Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra

TL;DR Highlight

RL post-training Kimi K3 put SWE-2 within 1 point of Fable 5.1 at 64% lower cost.

Who Should Read

Backend or AI developers choosing coding-agent models by cost and performance, or considering their own RL post-training.

Core Mechanics

SWE-2 is one point behind Fable 5.1
- On FrontierCode 1.1 Main, SWE-2 scores 50.0% and Fable 5.1 scores 50.9%.
- It costs 64% less than Fable 5.1.
Built on Kimi K3 with more RL
- It is post-trained from the 2.8T-parameter Kimi K3.
- RL added 5-6 points over K3 on many benchmarks.
Efficiency comes from less wandering
- On FrontierCode 1.1 Main, SWE-2 medium scores higher than SWE-1.7.
- It takes 58% fewer turns and costs 81% less on average.
- Its first real edit comes after a median of 18 steps versus 48 for SWE-1.7.
All effort levels trained in one RL run
- The reward is R = S - λe·C: success minus a cost penalty.
- λe is matched to the slope of the base model's cost-performance curve at each effort level.
- Only a linear penalty gives the same result whether averaging before or after, so they chose linear.

Community Reactions

Skepticism about the Terminal-Bench 4 gap
- Many commenters doubted the results because TB 2.1 is 92.8% but the newer TB 4 is 27.3%.
- They read it as a sign of how well the model generalizes to new problems.
Mixed hands-on experience
- One commenter found SWE-1.6 fast and good for small tasks, but SWE-1.7 thought longer than GLM 5.2 and ran in circles.
- Another said people around them in a Devin shop praise SWE only for being free.
Token efficiency should have been shown
- A commenter noted the benchmark page used to show output tokens and time but now shows only cost.
- They suspect most gains come from fewer output tokens.
Complaints about Devin lock-in and closed weights
- One commenter won't try it because it seems to require Devin.
- Another said they'd pick DeepSeek Flash 4.1 over a closed-weight model.

How to Apply

Plot cost and success rate per effort level for your agent
- Run the same 20 tasks today at medium and high.
- Compute extra cost per point of success rate and use it to choose effort.
Compare benchmarks across generations when picking models
- Put old and new benchmark scores side by side.
- Validate models with a large gap on your own tasks.
Measure steps to first edit in agent traces
- Find tasks with excessive explore (read/grep/ls) steps.
- Add a prompt limiting exploration scope and compare turns and cost.
Use a linear cost penalty in your own RL rewards
- Build the reward as success minus λ times cost.
- Set λ to the slope of the base model's cost-performance curve per effort level.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the paper's method: cost-performance slope)

Below are results from running our coding agent at each effort level. Each row is (effort, average cost in USD, success rate).
{{effort results table}}

Do the following.
1. Treat cost as x and success rate as y, and compute the slope between adjacent effort levels (success-rate gain / cost increase).
2. Find the segment with the lowest slope and judge whether moving up to it is cost-inefficient.
3. If we use reward R = S - λ·C with success S and cost C, give a table of λ values per effort level (the slopes above).

Next: rerun the same task set at different effort levels and update the table.

Terminology

Pareto frontierThe curve of choices where you can't improve cost or performance without giving up the other. Think of it as a map of the best value-for-money options.

RL (reinforcement learning)Training where a model tries things, gets scored, and reinforces high-scoring behavior. It's like taking a test and adjusting how you study based on the grade.

reward baselineA reference value, like an average, subtracted from rewards to reduce noise in the learning signal. It's grading on a curve to cut noise.

speculative decodingA small draft model proposes several tokens and the big model verifies them together to speed up decoding. Like a junior writing a draft and a senior skimming it.

KL divergenceA measure of how different two probability distributions are. Here it means the behavior gap between the inference-side and training-side models.

reward hackingWhen a model scores points by exploiting loopholes in the verifier instead of solving the real problem. Like finding holes in the grading rubric instead of answering the exam.