Which models do the real work, and what they cost
On Cognition's FrontierCode benchmark, Claude Opus 5.5 leads 46 models at 54.6%, spending $0.80 per task. The best value near the top is GPT-6.1 Sol: 92% of the leader's score for $0.36 a task.
SWE-2, Cognition's own model running in Devin's harness, places #8 at 50.0% for $1.18 a task.
Synced automatically · checked 2026-10-15 · figures last changed 2026-10-14
The FrontierCode leaderboard
Best score per model on the Main set (100 tasks), at whichever effort level scored highest, and what that run cost per task. The harder Extended set (150 tasks) is shown alongside.
| # | Model | Lab | Harness | Main score | Cost / task | Extended |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 medium | Anthropic | Claude Code | 54.6% | $0.80 | 65.3% |
| 2 | Claude Fable 5 xhigh | Anthropic | Claude Code | 53.5% | $13.09 | 64.9% |
| 3 | Claude Opus 5 medium | Anthropic | Claude Code | 53.4% | $4.31 | 63.6% |
| 4 | GPT-6 Astra max | OpenAI | Codex | 53.3% | $4.59 | 64.5% |
| 5 | Claude Sonnet 5.5 xhigh | Anthropic | Claude Code | 52.1% | $1.59 | 64.4% |
| 6 | Claude Fable 5.1 medium | Anthropic | Claude Code | 50.9% | $3.28 | 63.6% |
| 7 | GPT-6.1 Sol medium | OpenAI | Codex | 50.2% | $0.36 | 60.4% |
| 8 | SWE-2 max | Cognition | Devin | 50.0% | $1.18 | 62.5% |
| 9 | GPT-6 Sol max | OpenAI | Codex | 49.3% | $2.07 | 60.7% |
| 10 | Grok 4.6 high | xAI | Grok Build | 48.0% | $2.88 | 61.3% |
| 11 | Grok 4.7 high | xAI | Grok Build | 47.6% | $6.65 | 59.4% |
| 12 | Gemini 4 Argon medium | Chisel | 47.6% | $6.74 | 60.6% | |
| 13 | GPT-5.6 Sol max | OpenAI | Codex | 47.5% | $5.19 | 60.6% |
| 14 | Claude Opus 4.8 max | Anthropic | Claude Code | 46.5% | $9.62 | 59.6% |
| 15 | Claude Haiku 5.5 max | Anthropic | Claude Code | 46.4% | $1.33 | 58.4% |
| 16 | Kimi K3 | Moonshot | mini-SWE-agent | 44.2% | $3.82 | 58.2% |
| 17 | DeepSeek V4.1 Flash high | DeepSeek | Chisel | 43.9% | $0.62 | 56.6% |
| 18 | Gemini 3.7 Flash medium | Chisel | 43.6% | $1.82 | 56.3% | |
| 19 | GPT-5.5 xhigh | OpenAI | Codex | 43.0% | $4.03 | 56.7% |
| 20 | Claude Sonnet 5 xhigh | Anthropic | Claude Code | 42.7% | $10.07 | 56.2% |
| 21 | GPT-6 Luna max | OpenAI | Codex | 42.4% | $0.10 | 56.1% |
| 22 | Grok 4.5 high | xAI | Grok Build | 42.4% | $1.30 | 56.6% |
| 23 | SWE-1.7 | Cognition | Chisel | 42.0% | $1.97 | 54.3% |
| 24 | GPT-5.6 Terra max | OpenAI | Codex | 41.3% | $1.87 | 55.8% |
| 25 | Gemini 3.8 Flash medium | Chisel | 41.2% | $2.60 | 53.5% | |
| 26 | GLM 5.3 max | Zhipu | Chisel | 40.1% | $16.91 | 52.6% |
| 27 | GPT-5.6 Luna max | OpenAI | Codex | 39.8% | $0.37 | 55.1% |
| 28 | Mistral Large 4 (Preview) high | Mistral | Chisel | 39.4% | $4.13 | 53.2% |
| 29 | Claude Opus 4.7 max | Anthropic | Claude Code | 38.5% | $9.09 | 53.9% |
| 30 | Gemini 3.6 Flash medium | Chisel | 34.4% | $4.04 | 48.0% | |
| 31 | GLM 5.3 Flash max | Zhipu | Chisel | 31.8% | $1.15 | 46.2% |
| 32 | Kimi K2.7 | Moonshot | mini-SWE-agent | 30.1% | $3.01 | 45.4% |
| 33 | DeepSeek V4 Pro 0813 high | DeepSeek | Chisel | 28.6% | $1.81 | 44.5% |
| 34 | GPT-5.4-mini xhigh | OpenAI | Codex | 27.0% | $1.52 | 43.0% |
| 35 | Claude Opus 4.6 high | Anthropic | Claude Code | 26.6% | $3.98 | 43.7% |
| 36 | Composer 2.5 | Cursor | Cursor CLI | 25.6% | $3.09 | 40.8% |
| 37 | GLM 5.2 | Zhipu | mini-SWE-agent | 24.5% | $2.47 | 40.1% |
| 38 | Claude Sonnet 4.6 max | Anthropic | Claude Code | 24.3% | $2.90 | 40.0% |
| 39 | DeepSeek V4 Flash 0731 high | DeepSeek | Chisel | 18.8% | $1.53 | 31.7% |
| 40 | DeepSeek V4 Pro | DeepSeek | mini-SWE-agent | 17.6% | $1.55 | 31.0% |
| 41 | MiniMax M3 | MiniMax | mini-SWE-agent | 14.7% | $0.68 | 28.5% |
| 42 | Inkling | Thinking Machines | mini-SWE-agent | 14.0% | $3.60 | 24.8% |
| 43 | Nemotron 3 Ultra | NVIDIA | Chisel | 13.6% | $1.47 | 27.2% |
| 44 | Qwen 3.7 Plus | Alibaba | mini-SWE-agent | 10.2% | $0.24 | 21.8% |
| 45 | SWE-1.6 | Cognition | Chisel | 9.4% | $0.49 | 20.5% |
| 46 | Mistral 3.5 Medium | Mistral | Chisel | 8.0% | $1.35 | 18.7% |
Source: Cognition's FrontierCode leaderboard (data); method in the FrontierCode 1.1 post. We re-read the data every few hours and only republish when it parses cleanly, so a change upstream shows here the same day.
Live · refreshed every 3 hours
What Cognition is actually shipping
Every changelog entry and post we ingest, by product and theme. Unlike the benchmark below, this updates continuously — it comes from our own feed of Cognition's published sources, not a third-party snapshot.
- This month
- 9 -77%
- Last 90 days
- 93
- Tracked since
- 2024-03
Changelog entries per month, by product
Last 18 months · 593 entries tracked in total
Days since last ship
As of 2026-10-15
What they've been working on
Theme appearances across the last 180 days. Entries carry more than one theme, so these are appearances rather than shares.
Derived from Devin Central's own ingest of Cognition's published changelogs and blog (docs.devin.ai). Counts reflect published entries, which are a proxy for shipping activity — not a measure of engineering effort or code volume.
Method and caveats
- Synced automatically. We read Cognition's published leaderboard data every few hours. If it can't be read cleanly, or suddenly lists far fewer models, we keep the last good copy rather than publish a partial table. Last checked 2026-10-15; the figures last changed 2026-10-14.
- Score is the leaderboard's headline score on the Main task set. We show each model at the effort level where it scored highest, and the average cost per task of that same run. Cheaper effort levels of the same model usually score a little lower for much less money.
- Vendor benchmark. Cognition designed and runs FrontierCode and sells Devin, SWE-2 and SWE-1.x, which appear on it. We quote it because it is current; we do not adjust it.
- Cognition's Devin Fusion post (August 2026) charted Devin Fusion, its multi-model setup, at 63.1 on the Extended set; Fusion is not on the live leaderboard data, so it is not in the table.
Source and attribution
FrontierCode figures are Cognition's, from cognition.com/frontiercode.