Bench 4, it trails GPT-6 Astra by more than 30 points. The story is where SWE-2 sits on the price-performance curve. Cognition's own framing: it "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."
That positioning — "pushing the Pareto frontier," in Cognition's words — is the actual news, and it has implications for every team paying for AI coding agents today.
What SWE-2 Actually Is
SWE-2 is post-trained from Kimi K3, Moonshot AI's 2.8-trillion-parameter model that had already undergone extensive reinforcement learning for agentic coding. Cognition took that base and pushed it further with three technical moves worth understanding, because they represent where the whole industry is heading:
1. Cost penalties inside the RL run
Instead of training one model and hoping it's efficient, Cognition applied a linear cost penalty per reasoning-effort level in a single RL run, with each penalty tuned to the local slope of the base model's Pareto frontier. Translation: the model was explicitly rewarded for not wasting tokens at every price point. This is why SWE-2 offers medium, high, and max effort levels that each occupy sensible positions on the cost curve — the frontier was shaped during training, not after.
2. Length-weighted reward baselines
Training runs are notoriously unstable when reward signals vary with output length. Cognition's length-weighted reward baseline — in use since SWE-1.6 — significantly stabilizes training, which is what makes the multi-effort approach practical at all.
3. A dramatically bigger RL environment set
The team tripled the number of RL environments and added instruction-following overlays. Perhaps most interestingly, they built a "flywheel powered by previous checkpoints of SWE-2 that iteratively hardens our verifiers" — the model's own earlier versions help grade its later versions.
The Efficiency Story: 58% Fewer Turns
For anyone who has watched an AI coding agent burn 15 minutes grep-ing a codebase before writing a single line, this is the number that matters:
- SWE-2 medium takes 53 steps per run on FrontierCode 1.1 Main, versus 127 for SWE-1.7 — a 58% reduction
- Average cost per task dropped 81% versus the previous generation
- The model starts implementing sooner because it judges which parts of the codebase actually matter instead of exhaustively exploring everything
Cognition explicitly addressed the over-exploration complaint: SWE-1.7 was "exceedingly careful," which boosted performance but frustrated users on simple tasks. SWE-2's higher intelligence lets it skip the rituals. Fewer detours, fewer redundant reads, more complete first-pass solutions.
The Reaction: Impressed, But Skeptical
The launch hit the front page of Hacker News within hours and accumulated 400+ points, but the discussion surfaced a familiar tension:
The bullish case: SWE-2 beating Grok 4.6 on both score and cost on FrontierCode and DeepSWE is exactly what maturing RL infrastructure looks like. The base model (Kimi K3) is openly acknowledged, the training recipe is documented in unusual detail, and the multi-effort approach gives developers real control over spend.
The skeptical case: Terminal-Bench 4 — the benchmark designed to resist saturation — tells a different story. SWE-2 scores 27.3% while GPT-6 Astra scores 57.9% and Fable 5.1 scores 55.8%. On the hardest real-world terminal tasks, there's still a wide gulf between "frontier-adjacent at 25% of the price" and "frontier, period." Developers doing long-horizon autonomous work will still reach for the top-tier models; SWE-2 targets the enormous middle of the market where cost per merged pull request is the deciding metric.
There's also the provenance question: SWE-2 is a Cognition model built on a Moonshot base, trained with infrastructure originally developed for SWE-1.7. Whether "post-training house" is a durable business model — or a transitional one before base-model labs internalize these RL recipes — is the open question the comment threads kept circling.
Why This Matters for Your API Bill
Regardless of which model wins any given benchmark, the SWE-2 launch confirms a pricing trend that has been accelerating all year: the cost of competent agentic coding is collapsing.
Two years ago, frontier coding meant paying frontier prices. Today:
- Cognition claims SWE-2 medium matches models retailing at many multiples of its price
- Open-weight and open-recipe models (Kimi K3, GLM, DeepSeek) keep supplying strong bases that post-training shops can push up the curve
- Multi-effort models let you dial spend per task instead of per project
The practical consequence for developers: your default model choice is probably too expensive. If your agent workload is 80% routine refactors, test writing, and small bug fixes, a Pareto-frontier model at a fraction of frontier cost will handle it — and you can escalate only the genuinely hard tasks to GPT-6-class models.
How to Actually Act on This Today
- Audit your current agent spend per task type. Most teams have no idea what a "merged PR" costs them.
- Pick a cheap high-benchmark model and A/B it against your current default on your real backlog, not on public benchmarks.
- Route by difficulty, not by habit. A simple model router that sends easy tasks to an efficient model and hard tasks to a frontier model routinely cuts bills 40-70%.
On Qubax, you can run this experiment in minutes: browse qubax.ai/models, pick any two coding models, and test them on your own tasks at wholesale prices — an open market where compute providers compete on price, so you're not locked into one lab's retail rate card.
What to Watch Next
- Devin integration: SWE-2 is live in Devin Desktop and CLI today, rolling out on Devin Web and Fusion. Watch whether third-party API access follows — that's the moment it becomes directly comparable on price aggregators.
- The "post-training shop" thesis: If Cognition can reliably add 5-6 points to any strong open base, expect more companies to do the same, which compresses prices further.
- Terminal-Bench 4 as the real dividing line: The gap between SWE-2 (27.3%) and Astra (57.9%) defines exactly which workloads still justify frontier pricing.
The coding-model market just got a new reference point on the cost curve. If you're paying frontier prices for routine agent work, today is a good day to re-quote your stack.
Ready to test the new generation of coding models at wholesale prices? Compare SWE-2's base, Kimi K3, against GPT-5.6 Sol and Claude Fable 5.1 on [qubax.ai/models](https://qubax.ai/models).
FAQ
What is Cognition SWE-2?
SWE-2 is Cognition's newest AI coding model, announced September 10, 2026. It's post-trained from Moonshot AI's Kimi K3 (a 2.8T-parameter base) and achieves 50.0% on FrontierCode 1.1 Main, 73.0% on DeepSWE 1.1, and 92.8% on Terminal-Bench 2.1, at roughly 64% less cost than comparable frontier models.
How does SWE-2 compare to Claude Fable 5.1 and GPT-5.6 Sol?
SWE-2 scores within one point of Fable 5.1 on FrontierCode 1.1 Main (50.0% vs 50.9%) and slightly above GPT-5.6 Sol (47.5%), while Cognition claims a 64% cost advantage. On the harder Terminal-Bench 4 eval, however, both Fable 5.1 (55.8%) and GPT-6 Astra (57.9%) still lead by a wide margin.
Is SWE-2 based on Kimi K3?
Yes. Cognition confirmed SWE-2 is post-trained from Kimi K3, Moonshot AI's 2.8-trillion-parameter model, using an expanded RL training setup with tripled environments and per-effort-level cost penalties.
Where can I use SWE-2?
SWE-2 is available now in Devin Desktop and Devin CLI, with rollout underway on Devin Web and Fusion. Its base model, Kimi K3, is already available via multiple API providers, including Qubax, where you can compare it head-to-head with frontier coding models.
Does this mean frontier models are overpriced?
Not for every workload. SWE-2's own benchmarks show frontier models like GPT-6 Astra still lead by 30+ points on the hardest agentic tasks. But for routine coding work, the price-performance data strongly suggests most teams should route easy tasks to efficient Pareto-frontier models and reserve frontier models for genuinely difficult jobs.
