What is SWE-1.7?
SWE-1.7 is Cognition's in-house coding model, made by running its own reinforcement learning on Moonshot's Kimi K2.7 and training it inside the Devin agent for long, unattended jobs. Near the end of its context it writes a summary of its progress and carries on from it, and it reads widely through a codebase before it edits.
Cognition now calls it its previous flagship. SWE-2 replaced it, and Cognition says SWE-2 at medium effort beats it on FrontierCode while costing 81% less on average. SWE-1.7 and the faster SWE-1.7 Lightning are still in Devin's model list, and Cognition lists them only there.
Is SWE-1.7 still available?
Yes. SWE-1.7 and SWE-1.7 Lightning are still in Devin's model list, but Cognition now calls SWE-1.7 its previous flagship and points users to SWE-2.
- Successor
- SWE-2, Cognition's current coding model, in Medium, High and Max effort.
- Cost per task
- SWE-2 at medium effort costs 81% less on average than SWE-1.7 on FrontierCode 1.1 Main, by Cognition's figures.
- Speed to first edit
- SWE-2 makes its first real edit after a median of 18 steps, against 48 for SWE-1.7.
- Why keep it
- Standard SWE-1.7 has no listed token price on paid self-serve plans, and Lightning still offers Cerebras speed.
Who is SWE-1.7 best for?
SWE-1.7 is best for paid Devin users who hand off long bug fixes and unattended coding jobs and want an in-house model with no listed token price, or Cerebras speed from its Lightning version.
Paid Devin users watching their quota
Devin lists no token price for standard SWE-1.7 on its paid self-serve plans, and its docs suggest the SWE models for routine work to make the quota last.
Bug fixes that need a root cause
Cognition trained it to look past the reported symptom and test edge cases, often with small scripts, before it changes code.
Long unattended jobs
It summarises its working state and resumes from the summary, and single tasks in its training ran for up to six hours.
Fast back-and-forth
SWE-1.7 Lightning runs on Cerebras hardware for lower latency, at $2.50 / $12.50 per million tokens.
What's included with SWE-1.7?
Medium and Max effort, a faster Lightning version on Cerebras, self-compaction for long tasks, deep codebase reading and short reasoning, across Devin Web, Desktop and CLI.
- Medium and Max effort
- Devin offers it at two effort levels, Medium and Max, in its model picker.
- SWE-1.7 Lightning
- The same model served on Cerebras for lower latency; the Devin CLI's /fast command picks it when it is available.
- Self-compaction
- Near its context limit it writes a summary of where it is and continues from that summary.
- Deep codebase reading
- It makes more tool calls, file reads and searches per task than Kimi K2.7 Code, Opus 4.8 or GPT-5.5, by Cognition's count.
- Short reasoning
- Its chain of thought uses about half the words per sentence of its Kimi base.
- Across Devin
- Available in Devin Web, Devin Desktop and the Devin CLI.
How much does SWE-1.7 cost?
SWE-1.7 needs a paid Devin plan from Pro at $20 a month, where the standard model has no listed token price; SWE-1.7 Lightning costs $2.50 per million input tokens and $12.50 per million output tokens.
- Cache input $1.00 per million tokens
- Served on Cerebras for lower latency
- Draws down a paid plan's quota, then extra usage
- Standard SWE-1.7 with no listed token price
- Daily and weekly quota, extra usage at API prices
- Cloud agents in Devin Cloud
- Up to 200 users
- Central billing and an admin dashboard with analytics
- Priority support
- Everything in Pro
- Much higher usage quotas than Pro
SWE-1.7 is not on Devin's Free plan, which only runs SWE-1.6. On paid self-serve plans Devin lists no token price for standard SWE-1.7 at Medium or Max effort; enterprise plans billed in ACUs pay $0.50 per million input tokens and $2.50 per million output tokens for it. Usage past a plan's quota is billed at API list prices. Enterprise is priced on request and adds SAML or OIDC SSO and a dedicated deployment option. Devin shows monthly prices only. Full pricing on devin.ai. We checked these prices on .
How do readers rate SWE-1.7?
No one has rated SWE-1.7 yet. Be the first.
Sign in to rate. One rating per account.
SWE-1.7 pros and cons
It has no listed token price on paid Devin plans and digs into bugs carefully, but it runs only in Devin, overthinks simple tasks and has been overtaken by SWE-2.
Pros
- Devin lists no token price for standard SWE-1.7 on its paid self-serve plans, and Hacker News and Reddit users describe it as included with the $20 Pro plan.
- Cognition's tests put it level with GPT-5.5 at a much lower cost per task: 42.3% on FrontierCode 1.1 Main against 43.0%, and 81.5% on Terminal-Bench 2.1.
- Cognition found it much more likely than its Kimi base to chase a bug's root cause and check edge cases instead of patching the symptom.
- SWE-1.7 Lightning answers quickly on Cerebras hardware, at the 1,000 tokens per second Cognition quotes.
Cons
- It runs only inside Devin: Hacker News users who wanted it in OpenCode, OpenRouter or their own tools found no API and no downloadable weights.
- It overthinks simple jobs: Cognition itself cites that feedback, and Reddit and Hacker News users report it spinning on quick requests that SWE-1.6 handles at once.
- SWE-2 has overtaken it: on Cognition's newer tables SWE-1.7 scores 7.6% on Terminal-Bench 4 against 27.3% for SWE-2, and costs more per task.
- Its best numbers come from FrontierCode, Cognition's own benchmark, which several Hacker News commenters distrust; Cognition also says it touches more files and writes more tests than a task needs.
What are the best SWE-1.7 alternatives?
SWE-2 for Cognition's newer and cheaper-per-task model, Kimi K2.7 Code for open weights in any tool, and Composer 2.5 for Cursor users.
SWE-2PaidChoose it when you want Cognition's current model: at medium effort it beats SWE-1.7 on FrontierCode 1.1 Main while costing 81% less per task, at a list price of $3 / $15 per million tokens.
Kimi K2.7 CodeFreeChoose it when you want the open-weight base SWE-1.7 was built on, usable in any tool through Moonshot's API at $0.95 / $4 per million tokens, though it scores about 12 points lower on FrontierCode in Cognition's tests.
Composer 2.5FreeChoose it when you code in Cursor rather than Devin; it is Cursor's own agentic model with a 200K context, at $0.50 / $2.50 per million tokens.
Questions about SWE-1.7
What is the difference between SWE-1.7 and SWE-1.7 Lightning?
Cognition says they have the same intelligence; Lightning is served on Cerebras hardware for lower latency. Lightning costs $2.50 per million input tokens and $12.50 per million output tokens, while standard SWE-1.7 has no listed token price on Devin's paid self-serve plans.
Who is SWE-1.7 for?
SWE-1.7 is best for paid Devin users who hand off long bug fixes and unattended coding jobs and want an in-house model with no listed token price, or Cerebras speed from its Lightning version.
Is SWE-1.7 free?
No. Devin's Free plan only runs SWE-1.6, so SWE-1.7 needs a paid plan, starting with Pro at $20 a month. On paid self-serve plans Devin lists no token price for standard SWE-1.7, while SWE-1.7 Lightning costs $2.50 per million input tokens and $12.50 per million output tokens.
How much does SWE-1.7 cost?
SWE-1.7's paid plans are SWE-1.7 Lightning at $2.50 / $12.50 per million input / output tokens, Pro at $20 per month, Teams at $80 per month plus $40 per full dev seat and Max at $200 per month.
What are the best SWE-1.7 alternatives?
The closest alternatives on Recaply are SWE-2, Kimi K2.7 Code and Composer 2.5.
Where can I try SWE-1.7?
On its official website, cognition.com.
Can I use SWE-1.7 outside Devin?
No. Cognition offers SWE-1.7 only in Devin Web, Devin Desktop and the Devin CLI, and publishes no API or weights for it. If you need a model for your own tools, its base, Kimi K2.7 Code, has open weights on Hugging Face and its own API.
Which model is SWE-1.7 built on?
Cognition trained SWE-1.7 from Moonshot's Kimi K2.7 with its own reinforcement learning, inside the Devin agent. It reports 42.3% on FrontierCode 1.1 Main for SWE-1.7, against 30.1% for Kimi K2.7 Code, and says training made it explore more carefully and reason in shorter sentences.
Our verdict
SWE-1.7 is now a fallback inside Devin, not a first pick.
It fixes bugs carefully and has no listed token price on paid Devin plans, but Cognition's own tests show SWE-2 scoring higher at a fraction of the cost per task, and SWE-1.7 often overthinks simple work.
- Skip it for
- Anyone who works outside Devin, since it runs nowhere else, and anyone who wants quick answers to small requests.
- Check first
- Run a few of your own tasks on SWE-1.7 and SWE-2 and compare what each takes from your quota.
SWE-grep
Rnj-1
KAT-Coder-Pro V2
Mellum2
Ornith-1.0
GPT-5.3-Codex