A smaller model for the ordinary work
Cognition's pitch for SWE-2 Flash is not that it beats its flagship. The company says smaller models can now handle a significant share of what software agents are actually asked to do: automations, code review, feature development, documentation updates. Flash is built for that tier of work, and Cognition is positioning it as the lower-cost option. The company claims it costs one-fifth as much as SWE-2 within Devin and demonstrates a threefold speed advantage in a simulated replay. It is available in the same Devin Desktop and CLI surfaces, free through November 30.
The build is worth reading carefully. Cognition started from Inkling-Small, Thinking Machines' mixture-of-experts model with 276 billion total parameters and 12 billion active, and added a stage it has not used before: a large mid-training phase, on a few trillion tokens of scientific writing, blogs and code, before its usual reinforcement learning. The company says it chose the data mix by measuring how many tokens each candidate mix needed to match a baseline's validation loss, and saved the scarcest, highest-quality data for the final stage.
The scores, in their actual context
The headline gains are measured against the base model, not against rivals. Mid-training and RL together took Inkling-Small from 0.9% to 20.0% on Terminal-Bench 4.0, and from 3.2% to 24.0% on FrontierCode 2.0. Those are large improvements over the base model. Cognition additionally claims that Flash outperforms every American open-weight model included in its evaluation, including models three times larger. That comparison is also based on the company's own testing and has not been independently reproduced.
Two cautions belong next to the numbers. First, the absolute scores are still modest: 20.0% and 24.0% mean the model did not complete the large majority of tasks in these evaluations. That makes Cognition's routine-work positioning more relevant than any suggestion that Flash is a general-purpose coding leader. Second, the speed demonstration is a replay. Cognition's post shows each model on the same Terminal-Bench task "replayed at a simulated 200 vs. 60 tokens per second", with transcripts shortened to equal length. It illustrates the 3x claim; it does not document a measured throughput test. The demonstration should therefore be read as a controlled simulation of the speed difference, not evidence that Flash delivers three times the real-world throughput across ordinary Devin workloads. And the price claim has no dollar figure attached anywhere in the announcement. One-fifth of SWE-2's price is a ratio a reader cannot check against an invoice until Cognition publishes what either model costs on its own.
The open base is not the shipped model
The part of this launch that will outlast the benchmark chart is the weights decision. Flash is described by Cognition as a "fully American coding model", and it starts from Thinking Machines' openly published Inkling-Small. That makes the starting model inspectable, but not the complete training process reproducible. Cognition has withheld the trained Flash weights and has not published enough information to recreate the finished model independently. The trained model itself is a different matter. Cognition says it deliberately trained Flash for cybersecurity work, raising its ExploitBench score from 11.3% to 34.8%, and that the same capability could serve attackers. So the weights stay private for now.
In their place comes a promise: a program to share the weights securely with trusted industry partners and research labs, with details "in a few weeks". No start date, no eligibility criteria, no named partners. Enterprises and governments that want to host and inspect the model, the very customers Cognition says need that control for sensitive work, are being asked to wait for a program that does not yet exist in any checkable form.
That may be the right safety call. The ExploitBench score rising from 11.3% to 34.8% during training is a concrete reason Cognition gives for slowing down the release. The company argues that the resulting cybersecurity capabilities could also be useful to attackers. But readers should keep the categories separate. An open base model plus a closed post-trained model adds up to a closed product, and the gated program, if it arrives, will be a distribution channel for selected partners, not a public release.
What to watch
Three things will show whether this launch holds up. Independent runs of Terminal-Bench 4.0 and FrontierCode 2.0, to test the speed and score claims outside Cognition's harness. A published price, so the one-fifth claim can be checked in dollars. And the sharing program itself: who qualifies, under what terms, and whether "a few weeks" turns into a date. Until those land, Flash is a promising lower-cost option inside one company's product, announced with more candor about its limits than most, and verified by no one outside it.
Sources
- [1] Cognition — "Introducing SWE-2 Flash", November 4, 2026 (primary source, read first-hand in full at Gate 1: availability in Devin Desktop and CLI at one-fifth SWE-2's price and 3x its speed; free through November 30; Inkling-Small base at 276B total / 12B active parameters; a mid-training stage added for the first time before reinforcement learning; Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0% against the base model; generation speed shown as each model's own run "replayed at a simulated 200 vs. 60 tokens per second", transcripts shortened to equal length; ExploitBench from 11.3% to 34.8% and "not sharing the weights publicly yet"; a program to securely share the weights with trusted industry partners and research labs, "more on this in a few weeks"; no standalone dollar price and no API price published. All figures are Cognition's own, with no independent replication)Read source
- [2] Thinking Machines Lab — "Introducing Inkling-Small" and the Inkling-Small model card on Hugging Face (read first-hand at Gate 1: a sparse mixture-of-experts transformer, 276B total / 12B active parameters, released with full open weights; the repository's license field states Apache 2.0)Read source
- [3] Unite.ai — "Cognition Launches SWE-2 Flash for Faster, Lower-Cost Coding Agents", November 4, 2026 (secondary; launch-table context only, not used for this article's figures)Read source
- [4] Superpower Daily — "Cognition adds a lower-priced coding model to Devin, keeping its weights private" (secondary; corroboration of the weights decision and base-model context only)Read source