Cognition Ships SWE-1.7 Into Devin at 1,000 Tokens Per Second
On 8 July 2026, Cognition launched SWE-1.7, its most capable software engineering model to date, deploying it directly inside Devin — the AI software engineering agent that became commercially available in 2024 as the first fully autonomous coding assistant marketed to enterprise engineering teams. The delivery mechanism matters as much as the model: SWE-1.7 runs at 1,000 tokens per second through Cerebras wafer-scale hardware, giving Devin an inference speed that turns multi-step software engineering tasks from waiting periods into near-real-time feedback loops. At that throughput, iterative debugging passes, test generation cycles, and patch-and-verify sequences that would take minutes on GPU clusters complete in seconds.
How SWE-1.7 Was Trained: RL on Top of RL
SWE-1.7 was not trained from a raw pretrained base. Cognition built it on top of Kimi K2.7 Code — Moonshot AI's open-weight coding model, which had already undergone extensive reinforcement learning training targeted at software engineering tasks. Cognition then applied a second, large-scale reinforcement learning pass on top of that already-RL-trained foundation. The training spanned four data centres across three continents and represents what Cognition describes as RL on top of RL: a compound strategy that specialises an already-specialised model further, rather than training breadth first.
The rationale is that software engineering, unlike general language tasks, rewards iterative refinement — the ability to observe an error, adjust a strategy, and try again. That maps well to RL-based training that rewards completed objectives over supervised imitation. Stacking RL on a model already tuned for coding reinforces those iterative repair behaviours rather than diluting them with general-language training signals.
Benchmark Results: Near-Frontier at Under Two Dollars Per Task
Cognition reports SWE-1.7 achieving 42.3% on FrontierCode 1.1 Main, 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual. These scores position it near the frontier: it trails Anthropic's Opus 4.8 by a few points on each benchmark but beats GPT-5.5 on SWE-Bench Multilingual and sits within approximately one percentage point on FrontierCode 1.1 Main.
The cost picture is where the positioning becomes commercially meaningful. SWE-1.7 completes FrontierCode Main tasks at approximately 1.97 US dollars each, placing it on a better cost-performance trade-off curve than frontier models that achieve modestly higher benchmark scores at significantly higher per-task cost. For engineering teams running large volumes of automated coding tasks — test generation, code review, refactoring, documentation — the per-task cost compounds materially over a month of high-volume use.
Comparing SWE-1.7 Against the Benchmark Field
At 81.5% on Terminal-Bench 2.1, SWE-1.7 sits close to Grok 4.5 at 83.3%, while using roughly a quarter of the output tokens on similar tasks. Output token cost is the primary driver of per-task expenditure in agentic coding workflows, where models generate substantial intermediate reasoning and code before committing a final edit. SWE-1.7's token efficiency at competitive benchmark levels is its clearest commercial differentiator against models that score marginally higher but generate proportionally more tokens to do so.
Devin Security Swarm: Cognition's Other July Launch
Cognition shipped a second significant product in the first week of July 2026: Devin Security Swarm, an AI system built to find exploitable vulnerabilities in enterprise codebases, validate them at runtime, and generate the pull requests to remediate them. The timing reflects a growing problem: monthly security findings across enterprise engineering organisations are climbing from around 1,000 to more than 10,000, driven in part by the roughly 42% of code that is now AI-generated or AI-assisted.
On a benchmark of 50 real-world vulnerabilities drawn from published GitHub Security Advisories across 14 programming languages, Devin Security Swarm found 36 — more than any other AI-powered scanner tested — at 30% lower cost per finding than the next most accurate alternative. Three of those vulnerabilities were missed by every other tool tested. Cognition also offers a structured six-week Devin Security Programme to help enterprises assess their application security posture and clear existing vulnerability backlogs.
What This Means for Engineering Teams in India
For Indian software services firms and product companies managing large codebases, SWE-1.7 and Devin Security Swarm represent two converging points of leverage. Near-frontier benchmark performance at under two dollars per coding task makes high-quality AI-assisted engineering economically viable at the task volumes that characterise outsourced delivery or rapid-iteration product development. The 1,000 TPS inference speed means agentic coding loops run at a cadence compatible with developer flow rather than requiring asynchronous waits.
The Security Swarm capability is particularly relevant for teams delivering software to regulated sectors — banking, insurance, healthcare — where vulnerability backlogs represent compliance risk as much as engineering debt. The ability to automatically validate exploitability and generate remediation pull requests, rather than simply listing findings, closes the loop between discovery and resolution within the same developer workflow.
The Bottom Line
Cognition launched SWE-1.7 on 8 July 2026 — its most capable software engineering model — deployed inside Devin at 1,000 tokens per second via Cerebras hardware. SWE-1.7 is built using reinforcement learning applied on top of Kimi K2.7 Code, which was itself already RL-trained. Benchmark results of 42.3% on FrontierCode 1.1 Main, 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual position it near frontier model quality at approximately 1.97 US dollars per task. The same week, Cognition shipped Devin Security Swarm, which outperformed all other AI vulnerability scanners on a 50-issue real-world benchmark at 30% lower cost per finding, completing the picture of an agentic platform covering both build and security phases of software delivery.
Frequently Asked Questions
What is Cognition SWE-1.7 and when was it launched?+
Cognition SWE-1.7 is a software engineering AI model launched on 8 July 2026 and deployed inside Devin, Cognition's AI software engineering agent. It is Cognition's most capable model to date, built using reinforcement learning applied on top of Kimi K2.7 Code — an open-weight coding model that was itself already extensively RL-trained. SWE-1.7 runs at 1,000 tokens per second via Cerebras wafer-scale hardware, enabling near-real-time feedback loops for iterative coding tasks. Benchmark results include 42.3% on FrontierCode 1.1 Main, 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual.
What does RL on top of RL mean for SWE-1.7's training?+
RL on top of RL refers to the compound training strategy Cognition used to build SWE-1.7. Kimi K2.7 Code, the foundation model, had already undergone large-scale reinforcement learning training focused on software engineering tasks. Cognition then applied a second round of reinforcement learning on top of that already-RL-trained model, spanning four data centres across three continents. The result is a model that has been iteratively specialised for software engineering twice over, reinforcing iterative error-repair behaviours — the ability to observe a failure, adapt, and try again — that RL training is particularly effective at developing.
How does SWE-1.7 compare to frontier models on benchmarks?+
SWE-1.7 scores 42.3% on FrontierCode 1.1 Main, 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual. It trails Anthropic's Opus 4.8 by a few points on each benchmark but beats GPT-5.5 on SWE-Bench Multilingual. Its key differentiator is cost: it completes FrontierCode Main tasks at approximately 1.97 US dollars each, placing it on a better cost-performance curve than frontier models that achieve marginally higher benchmark scores. It also uses roughly a quarter of the output tokens that comparable-scoring models use on Terminal-Bench 2.1, which further reduces per-task cost in high-volume agentic workflows.
What is Devin Security Swarm and what can it do?+
Devin Security Swarm is an AI security product from Cognition, launched in the first week of July 2026, designed to find exploitable vulnerabilities in enterprise codebases, validate them at runtime to confirm they are genuinely exploitable, and generate the pull requests to fix them. On a benchmark of 50 real-world vulnerabilities from GitHub Security Advisories across 14 programming languages, it found 36 — more than any other AI-powered scanner tested — at 30% lower cost per finding than the next most accurate alternative. Three vulnerabilities found by Devin Security Swarm were missed by every other tool tested.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.