Cognition SWE-2: the vendor table, the cost claims, and what is missing
Cognition released SWE-2 on 2026-09-10, calling it their most advanced coding model yet, and aiming at the Pareto frontier of capability versus inference cost: 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while being 64% cheaper. It is post-trained from Kimi K3, and the post publishes no endpoint, no price and no rate limit. (checked 2026-09-20).
Updated 2026-09-20
One key across Claude, GPT and Chinese models, billed per token.
- Billed per token — live rates on /models
- Pay by card (Visa / Mastercard), Apple Pay, Google Pay or crypto
- Self-service: top up and activate instantly
The key numbers in the official table
FrontierCode 1.1 Main
The vendor's own table: on FrontierCode 1.1 Main, SWE-2 scores 50.0%, within one point of Fable 5.1 and 64% cheaper; the same row lists Kimi K3 44.2%, Grok 4.6 48.0%, Fable 5.1 50.9%, GPT-6 Astra 53.3% and SWE-1.7 42.0% - all vendor-reported.
Claimed reduction versus Fable 5.1
The comparison is the vendor's own: within one point of Fable 5.1 and 64% cheaper. That 64% is Cognition's own price comparison — the post gives no price per million tokens, and no third party has re-run the comparison.
Terminal-Bench 4 (the weak column in the same table)
In the Terminal-Bench 4 column of the same table: SWE-2 27.3%, Kimi K3 21.5%, Grok 4.6 20.3%, Fable 5.1 55.8%, GPT-5.6 Sol 37.3%, GPT-6 Astra 57.9%, SWE-1.7 7.6% - Cognition published the weak column as well, so read it next to the FrontierCode result.
Base-model parameters (Moonshot's figure)
The post states that SWE-2 is post-trained from Kimi K3, on a 2.8T-parameter base model. The 2.8T figure describes the base model K3, not SWE-2's own size; separately, the post says this round of RL was scaled to the multi-trillion-parameter regime for the first time.
What SWE-2 is
SWE-2 is the coding model Cognition (the company behind Devin) released, described as their most advanced coding model yet, with competitive agentic coding across multiple effort levels. It was not trained from scratch: the post says it is post-trained from Kimi K3 with RL, and that this round of RL still adds 5–6 points on many benchmarks. On release day it went live in Devin Desktop and Devin CLI, with Devin Web and Fusion following. The route is product subscription plus agent orchestration — the post lists no public endpoint, no price per million tokens and no rate limit.
Release cadence
Cognition positions SWE-2 as their most advanced coding model to date; it is available in Devin Desktop and Devin CLI the same day, with Devin Web and Fusion rolling out afterwards. At the time of the check that post is still the only source for the score table.
Timeline
2026-09-10: Cognition releases SWE-2. The official table gives 50.0% on FrontierCode 1.1 Main, 73.0% on DeepSWE 1.1, 92.8% on Terminal-Bench 2.1 and 27.3% on Terminal-Bench 4, and says it is 64% cheaper than Fable 5.1.
2026-09-10: Another point from the same post: SWE-2 is post-trained from Kimi K3 (2.8T), reusing the SWE-1.7 training infrastructure and recipe, and the company says RL still finds substantial headroom, adding 5-6 points on many benchmarks.
2026-09-20: Searching the public QCode model catalog for SWE-2, Devin and Cognition turns up no match, so SWE-2 is not offered here; the coding pages linked below point at models we actually sell.
Stated versus not yet sayable
What the post states
The comparison the post offers: SWE-2 matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of their price, and comes within a few points of GPT-6 Astra at a quarter of the cost. None of that is an independent measurement — no third party has reproduced it.
Do not read these as facts
The official post 'Introducing SWE-2: Pushing the Pareto Frontier' only gets as far as the product side: SWE-2 has entered Devin Desktop and CLI, and Devin Web and Fusion are rolling out. The post gives no public API endpoint or model id, no price per million tokens, no rate limit, no context length, no training data or license, and says nothing about open weights. Every score and cost comparison comes from Cognition's own table and its own cost basis, and no third party has re-run them; the post itself says this round of RL still adds 5–6 points on many benchmarks, so the numbers keep moving.
SWE-2 versus running your own coding agent
Cognition SWE-2 (inside the product)
Model plus Devin's agent orchestration in one product, available in Desktop and CLI from day one. The trade-off is subscription rather than per-token billing: with no endpoint, unit price or rate limit published, cost per task can only be computed on their terms.
Sold general models plus your own harness
Run coding tasks on models QCode sells: swap the model id, pay per token, keep rate limits and billing in your own console. The trade-off is that orchestration, context management and the verification loop are yours to build, and effort levels have to be approximated.
What to look at if you want the same setup
Task splitting: the vendor scores SWE-2 across effort levels, and a self-built stack has no such internal tiers, so the usual move is to split locating, patching and testing into separate calls and route the simple steps to cheaper models. The verification loop: the post describes the training-side change as a flywheel built from SWE-2's own earlier checkpoints that repeatedly hardens the verifiers rather than the model; in engineering terms, tests and lint have to be automated or every model's score drops. The cost basis: the reduction the vendor quotes is their own comparison, not a unit price — to compare properly, run the same repository and the same batch of issues, record tokens and retries, and then see what the run actually cost you.
What is actually available on QCode
SWE-2 is not offered. The public QCode model catalog returns no match for SWE-2, Devin or Cognition, and the 30-day metered-usage table does not list them either, so the site has neither an entry for it nor a price for it. To put coding-agent work on billable calls, the route is sold general models plus schema/tool calling with outer validation and retries. Billed per token; per-model rates are listed on /models. Rate limits and charges live in your own console. Want to try a few newer models on QCode first? Pay-as-you-go Packs start at $4 for a one-time quota you use within 7 days.
Questions
Can I call SWE-2 through QCode?
No. The public model catalog has no SWE-2 and no Cognition or Devin service; it is only served inside the Devin product line (Desktop, CLI, with Web and Fusion rolling out).
Are those benchmark numbers trustworthy?
Look at the source: every one of those numbers comes from Cognition's own post, so it is vendor-reported. To their credit the weak column is in the table too — on Terminal-Bench 4, SWE-2 is clearly behind — and they did not publish only the flattering benchmarks; there is no third-party re-run.
What is its relationship to Kimi K3?
The post is direct about it: SWE-2 is post-trained from Kimi K3 with RL, and K3 itself is a Kimi team model. In the same table the K3 row and the SWE-2 row are different - K3 itself scores 44.2% on FrontierCode 1.1 Main - so the K3 score is not the SWE-2 score.
Is there an API, a price, a context length?
The official post 'Introducing SWE-2: Pushing the Pareto Frontier' gives no public API endpoint or model id, no price per million tokens, no rate limit and no context length. To get those numbers you have to wait for the vendor, or evaluate the cost through the Devin product subscription.
So which route for my own coding agent?
It depends on what is missing for you: if you want a product to run the whole loop for you — locate, patch, verify — take the product subscription route; if you want billable calls where you pick the model and control the context, sold general models plus tool calling plus outer validation and retries are enough to get started.
Can this price reduction be used as a unit price?
Not as a unit price. It is Cognition comparing its own cost to Fable 5.1, with no published pricing and no third-party recalculation, and the same sentence carries within a few points of GPT-6 Astra at a quarter of the cost - also vendor framing. Measure on your own tasks.
Sources
Cognition's official post 'Introducing SWE-2: Pushing the Pareto Frontier', published 2026-09-10 (fetched 2026-09-20): https://cognition.ai/blog/swe-2; the model list on this site follows the public catalog at qcode.cc/models, and the 30-day metered-usage table lists none of them either. Every score, percentage and training description here is vendor-reported.
Skip the agent product and still run coding work on billable calls
SWE-2 is not in our catalog (checked 2026-09-20); what you can use is sold general models plus tool calling, billed per token.
Read next
Claude Fable 5 on SWE-Bench Pro
The comparator Cognition measures against: where the number comes from and its limits.
Choosing AI coding models, H2 2026
Budget, complexity and context for building a coding agent on sold models.
QCode pricing guide
Official price times service rate, plus how caching and retries are billed.
Crypto payment guide
Nine coins, 56 networks, zero platform fee for AI coding API credits.
No affiliation with Cognition or Moonshot AI. The scores, percentages and training descriptions quoted here paraphrase the Cognition post fetched on 2026-09-20 and are vendor-reported; QCode does not offer SWE-2 and does not evaluate the Devin product. Information follows the official post, and model availability and per-model rates are listed on /models.