How Much Does SWE-2 Cost?
Cognition has not published a standalone per-token price for SWE-2. The model is sold through Devin subscriptions, and the cheapest paid tier named in launch reporting is Devin Pro at $20 per month.
That is the answer most launch coverage skipped in favour of the benchmark chart. SWE-2 shipped on 10 September 2026, post-trained from Moonshot AI's Kimi K3, and the launch framing is about cost rather than raw capability.
The ladder below is what launch reporting states. Read every figure as reported rather than confirmed, because reported is the strength of the sourcing that exists for this release.
| Devin plan | Reported price | SWE-2 at launch |
|---|---|---|
| Pro | $20 per month | Reported free for one month from launch |
| Max | No price I can verify | Reported free for one month from launch |
| Teams | $80 per month plus $40 per developer seat | Reported free for one month from launch |
A frontier-class coding model handed to a $20 subscriber at no incremental cost for a month is an unusually direct offer, and it expires. AI Perks tracks the credit programs that cover the model and infrastructure bills sitting behind a coding agent.
What SWE-2 Actually Scores
SWE-2 is reported at 50.0% on FrontierCode 1.1 Main, within one point of Claude Fable 5.1, and Cognition claims it reaches that at roughly 64% lower cost than comparable frontier models.
Every number below is vendor-reported, arrives through launch coverage rather than through an independent evaluation, and has not been reproduced by anyone outside Cognition. That is the normal state of affairs for a model a week old, and it is the reason nothing here should be read as settled.
| Benchmark | SWE-2 reported score | Comparison available |
|---|---|---|
| FrontierCode 1.1 Main | 50.0% | Within one point of Claude Fable 5.1 |
| DeepSWE 1.1 | 73.0% | No rival figure I can verify |
| Terminal-Bench 2.1 | 92.8% | No rival figure I can verify |
Two claims are worth separating. The first is the absolute score: 50.0% on FrontierCode 1.1 Main places SWE-2 in the same conversation as frontier coding models. The second is the cost framing, which is the actual news. Parity is cheap to claim and expensive to verify; a large cost gap at near parity is what changes a build decision.
What this page cannot give you is the rest of the grid. Scores for the GPT-class models, for Grok, for the Kimi K3 base and for Cognition's own previous SWE generation are absent from the available sourcing, so they are absent from the table. A chart that fills those cells is asserting more than the public record supports.
How Far the 64% Cheaper Claim Actually Goes
The 64% figure is a relative cost claim made at launch, measured against unnamed comparable frontier models. There is no rate card underneath it, because no standalone SWE-2 price exists.
That matters more than it sounds. A per-token comparison needs two rates, and Cognition has published neither a rate for SWE-2 nor the baseline the 64% is measured against. The claim cannot be reconstructed from public numbers, only tested on your own workload.
Three consequences follow for anyone budgeting around it.
- The saving is directional, not contractual. Nothing in a Devin subscription is denominated in that 64%.
- The baseline is unstated. Comparable frontier models is a category, not a model, and the cost spread inside it is wide enough to move the figure materially.
- Benchmark parity is not task parity. Three strong benchmark rows do not tell you how a model behaves on long, unstructured agentic work, and no public comparison of that kind exists yet.
The sane configuration is both models. Run the cheap model by default, escalate to a frontier model when a task fails twice, and you capture most of the saving without betting a workflow on a claim you cannot audit. AI Perks tracks credits across the providers at both ends of that route.
What the Free Month Is, and What Happens After It
Launch reporting says SWE-2 was offered free for one month across Devin Pro, Max and Teams subscriptions. It is a promotional window inside a paid plan, not a free product.
Three mechanics matter more than the headline.
It is bounded in time. One month, running from the 10 September launch. No published statement I can verify says what replaces the window, whether it extends, or what the model consumes once it closes.
It is not a free tier. The offer rides on subscriptions people already pay for. There is no published route to SWE-2 at no cost for someone holding no Devin plan at all.
The surfaces differ. SWE-2 was available in Devin Desktop and the Devin CLI at launch, with Devin Web and Fusion described as rolling out. No date has been published for either of those two.
If you are evaluating Devin Pro, a month that includes the vendor's newest model on promotional terms is a materially different test from the month you will buy once the window closes. Time the evaluation accordingly, and fund the surrounding stack from AI Perks rather than from revenue.
What Cognition Has Not Published
The numbers people ask about SWE-2 most often are the ones with no public answer as of 17 September 2026. Being explicit about the gaps is what separates a page that stays right from one that quietly rots.
A standalone API price. There is no per-million-token rate for SWE-2 in any source I can verify. The 64% figure is a relative claim, not a rate card. An article quoting you a tidy per-token SWE-2 price is filling in a blank the vendor left empty.
The context window. No figure for SWE-2 appears in the available sourcing, and a base model's specification does not automatically carry over to the model post-trained from it.
Credit allowances and quota consumption. Nothing published on how much of a plan's allowance an SWE-2 task draws, how allowances refresh, or what usage beyond them costs.
The price of the Max plan. Max is named in the free-month offer, but the figure attached to that plan is not something I can verify here.
The variant lineup. Whether SWE-2 ships in more than one size, and at what relative cost, has not been published in any form I can confirm.
What happens after the free month. Nothing published, in either direction.
How to Cut the Real Bill Behind a Coding Agent
Start with the credit layer. getaiperks.com filters to the AI Tool, Developer Tools and Cloud Infrastructure categories. A coding subscription is the small bill. The model APIs, compute and infrastructure around it are the large one.
Spend the promotional month on the hard work. A window with no incremental cost is worth most on the tasks you would not normally risk quota on. Run the awkward refactors, not the autocomplete.
Route by your own evidence, not by brand. Vendor benchmarks and a vendor cost claim are a hypothesis. Push the same twenty real tickets through both models, log the retries, and let your own numbers choose the default.
Fund the API layer separately. Credits rarely pay a coding subscription directly, but they do cover the model calls, inference and hosting that agent workflows generate, and that is the line item that scales with usage.
Re-check when the window closes. Terms around a week-old release move quickly, and the ones on this page are reported rather than confirmed. Anything budgeted against a promotional month deserves a revisit once that month ends.
Frequently Asked Questions
How much does SWE-2 cost per token?
Cognition has not published a per-token price for SWE-2. It is consumed through Devin subscriptions rather than a standalone API rate card. Launch reporting puts Devin Pro at $20 per month and Devin Teams at $80 per month plus $40 per developer seat; prices for the other tiers are not something I can verify. Credits for the surrounding stack are tracked at getaiperks.com.
Is SWE-2 really free right now?
Launch reporting says SWE-2 was offered free for one month across Devin Pro, Max and Teams subscriptions, starting from the 10 September release. That is a dated promotion inside a paid plan rather than a free product, and no published statement I can verify describes the terms once the month ends.
Is SWE-2 better than Claude Fable 5.1?
The only comparison available is FrontierCode 1.1 Main, where SWE-2's reported 50.0% sits within one point of Claude Fable 5.1. That reads as parity rather than a win for either model. SWE-2's reported 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1 have no verified rival figures attached, so a broader ranking would be invention.
What model is SWE-2 built on?
SWE-2 is post-trained from Kimi K3, the Moonshot AI base model reported at 2.8 trillion parameters. How much that post-training adds over the base is not published in any figure I can verify. Credits for Moonshot and the wider stack are at getaiperks.com.
Where can I use SWE-2?
SWE-2 was available in Devin Desktop and the Devin CLI at launch. Devin Web and Fusion were described as rolling out, with no date published for either surface, so treat availability there as pending rather than live.
Can free AI credits offset a coding agent subscription?
Provider credits are usually scoped to that provider's own usage, so they rarely pay a Devin subscription directly. What they do cover is the model APIs, compute and hosting that agent workflows consume, which is the larger and faster-growing bill for most teams. AI Perks tracks credit programs across those layers at getaiperks.com.
Cheap models, expensive retries. Know which is which, and let someone else fund the stack at getaiperks.com.