QWE AI Academy The #1 mistake people are making with SWE-1.7 right now: they see the headline (“near GPT-5.5 and Opus intelligence”) and immediately swap it in as their default model. Then they wonder why their PRs balloon to 14 files touched when they asked for a two-line fix.
SWE-1.7 just dropped and the timeline is loud about it. Here’s what the model actually is, when it wins, when it doesn’t, and how to run it in Devin without setting your ACU meter on fire.
What just happened with SWE-1.7
Cognition – the company behind Devin – dropped SWE-1.7 on July 8, 2026. It launched inside Devin at 1,000 tokens per second via Cerebras, and Cognition pitched it as frontier-class agentic coding at a fraction of the cost.
The technical short version: SWE-1.7 is RL post-trained on a Kimi K2.7 Code base, done inside the real Devin agent use. That last part matters more than the base model – the training environment is the actual product, so the model learns behaviors that transfer directly to Devin sessions.
The headline numbers, per Cognition’s own launch data:
| Benchmark | SWE-1.7 | GPT-5.5 | Opus 4.8 |
|---|---|---|---|
| FrontierCode 1.1 Main | 42.3% | 43.0% | 46.5% |
| Terminal-Bench 2.1 | 81.5% | 84.2% | 86.9% |
| SWE-Bench Multilingual | 77.8% | 76.8% | 84.4% |
| Cost per task (FrontierCode) | $1.97 | higher | higher |
Notice the third row. On SWE-Bench Multilingual, SWE-1.7 actually beats GPT-5.5 (77.8% vs 76.8%) – the opposite ordering from the other two benchmarks. “Near frontier” is doing a lot of work in that phrase.
Why “just pick SWE-1.7” is the wrong move
Every tutorial writing about this today is going to show you that benchmark table and stop there. Here’s what they’re leaving out.
The benchmark pulling the most marketing weight – FrontierCode – is Cognition’s own. That doesn’t make it fake, but it’s worth being honest about the setup. Both Cognition and Cursor tend to rank their own models highly on their own evals (HN thread, July 2026): training data and evaluation appear to come from the same Devin/Cursor interaction logs, so the models naturally overfit to their own use. That’s the skeptic’s read, anyway.
The other thing that will surprise you in real usage: SWE-1.7 thinks more than the task asks. The tradeoff for its “long-horizon reasoning” is increased change scope – it writes additional test cases and touches more files than the task naively requires (per the AlphaSignal writeup). If you’re trying to open a tight, reviewable PR, that’s a bug, not a feature.
Pro tip: Before asking SWE-1.7 to fix a bug, add an explicit constraint to your prompt: “Change only the minimum files needed. Do not add new tests unless they’re required to verify the fix.” This kills 80% of the scope creep that would otherwise burn extra ACUs and expand your diff.
The honest decision matrix (when to pick what)
Forget picking a “best” model. Pick by task shape.
- Long-running async task that Devin runs while you’re at lunch – SWE-1.7 wins. It’s specifically tuned for long-horizon asynchronous software engineering tasks, and the 1,000 tok/s speed compounds over multi-step trajectories.
- Long terminal session, sandboxed CI reproduction, shell workflows – GPT-5.5. It leads SWE-1.7 by 2.7 points on Terminal-Bench 2.1 (84.2% vs 81.5%), which is the clearest single-benchmark separation between the two models.
- Cross-language migration or a codebase with 6+ languages – SWE-1.7 (surprisingly). It’s the only benchmark where it edges GPT-5.5.
- Ambiguous architectural refactor across a large codebase – Opus 4.8, if you can afford it. Higher raw scores across the board and better MCP tool use for agentic workloads.
- Quick chat-style code review or explanation – none of the above. You’re wasting compute on a specialist model.
How to actually run SWE-1.7 in Devin
SWE-1.7 lives inside Devin. You can’t hit it as a standalone API the way you would GPT-5.5.
Basic path:
- Go to devin.ai and grab a plan. As of July 2026, the tiers are: Free, Pro at $20/month, Max at $100/month, Teams at $500+, and Enterprise custom – check the pricing page for any updates, as these change frequently.
- Open Devin (Web, Desktop, or the CLI – all three support SWE-1.7).
- In the model dropdown above the input, select SWE-1.7. As of the July 2026 launch, it’s free for paid users for approximately one month – the cheap window to test it side by side with Opus and GPT-5.5.
- Write your task the way you’d write a ticket, not a chat message. Devin was built to consume tickets. Give it: goal, acceptance criteria, files to focus on, files to avoid.
- Let it plan first. Don’t accept the first plan blindly – SWE-1.7 tends to over-plan.
A real example: bug fix with scope control
Say you have a Node service where parseUserAgent() mis-classifies iPad-on-Safari as desktop. Here’s the wrong prompt and the right prompt.
❌ Wrong:
"Fix the parseUserAgent bug where iPads are treated as desktop."
✅ Right:
"In src/utils/parseUserAgent.js, the current regex fails to
match iPad Safari (which reports MacIntel + touch support).
Update ONLY parseUserAgent.js and its existing test file
(src/utils/parseUserAgent.test.js). Add one new test case
for iPad Safari. Do not touch any other files. Do not
refactor unrelated code."With the second prompt, SWE-1.7’s long-horizon instinct gets a fence around it. You’ll still get its edge-case reasoning (it’ll probably ask about iPadOS 13+ user-agent spoofing), but the diff stays reviewable.
The ACU trap nobody’s warning you about
SWE-1.7 costing $1.97 per task on Cognition’s benchmark does not mean your monthly Devin bill will be low.
Devin bills in Agent Compute Units. One ACU is roughly 15 minutes of active Devin work; a typical bug fix runs 2-3 ACUs, but a multi-file migration runs 30+. The overage rate sits at $2.25/ACU (as of July 2026 – verify at devin.ai/pricing). And because consumption is non-linear and can’t be paused mid-task, complex or iterative tasks can burn through your monthly allocation fast.
SWE-1.7’s tendency to expand scope makes this worse, not better. More files touched = more tool calls = more ACUs. The per-task benchmark number is a floor, not a ceiling.
Watch your first ten tasks like a hawk. Track ACU spend per ticket. If a “simple bug fix” is costing you 8+ ACUs, something in your prompt is letting SWE-1.7 wander.
The bigger question the launch doesn’t answer
If a specialized coding model trained inside a specific agent use can land within a few points of frontier general-purpose models on coding benchmarks – and beat them on cost – what does that say about the value of the frontier labs’ generalist strategy for coding specifically?
Nothing definitive, honestly. The benchmarks are self-selected and the sample is one model. But it’s the first release where a scrappy, application-focused lab has meaningfully closed the gap on Anthropic and OpenAI for the specific task of writing mergeable code. Worth watching.
FAQ
Is SWE-1.7 open source or available on other IDEs?
No. It’s proprietary and only runs inside Devin (Web, Desktop, CLI). No weights, no API for third-party tools.
Should I cancel my Claude or ChatGPT subscription now?
Almost certainly not. SWE-1.7 is a specialist addition for agentic coding tasks inside a specific use – not a swap-out for general use. Opus 4.8 is still ahead on raw coding benchmarks and dominates tool-heavy MCP workloads; GPT-5.5 leads terminal-heavy work. The practical move is running the same task through two models in Devin’s Arena Mode for a week before making any subscription decisions.
Why did SWE-1.7 jump so much from previous versions?
Two reasons, per Cognition. They switched to Kimi K2.7 Code as the base model – a stronger starting point than what previous versions used. Then they refined the RL pipeline and ran it inside the actual Devin agent use, so training feedback now matches production behavior rather than a synthetic environment. The AlphaSignal writeup called the FrontierCode jump “too large to be a rounding artifact” – it’s a fundamentally different training run.
Next step: Open Devin, pick a real ticket from your backlog (not a toy task), and run it once with SWE-1.7 and once with your current default. Compare diff size, ACU cost, and review time. That’s the only benchmark that matters.