SWE-2 Flash vs DeepSeek V4 Flash: One Has a Rate Card, the Other Has a Ratio

The first thing that breaks when you put SWE-2 Flash and DeepSeek V4 Flash side by side is the unit. DeepSeek publishes dollars per million tokens and a checkpoint you can download. Cognition publishes a fraction — one-fifth of SWE-2 — and a seat inside Devin. One of these is a model you can build on; the other is a model you can use. Every comparison that follows has to survive that asymmetry, and most of the ones circulating on launch day did not.

That is not a reason to skip the comparison. It is the reason the comparison is worth making carefully. The two models are aimed at the same buyer — a team that wants frontier-adjacent coding behaviour without frontier pricing — and they make opposite bets about how that buyer wants to receive it. One bet is disaggregated: weights, an API, a published context window, and a bill you can forecast. The other is integrated: a model tuned for one harness, sold as a fraction of another model's cost, with the serving stack included in the number.

What each one actually is, and the checkpoint trap

SWE-2 Flash was announced on 4 November 2026. It is post-trained from Inkling-Small, Thinking Machines' 276B-total, 12B-active mixture, and it lives in Devin Desktop and Devin CLI. It is free there through 30 November. There is no API, no public weights, and no published context window.

DeepSeek V4 Flash is a 284B-total, roughly 13B-active mixture under the MIT licence, with a 1M-token context window and up to 384,000 output tokens. It takes text only; the vision variant is a separate experimental ID. It is served over an OpenAI-compatible API and has been callable since the spring, with a re-post-trained 0731 checkpoint that did most of the work on agentic coding.

Then the trap. DeepSeek's own models and pricing page, read on 5 November 2026, lists deepseek-v4-flash among the legacy names still accepted by the API and says it is now served by DeepSeek-V4.1-Flash-1017. That is a different and larger model: Cognition's comparison table puts V4.1 Flash 1017 at 763B total and 16B active. So "DeepSeek V4 Flash" resolves to at least three things depending on who is speaking — the April build, the 0731 checkpoint, or the October V4.1 checkpoint the ID now points at. On OrcaRouter the pinned IDs are separate routes, which is the only reason a comparison can be reproducible at all.

This is not pedantry. It changes which benchmark row you are entitled to quote. Public leaderboard figures for "DeepSeek V4 Flash" were measured on the 0731 build. Cognition's table measured V4.1 Flash 1017. They are not the same model, and the numbers differ accordingly.

Price, in the two units each vendor actually publishes

DeepSeek's card is arithmetic. Per million tokens for the model the deepseek-flash ID now serves — DeepSeek-V4.1-Flash-1017, which the page names explicitly — read on 5 November 2026: input at $0.15 off-peak and $0.30 in the peak windows, output at $0.60 off-peak and $1.20 in peak, cached input at $0.003 and $0.006. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Outside those windows the cheaper rate applies.

The pinned older checkpoint is priced differently. OrcaRouter's catalogue carries deepseek/deepseek-v4-flash at $0.22 per million input tokens and $0.66 per million output, a flat rate with no peak tier. That is the price of reproducibility: pin the April-lineage model instead of calling the current one and you pay more, in exchange for the model most of the published benchmark rows were actually measured on. Neither choice is wrong; they answer different questions.

Cognition's card is a comparison. "One-fifth the price and 3x the speed of its much larger sibling" is the entire public statement, with no absolute figure attached, and free through 30 November on top of it. There is no way to compute what a task costs on SWE-2 Flash from public information, because the denominator is a model whose own token price was never published either.

Two consequences follow, and both are practical rather than pedantic.

• Forecastability — DeepSeek's cost per million tokens is knowable to the cent, including the peak multiplier. SWE-2 Flash's cost is knowable only as a fraction of an unpublished number, inside a subscription you are already paying for.

• Flat quotes are misleading — the pinned V4 Flash route is a single flat rate, while the vendor's current card is two-tier. A flat number is therefore wrong for part of the week, and it describes a different checkpoint than the one most benchmark rows name. Plan against the vendor card; reproduce against the pinned route.

Both matter more than they look. For an agent that runs thousands of short calls, the peak multiplier is a scheduling problem you can engineer around; for a coding agent that runs a handful of long sessions, it is a rounding error. Which of those you are determines whether DeepSeek's card is two prices or one.

Speed: a replay against a measurement

Cognition's 200 tokens per second is not a live timing. The post explains it: each model's own run on Terminal-Bench 4.0's gsea-proteomics task, "replayed at a simulated 200 vs. 60 tokens per second", with transcripts shortened to equal length. SWE-2's 60 tok/s comes from the same replay. It describes the serving profile Cognition is targeting, which is a legitimate thing to publish, but it is not a stopwatch on two endpoints.

Artificial Analysis measures DeepSeek V4 Flash at 112 output tokens per second on its current model page, with a 32.41-second time to first answer token. That is a measurement of a live API, including whatever queueing the provider does. The two numbers are not comparable, and subtracting them would be meaningless: one is a simulated replay inside a product, the other is an observed serving rate across the public endpoint.

What survives the caveat is the direction. If a 12B-active model really serves at roughly 200 tok/s inside Devin while a 13B-active model measures 112 tok/s on a public API, the smaller-active model is not obviously the faster one — the serving stack is doing as much work as the architecture. That is a point in favour of integrated products and against reading speed off a parameter count.

The one same-table run, and who ran it

Cognition's launch table includes DeepSeek's model directly. The column is DeepSeek V4.1 Flash 1017 — the checkpoint the deepseek-v4-flash ID now resolves to — which makes this the only comparison of the pair produced under one roof. Every figure below is Cognition's, on Cognition's benchmarks, with no harness disclosed.

• FrontierCode 2.0 — SWE-2 Flash 24.0 vs DeepSeek V4.1 Flash 1017 19.3

• FrontierCode 1.1 Main — 45.5 vs 43.9

• DeepSWE 1.1 — 60.1 vs 77.0

• Terminal-Bench 4.0 — 20.0 vs 42.3

• Terminal-Bench 2.1 — 89.2 vs 91.0

• SWE-Atlas QA — 50.0 vs 71.0

• SWE-Bench Pro v2 Hard — 89.9 vs 86.3

• ExploitBench — 34.8 vs 31.1

The split is four rows each, and the margins are nothing like symmetric. SWE-2 Flash wins FrontierCode 2.0 by 4.7 points, FrontierCode 1.1 Main by 1.6, SWE-Bench Pro v2 Hard by 3.6 and ExploitBench by 3.7. DeepSeek wins DeepSWE 1.1 by 16.9, Terminal-Bench 4.0 by 22.3, SWE-Atlas QA by 21.0 and Terminal-Bench 2.1 by 1.8. Every Flash win is a handful of points. Two of the DeepSeek wins are north of twenty.

Weight the comparison by margin rather than by win count and it stops being close. The Flash is genuinely competitive on FrontierCode, which rewards a bounded patch against a written specification, and it is far behind on Terminal-Bench 4.0, which measures a terminal agent carrying a task to completion. That is also the row where the rest of the field separates: GLM 5.3 Flash scores 41.8 on the same benchmark and Claude Haiku 5.5 scores 39.2, against the Flash's 20.0.

Now the caveats that belong in the same breath. This is a competitor's table. FrontierCode is Cognition's own instrument, and Cognition grades it. The post discloses no harness, no effort setting and no run count, so "same table" is not the same as "same conditions" — it is one vendor's methodology applied uniformly, which is better than nothing and worse than independent. And the DeepSeek column is V4.1 Flash 1017, not the 0731 build that most public numbers describe.

There is a useful sanity check on how much the instrument matters. Artificial Analysis' evaluation of the 0731 build of DeepSeek V4 Flash, last run on 31 July 2026, records 12.1% on Terminal-Bench 4.0. Cognition's table records 42.3% for the V4.1 Flash 1017 checkpoint. Thirty points apart on the same named benchmark. Part of that is the checkpoint — V4.1 is a different and larger model — and part is the harness, since one is a standardised third-party run and the other is a vendor's own. Either way, no single row in this comparison is a property of a model alone, and the rows where the two vendors disagree most are the ones to hold most loosely.

Context, weights and what you can build

• Context window — DeepSeek 1,048,576 tokens published vs SWE-2 Flash not published

• Maximum output — DeepSeek 384,000 tokens vs SWE-2 Flash not published

• Weights — DeepSeek MIT, downloadable vs SWE-2 Flash gated, partner program promised "in a few weeks"

• Modality — DeepSeek V4 Flash text-only, with vision arriving on the V4.1 checkpoint and on a separate experimental ID vs SWE-2 Flash not stated

• Distribution — DeepSeek API and third-party platforms vs Devin Desktop and Devin CLI only

• Effort control — DeepSeek thinking and non-thinking modes vs SWE-2 Flash not documented

• Licence to fine-tune — DeepSeek yes vs SWE-2 Flash no, not yet

For a team building a product, five of those seven rows are disqualifying on their own. You cannot plan a retrieval pipeline around an unpublished context window, you cannot fine-tune weights you do not have, and you cannot put a dependency in your architecture that only exists inside somebody else's IDE. SWE-2 Flash is not competing for that slot. It is competing for the seat a developer already occupies.

The one row where the Flash has an advantage is the one Cognition is least clear about: cost per task. If a fifth of SWE-2's price is a fifth of the cost of getting a job done, it beats DeepSeek's per-token card for interactive work inside Devin, because the subscription is already paid. If it is a fifth of the token price on a model that generates more tokens per task, the advantage shrinks. Cognition does not publish enough to tell which.

Where a router changes the arithmetic

DeepSeek V4 Flash is routed on OrcaRouter as deepseek/deepseek-v4-flash, at the provider's list price with no markup added, which matters for a model whose price changes twice a day: a vendor cut or a peak-window change reaches the route rather than waiting for a platform to reprice it. The same key reaches more than 200 models, so the sensible version of this comparison — run your queue against DeepSeek V4 Flash, GLM 5.3 Flash and Kimi K3 and see which one your tasks actually like — costs a model-ID change rather than three contracts. Automatic failover across providers is the other half: DeepSeek's peak windows are exactly when a second route earns its place, and a fallback chain is cheaper than a pager alert.

SWE-2 Flash is not on OrcaRouter and cannot be. There is no endpoint to route to, which is the honest summary of the whole matchup. If you want it, you buy Devin. If you want to call a coding model from your own software, the choice was made for you before the benchmarks were published.

Which one to pick

If you need a coding model in production — an API, a context window you can size against, weights you can pin, a bill you can forecast — DeepSeek V4 Flash is the only one of the two that qualifies, and the October checkpoint is the one to pin. The peak-window discount is worth engineering around, and the 1M context is worth having even if you rarely fill it.

If you already pay for Devin, SWE-2 Flash is close to a free option until 30 November. Run it on the bounded end of your queue, keep SWE-2 for the multi-step work the DeepSWE gap describes, and treat the free window as the evaluation period it is. The risk is not that it is bad. The risk is that a ratio is being used where a measurement should be.

If you are standardising a team on one coding model, this is not the matchup to decide it. One candidate can be tested and one cannot, and a comparison in which only one side is observable is a distribution decision wearing a benchmark costume. Pick the one you can measure, then revisit when Cognition ships an endpoint or a partner programme that puts the weights somewhere you can reach.