SWE-2 pricing: real cost per million tokens
What Cognition's SWE-2 really costs across 1 plan from 1 provider: monthly fees, usable allowances and real per-token prices, against the official API list price and independent leaderboards.
Data updated · How real prices are computed
As of 2026-10-08, SWE-2 is unmetered on every channel we track (real cost ≈$0).
- Lowest real price
- ≈$0unmetered on every channel
- API list price
- —
- Most monthly tokens
- —
- Plans & APIs
- 11 subscription · 0 metered APIs
Plot SWE-2 on the price–capability chart →
SWE-2 plans and real prices
Real price = monthly fee ÷ tokens you can actually use in a month. “vs API” shows how many times cheaper than the official API list price (standard workload mix) the plan is. “Evidence” opens the full calculation and sources for that row.
| Plan | Provider | Billing | Monthly fee | Monthly tokens | Real $/MTok | vs API | Confidence | Data date | |
|---|---|---|---|---|---|---|---|---|---|
| Devin Pro (promo until 10/31) | Devin | Subscription | $20/mo | — | ≈$0 promo until 2026-10-31, unmetered | — | Medium | Evidence ↓ |
Behind each number
For every plan: the data date, the exact calculation, the quota basis, how leaderboard scores were mapped, and the original adoption evidence with its sources.
Devin Pro (promo until 10/31)Devin≈$0/MTok
- Data date
- 2026-09-10 · official source date
- Real price / MTok
- ≈$0
- Quota confidence
- Medium
- Calculation
- $20 / month ÷ unbounded usage → ≈$0 / MTok · promo until 2026-10-31, unmetered · promotional price, not a permanent allowance
- Benchmark mapping
- Terminal-Bench 4.0: Low (1 configuration)Mapping confidence: how directly each score refers to the model served here.
Adoption evidence · source and rationale (translated from the Chinese record)
Official post 2026-09-10: SWE-2 free for all Pro, Max & Teams subscribers for the next month; in the maintainer's panel the same-segment swe-2 45.5M tokens drew no quota; docs.devin.ai/admin/billing/usage shows no concurrency cap; devin-swe2-round1-2026-09-12.json
Added ≈$0/MTok (recorded as 0): during the SWE-2 promo it draws no quota and is not billed for Pro/Max/Teams, with no concurrency cap → unbounded denominator; the ruled it enters the frontier at the promo price and changes the frontier, ending 2026-10-31 (specified; the official post says only 'for the next month'); the cheapest available tier Pro $20 is used — Max has the same Y, is pricier, and is not drawn twice; after the promo ends the billing weight must be re-checked — the pricing page's permanent-free basis is SWE 1.7, not SWE-2
devin-swe2-round1-2026-09-12.jsonBrowse the public evidence archive
SWE-2 benchmark scores
Highest archived configuration per leaderboard; ranks are among all models each board archives in this dataset. Scores are never mixed across boards.
| Leaderboard | Metric | Score | Rank | Configuration | Effort | Harness | Configs | Snapshot |
|---|---|---|---|---|---|---|---|---|
| Terminal-Bench 4.0 | Resolution Rate % | 27.3 self-reported | #10 / 18 | Devin CLI - SWE-2 (best effort) | — | Devin CLI | 1 | 2026-09-21 |
Every archived configuration
Terminal-Bench 4.0 1 configuration
| Configuration | Score | Score interval | Harness / effort / mode | Source task cost (mean / median) | Checked | Links |
|---|---|---|---|---|---|---|
| Devin CLI - SWE-2 (best effort) [vendor self-report]Devin CLI - SWE-2 (best effort)Vendor self-report, not an official tbench.ai leaderboard run. The Cognition blog's Coding benchmark results table lists Terminal-Bench 4 at 27.3%. The same table's Astra/Sol/Grok 4.6 agree with the official leaderboard; Fable 5.1's 55.8 does not match the official 57.88. No CI, no cost/token data. See devin-swe2-round1-2026-09-12.json. | 27.3 vendor self-report | — | Devin CLI / — / — | — / — | 2026-09-12 | Source · Archive |
These are reference mappings, not benchmarks of a subscription or API channel. Harness and quota-measurement effort alignment are unverified. Source task costs are not subscription task costs.
- Exact served-model reference only; product harness and quota-measurement effort are unverified. Not a benchmark measurement of this subscription or API channel. Vendor self-reported score, not an official leaderboard run.
Frequently asked questions
How does SWE-2 score on independent leaderboards?
SWE-2 has scores on 1 leaderboard; its strongest placement is 27.3 (Resolution Rate %) on Terminal-Bench 4.0, #10 of 18 models tracked.
Who makes SWE-2, and where can I use it?
SWE-2 (model ID swe-2) is made by Cognition and is available through Devin.