SWE-2 Flash vs Qoder Cantus: Two Coding Models You Can Only Rent

SWE-2 Flash and Qoder Cantus have almost nothing in common except the thing that decides this comparison: you cannot call either of them. Cognition's SWE-2 Flash shipped on 4 November 2026 and runs only inside Devin Desktop and the Devin CLI. Qoder Cantus, Alibaba's built-in top-tier model for its Qoder coding platform, has run only inside Qoder's own surfaces since July 2026. No endpoint, no model ID, no weights you can download. Two models, two agent products, and no shared benchmark on which either has been independently tested.

That sounds like a dead end for a versus page. It is actually the whole story, because it changes what you are comparing. This is not a bake-off between two checkpoints. It is a choice between two rented rooms, and the useful questions are which room publishes its meter, which one publishes its homework, and which of those two things you would rather have when the invoice arrives.

Eight rows against an empty column

Cognition's launch post is unusually generous with numbers. SWE-2 Flash reports eight benchmark rows, two training deltas, a speed figure and a price figure. Qoder has published no benchmark score for Cantus at all — not on its own site, not on Artificial Analysis, not on LMArena — along with no parameter count, no architecture, and no context window.

It is tempting to read that as eight rows to zero. It is closer to eight claims to zero claims, and the distinction matters more than the count.

• FrontierCode 2.0 — 24.0, against 23.8 for the larger SWE-2 and 3.2 for the Inkling-Small base it was post-trained from.

• FrontierCode 1.1 Main — 45.5, against SWE-2's 50.0.

• DeepSWE 1.1 — 60.1, against SWE-2's 73.0.

• Terminal-Bench 4.0 — 20.0, against SWE-2's 27.3 and 0.9 for the base model.

• Terminal-Bench 2.1 — 89.2, against SWE-2's 92.8.

• SWE-Atlas QA — 50.0, against SWE-2's 56.3.

• SWE-Bench Pro v2 Hard — 89.9, against SWE-2's 90.6.

• ExploitBench — 34.8, against SWE-2's 36.1.

Every one of those is vendor-reported. Cognition does not publish the harness, the reasoning-effort setting, or whether the figures are single-attempt or best-of-N, and FrontierCode is Cognition's own benchmark. So the eight rows are not eight independent facts. They are eight statements about which a reader can at least form a view, and that is a different and lesser thing.

Cantus's empty column is not the same kind of weakness. A missing number can be filled in later. A claim that was never made cannot be checked, and Qoder has had since July to make one.

The row that actually predicts this matchup

Read the SWE-2 Flash row against its own larger sibling and a shape appears. On SWE-Bench Pro v2 Hard, a benchmark of well-specified patches, Flash scores 89.9 to SWE-2's 90.6 — a gap of seven-tenths of a point, which is nothing. On DeepSWE 1.1 and Terminal-Bench 4.0, where the agent has to work unsupervised for longer with less scaffolding, the gaps open to 12.9 points and 7.3 points.

That is the signature of a model that holds up on scoped work and thins out as the horizon lengthens. It is a perfectly reasonable design choice for a fast, cheap tier, and it is the single most decision-relevant thing in the launch post.

Now hold that next to how Cantus is sold. Qoder's own launch page describes it as built for "extended autonomous task execution" — the long-horizon, low-supervision end of the spectrum. That is precisely the axis where SWE-2 Flash's own numbers are weakest.

So the one dimension on which these two models genuinely compete is also the one where one side has a mediocre published score and the other has no score at all. You cannot resolve it with the available evidence, and anyone who tells you otherwise is comparing a vendor table to a marketing adjective.

What each one costs, in the only units published

Here the asymmetry flips, and it flips hard.

SWE-2 Flash is free through 30 November 2026, which is the best possible price and also the least durable one. After that, Cognition's only pricing statement is that Flash costs "one-fifth the price" of SWE-2. SWE-2 has no published rate card either. The ratio is real inside Devin's billing and terminates in nothing a reader can compute, so the actual post-November cost is unknown.

Cantus works the other way. Qoder publishes a reference factor per model and a median credit cost per task mode, which is enough to do arithmetic. On Qoder's credits page, an Editor agent-mode request at a 200K context window is estimated at about 12 credits; Editor Ask at about 4; a Qoder agent-mode request at 200K at about 50.

The multiplier on top of that moved. Qoder's release notes set a new reference factor effective 00:00 Singapore time on 21 September 2026, raising Cantus from 3.2x to 4x. So the same Editor agent turn that cost roughly 38 reference credits in August now costs 48 — a 25% increase for an unchanged model. Qoder publishes no official dollar-per-credit rate; top-up packs are widely reported at around $20 per 1,500 credits, which puts that turn near $0.64 at add-on rates, and lower when paid from a plan allowance.

Put plainly: the model with eight benchmark rows has a price you cannot calculate, and the model with zero benchmark rows has one you can. That is the most useful sentence on this page.

And the factor table is worth reading in full, because it prices Cantus against everything else in the same building. At 4x it sits five times GLM-5.3's 0.8x, eight times Qwen3.8-Max's 0.5x, and forty times DeepSeek-Flash's 0.1x. If your work inside Qoder is mostly scoped edits rather than long autonomous runs, Cantus is the most expensive way to do it by a wide margin — a comparison Qoder publishes itself and that almost no coverage of the model bothers to make.

Context: 200K documented, the other undisclosed, and one assumption to drop

Cantus has no published context window, but Qoder's credit estimates are denominated at 50K and 200K, and 200K is the operative figure — that is the window Qoder's own billing model is calibrated against.

SWE-2 Flash's context window is not stated anywhere in its launch post. The trap here is inheritance: SWE-2 runs on Kimi K3, which carries a 1M-token window, and it is easy to assume Flash does too. It does not follow. Flash is post-trained from Inkling-Small, a different base with a different lineage, and Cognition has said nothing about the window it serves. Treat it as unknown rather than large.

• Context — Cantus 200K in Qoder's billing model vs SWE-2 Flash undisclosed, and not inheritable from SWE-2.

• Parameters — Cantus undisclosed vs Inkling-Small's 276B total / 12B active upstream.

• Weights — Cantus never promised vs SWE-2 Flash withheld, with a partner programme promised "in a few weeks".

• Independent testing — neither has any.

Two locked rooms, two meters

Both models are single-surface, and the surfaces are not interchangeable. Cantus appears across Qoder's IDE, JetBrains plugin, CLI, QoderWake, Cloud Agents and mobile/web app — broad inside Qoder, zero outside it. SWE-2 Flash appears in Devin Desktop and the Devin CLI — narrower, but for teams already on Devin it is a model switch, not a migration.

The meters are also more volatile than a rate card. Cantus's factor moved 25% in one notice with no model change, and its newer sibling Sonus launched in September at 8x. Cognition's "one-fifth" ratio is a launch-period number that will be repriced once the free window closes on 30 November.

One practical wrinkle worth knowing if you are on Qoder: users reported through October that Cantus and Sonus disappear from the desktop model picker when a proxy is configured, with disabling the proxy as the workaround. It is a client bug, not a model problem, but it is the kind of thing that turns a long autonomous run into a wasted afternoon.

For completeness, since it is the obvious next question: neither SWE-2 Flash nor Qoder Cantus is on OrcaRouter, because neither vendor offers an API to route. What is on OrcaRouter is the set of models the same Cognition table scores on Terminal-Bench 4.0 at 41.8 and 42.3 — GLM 5.3 Flash and DeepSeek V4.1 Flash 1017 — at $0.15 per million input tokens against $0.50 and $0.60 per million output, with 1M-token context windows. Those are Cognition's measurements in Cognition's harness, so read them as directional rather than certified, but the direction is not subtle: both are callable, both outscore SWE-2 Flash's 20.0 on that row, and both sit behind one key with 0% markup on provider list price and automatic failover if a provider degrades. If your reason for reading this page was finding a coding model you can put in a pipeline rather than a seat, that is the honest answer, and it is not either of the two in the title.

Who should pick which

• Already on Devin, work is scoped — use SWE-2 Flash now. It is free until 30 November, it is close to SWE-2 on well-specified patching, and the cost of trying it is a model-selector change.

• Already on Devin, work is long and unsupervised — stay on SWE-2, or route a callable model. Flash's 20.0 on Terminal-Bench 4.0 is the row that governs this, and it is 7.3 points behind the model it is cheaper than.

• On Qoder for compliance or region reasons — Cantus is the top tier and the only one positioned for extended autonomy, but at 4x it is eight times Qwen3.8-Max's factor for the same task estimate. Reserve it for the runs that need it.

• Need an API, or need to compose models, or need to fail over — neither. That requirement rules both out before any benchmark is consulted.

The thing to watch is not a benchmark. It is whether Cognition attaches a rate card to SWE-2 Flash when the free window closes on 30 November, and whether Qoder ever publishes a single score for Cantus. Until then the honest verdict is that one model is better documented and the other is better priced, and neither is verifiable — which is a strange pair of facts to have to weigh, and the reason this matchup resists the usual answer.