A model that is one-fifth the price of its bigger sibling sounds like an easy switch. On November 4, 2026, Cognition, the company behind the Devin coding agent, launched SWE-2 Flash with exactly that pitch: one-fifth the price and 3x the speed of SWE-2. The same post carries an 8-row benchmark table, and that table is where the decision actually gets made. Flash beats SWE-2 on 1 of the 8 rows, and it beats Anthropic's budget model, Claude Haiku 5.5, on 1 of 8 as well.
That does not make it a bad model. A 276-billion-parameter model landing within 0.7 points of a 2.8-trillion-parameter one on a hard coding test is a real result. It does mean the headline needs reading closely, because "one-fifth the price" has no per-token number behind it, the 200 tokens per second is a simulated replay, and the weights are not public.
Quick summary: SWE-2 Flash launched November 4, 2026, in Devin Desktop and Devin CLI, free through November 30. It has 276B total and 12B active parameters and is built on Thinking Machines' Inkling-Small. Cognition reports one-fifth the price and 3x the speed of SWE-2, but publishes no per-token rate for Flash and no public API, and every benchmark here is Cognition-reported. Weights are withheld, with a partner program promised "in a few weeks."
What Cognition actually shipped
SWE-2 Flash is a mixture-of-experts model with 276B total parameters and 12B active, the same size as its base, Inkling-Small, which Thinking Machines Lab released on July 30, 2026. Cognition's size chart puts SWE-2 at 10.1x larger, at 2.8T total and 104B active. Cognition pitches Flash at the everyday share of agent work, naming automations, code review, feature development and documentation updates. (Sources: Cognition, Thinking Machines)
The bigger change may be the base model. SWE-2, launched September 10, 2026, was post-trained from Moonshot's Kimi K3, a Chinese open-weight model. Flash starts from an American one, and Cognition calls it "a fully American coding model." Coverage describes Inkling-Small as Apache 2.0 licensed. Why that matters is our inference, not Cognition's statement: US companies and agencies that are wary of Chinese-origin models now have a Devin option that sidesteps the question. (Sources: Cognition on SWE-2, Superpower Daily)
The training recipe is also new for the company. Cognition says it added a large-scale mid-training stage before reinforcement learning for the first time, using a few trillion tokens of scientific articles, blogs, code and other sources. It credits most of the reinforcement-learning gains to better and more varied tasks, not to changes in the recipe itself. The result it highlights: Terminal-Bench 4.0 rose from 0.9% on the base model to 20.0%, and FrontierCode 2.0 rose from 3.2% to 24.0%. The post does not separate how much each stage contributed. (Sources: Cognition, Unite.AI)
The 8-row table, in plain numbers
Here are 5 of the 8 benchmarks from Cognition's own comparison. Higher is better. These are vendor-reported, none has been independently reproduced in the coverage we reviewed, and FrontierCode is a benchmark Cognition itself built. Unlike the September SWE-2 post, this one does not explain which harness or reasoning-effort settings were used.
| Model | FrontierCode 2.0 | DeepSWE 1.1 | Terminal-Bench 4.0 | SWE-Atlas QA | ExploitBench |
|---|---|---|---|---|---|
| SWE-2 Flash | 24.0 | 60.1 | 20.0 | 50.0 | 34.8 |
| SWE-2 | 23.8 | 73.0 | 27.3 | 56.3 | 36.1 |
| Claude Haiku 5.5 | 27.7 | 59.3 | 39.2 | 56.5 | 38.0 |
| GPT-6 Luna | 21.9 | 66.6 | 16.4 | 45.2 | 44.7 |
| DeepSeek V4.1 Flash 1017 | 19.3 | 77.0 | 42.3 | 71.0 | 31.1 |
| GLM 5.3 Flash | 14.2 | 63.4 | 41.8 | 56.5 | 36.0 |
| Muse Spark Open | 6.2 | 60.0 | 8.0 | 54.0 | Not listed |
| Beam 1.0 | 2.9 | 43.3 | 4.0 | 33.9 | 19.5 |
| Inkling-Small (base) | 3.2 | 17.4 | 0.9 | 27.4 | 11.3 |
Counting all 8 rows in Cognition's table, including FrontierCode 1.1 Main, Terminal-Bench 2.1 and SWE-Bench Pro v2 Hard, Flash wins 1 of 8 against SWE-2, 1 of 8 against Haiku 5.5, 4 of 8 against DeepSeek V4.1 Flash 1017, 4 of 8 against GLM 5.3 Flash, and 6 of 8 against GPT-6 Luna. Three patterns stand out. It is close to SWE-2 on bounded tasks, such as 89.9 versus 90.6 on SWE-Bench Pro v2 Hard, but falls further behind on DeepSWE 1.1 and Terminal-Bench 4.0. Its one win over SWE-2, on FrontierCode 2.0, is by 0.2 points. And for scale, Claude Opus 5.5 leads FrontierCode 2.0 at 45.5%, so nothing in this budget tier is near the top of that test. (Sources: Cognition, BenchLM)
Three claims that need an asterisk
"One-fifth the price"
Cognition's wording does not say per token, and the launch post gives no rate for either model. Devin's model docs, which do not name Flash yet, do list SWE-2 at $3 per million input tokens and $15 output after its promotion. If one-fifth applied per token, Flash would land near $0.60 and $3, but that is our arithmetic, not a published price. There is no standalone API, so you cannot compare this directly with a rival's price sheet. If the one-fifth holds per task, a rough adjustment for success rates still favors Flash: our arithmetic gives about 4.1x cheaper per solved task on DeepSWE 1.1 and about 3.7x on Terminal-Bench 4.0. That rests on an unconfirmed assumption and ignores the time spent reviewing failed runs. (Sources: Cognition, Devin docs, OrcaRouter)
"3x the speed"
The 200 versus 60 tokens per second comparison is, in Cognition's own caption, each model's run on one Terminal-Bench 4.0 task, "replayed at a simulated" speed with transcripts shortened to equal length. That illustrates a serving target. It is not a measured live throughput, and it says nothing about full task completion time. (Sources: Cognition, Superpower Daily)
"Outperforms all American open-weight models"
Cognition adds "even models that are 3x larger," which points at Meta's Muse Spark Open, listed at 816B total. The DeepSWE 1.1 margin there is 60.1 to 60.0, and the same table shows Muse Spark Open ahead on SWE-Atlas QA, 54.0 to 50.0. Flash also is not openly available itself. On the same day, Ricursive launched Ripple-65B Preview and claimed the US open-weight lead on Terminal-Bench 2.1 at 80.52, a claim that holds only among public models, since Cognition lists gated Flash at 89.2. (Sources: Cognition, Ricursive)
The weights decision
Cognition deliberately trained Flash for security work, lifting ExploitBench from 11.3% to 34.8%, and says those same skills could help attackers, so "we are not sharing the weights publicly yet." It is building a program for trusted industry partners and research labs, with fine-tuning and sensitive-environment use as the stated goal, and promises more "in a few weeks." Cognition's table puts Flash below GPT-6 Luna (44.7), Haiku 5.5 (38.0) and SWE-2 (36.1) on that benchmark, so the risk argument is about downloadable, unmonitored weights rather than Flash being unusually capable. That last reading is ours. (Sources: Cognition, Unite.AI)
What the cheap tier costs elsewhere
Since Flash has no published token price, the fair way to frame it is against rivals you can actually buy by the token. Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 per million input tokens and $0.50 per million output, though Haiku jumps to $0.50 and $2.50 once a prompt passes 100,000 tokens, which long agent sessions can reach. DeepSeek V4.1 Flash 1017 lists at $0.30 and $1.20 at peak hours, and half that off-peak. These API rates are not directly comparable with a price inside Devin, which depends on your plan and quota. Even so, the $0.60 and $3 implied above would be 6x the list price of Haiku 5.5 or GPT-6 Luna. (Sources: Anthropic pricing, OpenAI pricing, DeepSeek pricing, VentureBeat)
Who should try it before November 30
Cognition's post says only "Free through 11/30." Alpha Signal and Runtime Wire both describe that as a benefit for paying Devin users, and Alpha Signal names Pro, Max and Teams, so check your own plan's model picker. Devin's docs say SWE-2 itself is also free on self-serve plans in Desktop and CLI through November 30, so the one-fifth saving only matters once that promotion ends. Devin Pro is $20 a month and Max is $200, per Devin's docs. (Sources: Alpha Signal, Runtime Wire, Devin billing docs, Devin model docs)
- Good fit for a test: scoped pull requests, code review passes, documentation updates and scheduled automations, the jobs Cognition itself names.
- Keep a stronger model for long, multi-step work. Flash's biggest gaps to SWE-2 are on DeepSWE 1.1 (60.1 vs 73.0) and Terminal-Bench 4.0 (20.0 vs 27.3).
- Poor fit if you need an API, since Devin Desktop and Devin CLI are the only surfaces, or if you must self-host today, since the weights are not available.
- Measure your own numbers during the free window: tasks merged without rework, reviewer minutes per task and retries per task. A benchmark table cannot tell you those.
Run the same 10 real tickets from your backlog through Flash and your current default model before November 30, then compare merged results and review time. Do not budget for December until Cognition's post-promotion pricing is clear in your plan.
01What is SWE-2 Flash?
A coding model Cognition launched on November 4, 2026, with 276B total and 12B active parameters, built on Thinking Machines' Inkling-Small and available in Devin Desktop and Devin CLI.
02How much does SWE-2 Flash cost?
Cognition says one-fifth the price of SWE-2 and free through November 30, 2026. It has not published a per-token rate in the launch post, and reports indicate the free access applies to paying Devin users. SWE-2 is also free on self-serve Desktop and CLI plans through that date, per Devin's docs, then lists at $3 input and $15 output per million tokens.
03Can I use SWE-2 Flash through an API?
Not based on the launch materials. Coverage says Cognition has not announced a public API, and it runs only in Devin Desktop and Devin CLI.
04Are the weights open?
No. Cognition says it is not sharing them publicly yet because of cybersecurity concerns, and it plans a gated program for trusted partners and research labs. Its base model, Inkling-Small, has public weights.
05Is SWE-2 Flash better than Claude Haiku 5.5?
On Cognition's own table, usually not: Haiku 5.5 leads on 7 of 8 rows, while Flash leads on DeepSWE 1.1 by 60.1 to 59.3. Costs cannot be compared directly because Flash has no token price.
06Is the 200 tokens per second figure real?
It comes from a simulated replay of one Terminal-Bench 4.0 run, so treat it as an illustration of target speed rather than a measurement.
Comparing budget coding models on your own prompts is the only test that counts. LumiChats gives you 40+ AI models under one subscription, so you can see how the fast tier handles your work before you commit to any single tool.
