SWE-1.6 is Cognition’s software engineering model for Windsurf, and as of September 2026 it is no longer the current one. Windsurf’s official model documentation now lists it as the previous generation, replaced by SWE-1.7 and SWE-1.7 Lightning.
It remains selectable, and it is still fast. But the more interesting thing about SWE-1.6 is how it was built: Cognition took the exact same pre-trained base as SWE-1.5 and got roughly ten points more on SWE-Bench Pro through post-training alone, with no new pretraining run. That is a result worth understanding even now that a newer model exists.
Where SWE-1.6 Sits in the Lineup
Windsurf’s model documentation currently lists a full family:
| Model | Role |
|---|---|
| SWE-1.7 | Current flagship software engineering model |
| SWE-1.7 Lightning | Same intelligence, lower latency, served on Cerebras |
| SWE-1.6 | Previous generation, optimised for intelligence and model UX |
| SWE-1.6 Fast | Faster SWE-1.6 for paying users |
| SWE-1-mini | Powers passive Tab suggestions, tuned for real-time latency |
| swe-grep | Handles context retrieval and Fast Context |
| swe-check | Powers Quick Review, lightweight checks for common issues |
That split matters. Cognition is not shipping one general model, it is shipping specialised small models for retrieval, review, and autocomplete alongside the main agent. Most of the cost in an agentic coding session goes on context gathering and verification, not on the reasoning itself.
What Actually Changed From SWE-1.5
This is the part worth your attention.
SWE-1.6 was post-trained from scratch on the same pre-trained model as SWE-1.5. Same foundation weights. Same inference speed. No larger parameter count, no fresh pretraining run, no new architecture.
From that identical starting point, Cognition reported an improvement of more than 10 percent on SWE-Bench Pro, with independent reporting putting the figures at roughly 50 percent against 40 percent for SWE-1.5. Specific numbers vary a little between the preview checkpoint and general availability, so treat the ten-point gap as the durable claim rather than any single percentage.
The company’s own framing was intelligence per token. The model does not think longer, it thinks less wastefully, completing tasks in around 40 percent fewer turns.
For anyone tracking where capability gains are coming from in 2026, that is the headline. The assumption that better coding models require bigger pretraining runs is looking weaker. A stable base plus better reinforcement learning produced a generational jump here.
The Length Penalty
The mechanism behind it is unusually concrete for a model release.
Cognition added a length penalty to the reinforcement learning reward, which discourages unnecessarily long trajectories. That single change cascaded into several of the behaviours users complained about.
Overthinking dropped, because rambling was penalised. Reasoning loops became rarer, for the same reason. And the model started using parallel tool calls far more often, because gathering context in one shot is shorter than gathering it sequentially. Cognition demonstrated this with a question about the PyTorch codebase, where SWE-1.6 answered faster than the preview by fetching context in parallel.
Their training charts showed response length growing much more slowly while task solve rate held steady, which is the outcome you want: same capability, less output.
One related fix deserves a mention because it is a daily annoyance rather than a benchmark. Cognition significantly reduced cases where the model reached for the terminal inappropriately, which is what forces Windsurf users to keep manually approving commands instead of letting the agent run.
The Benchmark Numbers, In Context
Cognition evaluated on SWE-Bench Pro, choosing it on OpenAI’s recommendation as the successor to SWE-Bench Verified.
Here is the part most coverage skipped. In Cognition’s own published comparison, SWE-1.6 did not top the table. They measured GPT-5.3-Codex at 54.0 percent at best across three harnesses, above their own model. They also tested Claude Opus 4.6 and Sonnet 4.6 at high reasoning effort across Claude Code, Cascade, and Devin, plus GLM-5, Kimi K2.5, and Composer-1.5 under Cursor’s CLI.
So SWE-1.6 was never pitched as the smartest available model. The pitch is speed, cost, and behaviour. If raw benchmark position is your criterion, our comparison of GLM-5.1 against Claude Opus 4.6 covers two of the models sitting in that same table.
Speed and Cost
SWE-1.6 shipped in two tiers. A free tier at 200 tokens per second, served through Fireworks, and a fast tier at 950 tokens per second for paying users, served through Cerebras. Cognition made the free tier available to everyone in Windsurf for three months at launch.
Nine hundred and fifty tokens per second is genuinely fast, and speed changes how an agent feels to work with more than a benchmark point does. A model that finishes while you are still reading the plan keeps you in flow. One that takes ninety seconds sends you to another tab.
Check current availability before planning around the free tier. That three-month window opened in May 2026 and has since closed, and pricing on the SWE-1.7 line is a separate question.
Credit Where It’s Due on Methodology
Vendor benchmarks usually deserve suspicion. Cognition’s disclosure was better than most, and it is worth saying so.
They manually read hundreds of agent trajectories and cross-checked against Scale-reported SWE-Bench Pro runs. They documented the specific problems they fixed, including dependency issues in the grading environment, inconsistent timeout handling across harnesses, edge cases in patch application, and out-of-memory failures during grading. They confirmed no repository overlap between their training tasks and the benchmark tasks, which is the contamination question that invalidates most coding benchmarks.
And they reported a competitor scoring above them, noting that their GPT-5.3-Codex measurement came out slightly worse than the figure OpenAI reported.
None of that makes it independent evaluation. It is still a company grading its own model. But publishing a number that beats you is not what marketing does.
Who Should Care Now
Windsurf users: SWE-1.7 is the default choice. Reach for 1.6 only if you hit a specific regression, and check whether SWE-1.7 Lightning covers your latency need first.
Anyone choosing a coding assistant: the model is one input and the harness is another. Cognition’s own evaluation showed the same model scoring differently across Claude Code, Cascade, and Devin, which tells you the surrounding agent matters as much as the weights. Our roundup of the best AI for coding covers that trade-off, and OpenCode versus Cursor looks at two of the competing environments directly.
People tracking where AI capability comes from: this is the interesting case. Same weights, better post-training, ten points. If that pattern holds, the gap between labs with enormous pretraining budgets and labs with good RL pipelines narrows considerably.
Frequently Asked Questions
Is SWE-1.6 still available?
Yes, it remains selectable in Windsurf, listed as the previous-generation model alongside SWE-1.6 Fast for paying users. SWE-1.7 is the current default.
Is SWE-1.6 better than Claude or GPT for coding?
Not on raw benchmark score. Cognition’s own published comparison put GPT-5.3-Codex above it on SWE-Bench Pro. Its advantages were speed, cost, and smoother agent behaviour.
Is SWE-1.6 free?
It launched with a three-month free window at 200 tok/s for all Windsurf users, which opened in May 2026. Check current plan terms rather than assuming that offer still stands.
The Practical Call
If you are in Windsurf, use SWE-1.7 and stop thinking about this. SWE-1.6 was superseded within months, which is the normal rhythm of this category now.
What is worth carrying forward is the lesson underneath it. Cognition demonstrated that a coding model can gain a full generation of capability from post-training on a fixed base, and that penalizing verbosity in the reward makes an agent both smarter per token and more pleasant to work with. Those two findings will outlast the model they came from, and they are a better reason to have read about SWE-1.6 than any benchmark number in it.
Published: September 10, 2026
