SWE-2: Cognition's Frontier-Near Coding Model Lands in Devin

On September 10, 2026, Cognition launched SWE-2 — the successor to SWE-1.7 and the company's closest model yet to frontier-tier coding performance at a fraction of the cost. The headline number is 50.0% on FrontierCode 1.1 Main — within one point of Fable 5.1 (50.9%) while Cognition claims 64% lower cost at that score level.

The launch lands one day after Marc Andreessen's second a16z Cognition investment post, which argued Devin now writes 90%+ of Cognition's own production code. SWE-2 is the engine behind that dogfooding story — cheaper, sharper rollouts inside the Devin harness.

TL;DR

What Cognition Claims — and What It Means for Builders

Cognition's pitch has not changed since SWE-1.7: optimize the entire cost–performance Pareto curve, not just chase a single leaderboard crown. SWE-2 is the first time they have pushed RL into the multi-trillion-parameter regime, building on the SWE-1.7 infrastructure (multi-continent rollouts, entropy control, verifier hardening) with a new algorithmic centerpiece.

The base is Kimi K3 — a 2.8T-parameter MoE model that had already undergone extensive RL for agentic coding. Cognition's additional pass still finds 5–6 points on many benchmarks, echoing the SWE-1.7 lesson that post-training headroom persists even after a heavily RL'd open base.

On Cognition's published table:

FrontierCode measures whether AI-written pull requests would merge — Cognition's own benchmark, so treat cross-vendor ordering as directional. DeepSWE 1.1 targets deeper repository work. Terminal-Bench 2.1 is the industry-standard terminal agent suite; SWE-2's 92.8% is the highest number in Cognition's table and a meaningful jump from SWE-1.7's 81.5%.

The soft spot is Terminal-Bench 4 at 27.3% — long-horizon agentic tasks where Fable 5.1 (55.8%) and GPT-6 Astra (57.9%) still lead by a wide margin. If your work looks like multi-hour autonomous repo refactors, SWE-2's headline FrontierCode parity does not automatically translate.

What People Are Asking

Is SWE-2 actually cheaper, or is this benchmark math?

Cognition's cost story is task-level, not a public per-token rate card. They plot mean USD per rollout against solve rate on FrontierCode and DeepSWE, claiming SWE-2 matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of the price and sits within a few points of GPT-6 Astra at roughly a quarter of the cost.

OfficeChai and other coverage note savings up to 70% at the top end across benchmarks Cognition tested — the 64% figure is specifically the Fable 5.1 FrontierCode parity point. Until Cognition publishes Devin task pricing tied to effort level, reproduce cost on your task mix inside Devin rather than extrapolating from blog charts.

Why does SWE-2 explore less than SWE-1.7?

SWE-1.7 drew user feedback for over-exploring simple tasks — thorough greps and reads before the first edit. SWE-2's efficiency gains come from focused exploration: stronger judgment about which parts of a codebase matter.

On FrontierCode 1.1 Main, SWE-2 medium scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average. Median first real code edit drops from step 48 (SWE-1.7) to step 18 (SWE-2 medium). High and max effort levels still plan longer and probe deeper — Cognition reports real behavioral separation across effort tiers, not just a speed knob on identical policies.

How is the training different from Kimi K3's approach?

Kimi K3 trains separate experts per domain × effort level, then consolidates via multi-teacher on-policy distillation with problem-specific token budgets. Cognition instead runs one RL pass for all effort levels using a linear cost-penalized reward:

R = S − λ_e C

where S is rollout success, C is cost (inference USD + time), and λ_e is tuned to the local slope of the base model's Pareto frontier at effort level e. Cognition derives this from first principles — only a linear penalty preserves the same expected reward whether you average cost before or after penalizing individual rollouts.

The geometric intuition: set λ_e equal to the frontier's tangent slope so iso-reward lines sit tangent to the curve. Too large a penalty and high-effort collapses toward medium-effort behavior — reward rises without moving the frontier. Matching the slope ensures gradient steps push the frontier upward rather than sideways.

They also publish the length-weighted group baseline used since SWE-1.6 — approximating the optimal variance-reducing baseline without an extra backward pass per rollout, which Cognition says significantly stabilizes training and keeps inference–training KL lower.

Does SWE-2 fix the long-horizon gap?

Partially, not completely. SWE-1.7 scored 7.6% on Terminal-Bench 4; SWE-2 jumps to 27.3% — a real improvement, but still roughly half of Fable 5.1 and GPT-6 Astra. Cognition tripled RL environments, added instruction-following overlays, and built a verifier flywheel from SWE-2 checkpoints to harden reward signals — infrastructure that should help long tasks over time, but the published numbers say frontier labs retain a moat on the hardest agentic horizons.

Training Infrastructure: What Scaled with K3

Moving from Kimi K2.7 (SWE-1.7) to K3 at ~3× the parameters required rollout-serving upgrades:

• Prefill delayer — batches nearby prefill requests in the GPU scheduler for 10–20% TPM/TPS gains at acceptable TTFT tradeoff
• DSpark speculative decoding with online draft-model training — acceptance rates degrade as the policy shifts during RL; SpecForge-trained drafts plus continuous online updates recover accept length
• NVFP4/FP8 kernels + quantization-aware training — lower memory, more concurrent rollouts, lower train–inference mismatch than SWE-1.7 despite the larger base

Data-side, Cognition tripled RL environments, expanded repo sourcing, layered instruction-following requirements onto existing tasks, and iteratively patched verifier false positives/negatives using rollouts from earlier SWE-2 checkpoints — a recursive flywheel Kimi K3's resourcefulness reportedly forced them to harden.

Behavioral Patterns Cognition Highlights

From internal testing, Cognition reports SWE-2 differs from SWE-1.7 in ways that matter for production agent work:

• Test coverage — end-to-end tests that catch regressions and edge cases, not just happy-path stubs
• Resourcefulness within boundaries — when an MCP integration was unavailable, SWE-2 reconstructed needed data from Slack channel history it already had access to
• Verification discipline — re-derives conclusions when challenged instead of agreeing; runs artifacts for evidence rather than trusting surface prose

On trustworthiness evals Cognition reran from their July post, SWE-2 passed 98.0% of propaganda/censorship attempts (145 politically sensitive prompts in English, Simplified Chinese, and Traditional Chinese). A context-dependent vulnerability framing test showed no statistically significant framing effects — same null result as their earlier open-source-derived model study.

What Hacker News pushed back on

The launch thread (393 points, 165+ comments as of publication) raised three objections worth stating plainly rather than smoothing over:

• The Terminal-Bench 2.1 vs TB4 gap reads as benchmaxxing to skeptics. One top comment framed it directly: TB2.1 is saturated (multiple frontier models score 88-92%+), so SWE-2's 92.8% there proves little, while its 27.3% on the newer, harder TB4 is the more honest signal — and it trails GPT-6 Astra's 57.9% by roughly half. Replies pushed back that every current model shows the same TB2.1-to-TB4 drop (GPT-5.6 Sol: 90% → 37%), so the pattern reflects benchmark staleness across the industry, not SWE-2 specifically.
• DeepSeek V4.1 Flash, released the same day, scored higher on TB4 (31.2%) at open-weight pricing — several commenters used it as the direct comparison Cognition's own post doesn't make, since SWE-2 ships closed-weight despite being post-trained on an open base (Kimi K3).
• Devin's product reputation is still recovering. Multiple commenters described Devin CLI as buggy or inconsistent between its desktop and cloud harnesses, a legacy of Cognition's much-criticized original 2024 demo; others reported the opposite experience after switching from Cursor or Windsurf. Take both sets of anecdotes as exactly that — anecdotes, not a resolved verdict — and verify on your own repos before treating either side as the final word.

SWE-2 vs the Agent Coding Landscape

Teams choosing between paths should ask whether they need self-driving ticket completion (Devin's unit of work) or IDE-native autocomplete plus commands (Claude Code, Cursor, Codex). SWE-2 only exists inside the first column.

Andreessen's post is the enterprise proof layer: Mercedes COBOL migration in eight days, Rivian 10× test generation, Itau 70% automatic vulnerability remediation. SWE-2 is the model refresh that makes those rollouts cheaper per attempt — if the efficiency metrics hold outside Cognition's eval harness.

Honest Limitations

• Self-reported evals — FrontierCode and DeepSWE are Cognition-owned; Terminal-Bench runs use vendor-preferred harnesses (Devin CLI for open-weight models, Claude Code for Anthropic, Codex for OpenAI). Cross-vendor ordering needs independent replication.
• Terminal-Bench 4 gap — 27.3% vs ~56–58% for frontier labs. Do not assume FrontierCode parity implies long-horizon parity.
• No open weights — you cannot audit SWE-2 weights, run ablations, or self-host. Compare Kimi K3 open weights if control matters more than Devin integration.
• Devin lock-in — evaluate against Claude Code, Cursor, and Codex on your repos before procurement.
• Cost claims are curves, not SKUs — Cognition assumes list pricing including public discounts; your Devin plan and task mix will differ.
• Andreessen dogfooding stats — 90%+ internal code from Devin is Cognition self-report, not third-party audit. Directionally interesting; not a universal customer benchmark.

The Bottom Line

SWE-2 is Cognition's answer to a question SWE-1.7 already posed: can RL post-training on an open base close the gap to closed frontier models without closed-model inference prices? At 50.0% FrontierCode within a point of Fable 5.1 and 92.8% Terminal-Bench 2.1, the scoreboard says yes for a large class of agentic coding tasks — especially when medium effort cuts turns and cost versus SWE-1.7.

The remaining gap is visible in the same table: Terminal-Bench 4 and the long tail of multi-hour autonomous work. Cognition is betting that Pareto-informed RL, verifier flywheels, and Devin harness co-design will keep closing that gap faster than frontier labs drop prices. Builders should verify on their repos, at their effort settings, with their security boundaries — not on blog charts alone.

Related on explainx.ai

• SWE-1.7: Cognition's frontier coding model at 1000 tok/s (Jul 2026) — predecessor, training recipe, SWE-1.7 benchmarks
• Marc Andreessen: Devin writes 90%+ of Cognition production code (Sep 2026) — a16z thesis and enterprise proof points
• Terminal-Bench 2.0: The AI Agent Benchmark That Actually Matters — benchmark context for SWE-2's 92.8% TB 2.1 score
• Agent Harness Engineering — why harness beats model alone
• Cognition Acquires Poke Maker Interaction — Devin messaging expansion
• How to read an AI benchmark and not get fooled — interpreting self-reported scores
• Claude Code vs Codex vs GLM routing guide — alternatives to Devin
• Loop engineering for coding agents — agent loop design patterns

Sources: Cognition — SWE-2 · Cognition — SWE-1.7 · Kimi K3 paper · Cognition on X (September 10, 2026).

Benchmarks, availability, and cost curves per Cognition's September 10, 2026 post — verify in Devin before procurement decisions.