Data as of November 5, 2026 · How the score is built
SWE-2 Flash
Decision readingSWE-2 Flash is tracked, but not publicly ranked yet. The profile exposes 8 sourced benchmark rows and leaves unsupported fields blank until a published record exists.
Released Nov 4, 2026 — see all recent releases
Decision snapshot
Each value carries a field reference instead of floating alone. Markers compare this model with the current ranked and priced catalog; they are not absolute quality thresholds.
Capability
Unranked
field median 50.1Not eligible for a public rank
Price
API rate not published
input median $0.65No comparable first-party hosted token rate
Speed
Not measured
field median 92 tok/sTime to first token not measured
Context
1Mtokens
field median 256,000Maximum output length is tracked separately
Strongest published evidence
Published rows are visible, but no category has enough eligible evidence for a comparative rank.
Validate before choosing
8 published rows leave some tracked benchmark slots empty. No comparable first-party API token rate is published.
Source-linked · 8 displayable benchmark rows
Category score record
Scores and ranks appear only where published evidence can be displayed. The table keeps the score, weight, cohort, and evidence state together.
| Category | Score | Rank | Percentile | Weight | Benchmarks | Evidence |
|---|---|---|---|---|---|---|
| AgenticWeight 22%2 benchmarksVerified | Score pending | 2 benchmarks | Verified | |||
| CodingWeight 20%5 benchmarksVerified | Score pending | 5 benchmarks | Verified | |||
| ReasoningWeight 17%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| MultimodalWeight 12%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| KnowledgeWeight 12%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| MultilingualWeight 7%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| Inst. FollowingWeight 5%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| MathWeight 5%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured |
8 of 802 tracked benchmark slots have displayable evidence · bars run 0–100
Coverage detailsHow much of this is verified
Coverage is split by category so a strong number never hides a thin evidence base. Verified means the row is tied to a published source; provisional rows remain visible but separate.
- Agentic2/2 verified
- Coding5/5 verified
- ReasoningNot measured
- MultimodalNot measured
- KnowledgeNot measured
- MultilingualNot measured
- Inst. FollowingNot measured
- MathNot measured
Benchmark ledger
Coding opens by default. The marker compares each value with the best source-verified result in the catalog; provisional leaders do not set the reference. Expand the remaining categories for every published row.
Coding5 rows
| Benchmark | Score | Versus best verified row | Gap | Weight | Evidence |
|---|---|---|---|---|---|
| DeepSWE | Score60.1% | Versus best verified row Best verified: Muse Wonder Full · 79.0% | Gap18.9 behind | WeightDisplay only | Provider exact |
| FrontierCode 1.1 Main | Score45.5% | Versus best verified row Best verified: Claude Opus 5.5 · 54.4% | Gap8.9 behind | WeightDisplay only | Provider exact |
| Terminal-Bench 2.1 | Score89.2% | Versus best verified row Best verified: SWE-2 · 92.8% | Gap3.6 behind | WeightDisplay only | Provider exact |
| SWE-Atlas Codebase QnA | Score50.0% | Versus best verified row Best verified: Ling 3.1 Flash · 64.5% | Gap14.5 behind | WeightDisplay only | Provider exact |
| FrontierCode 2.0 | Score24.0% | Versus best verified row Best verified: Claude Opus 5.5 · 45.5% | Gap21.5 behind | WeightDisplay only | Benchmark exact |
Agentic2 rows
| Benchmark | Score | Versus best verified row | Gap | Weight | Evidence |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | Score89.2% | Versus best verified row Best verified: SWE-2 · 92.8% | Gap3.6 behind | WeightDisplay only | Provider exact |
| Terminal-Bench 4.0 | Score20.00% | Versus best verified row Best verified: Claude Sonnet 5.5 · 70.60% | Gap50.6 behind | WeightDisplay only | Provider exact |
External signals1 row
| Benchmark | Score | Versus best verified row | Gap | Weight | Evidence |
|---|---|---|---|---|---|
| ExploitBenchExploitBench v8-bench | Score35% | Versus best verified row Best verified: GPT-6 Astra · 100% | Gap65.2 behind | WeightDisplay only | Provider exact |
Lineage
The sequence follows explicit supersedes links. Each score is estimated for that model; a relative can inform a sparse estimate but never sets a floor, so a newer release can score below an earlier one. Scores and prices remain blank when the corresponding public row or first-party rate is unavailable.
Spec sheet
Each documented value carries its source. Missing fields stay visible as not sourced or not published, rather than disappearing from the page.
- API model ID
- Not publishedCognition SWE-2 Flash launch post
- Context window
- 1MCognition SWE-2 Flash launch post
- Maximum output
- Not sourced yet
- Knowledge cutoff
- Not sourced yet
- Input modalities
- Not sourced yet
- Output modalities
- Not sourced yet
- Parameters
- Not disclosed by the provider
- Availability
- Released November 4, 2026 as Cognition's small, fast coding model: Inkling-Small (276B total, 12B active parameters) from Thinking Machines Lab with an added mid-training stage and large-scale reinforcement learning. Available in Devin Desktop and Devin CLI, free through November 30, 2026; Cognition says it costs one-fifth of SWE-2 and runs about 3x faster (200 vs. 60 tokens per second in its replay demo). No standalone API or per-token list price yet, and the weights are not public; Cognition plans a gated weight-sharing program for industry partners and research labs.
- Cloud regions
- Not tracked yet
- Lifecycle
- Current
- API capabilities
- Tool calling, structured outputs, and batch support are not tracked yet
- Prompt caching
- Not documented in the pricing recordCognition SWE-2 Flash launch post
- Self-host
- Weights are not published
- Rate limits
- Not tracked yet
How to read this profile
The visual layer above carries the decisions. These notes preserve the model, ranking, coverage, and family context behind the numbers.
We track SWE-2 Flash, but the public leaderboard excludes this profile until enough non-generated benchmark coverage is available. Only published rows appear above.
SWE-2 Flash is a proprietary model with a 1M context window. It uses an explicit reasoning mode, which can improve complex problem solving while adding latency and token use.
Released November 4, 2026 as Cognition's small, fast coding model: Inkling-Small (276B total, 12B active parameters) from Thinking Machines Lab with an added mid-training stage and large-scale reinforcement learning. Available in Devin Desktop and Devin CLI, free through November 30, 2026; Cognition says it costs one-fifth of SWE-2 and runs about 3x faster (200 vs. 60 tokens per second in its replay demo). No standalone API or per-token list price yet, and the weights are not public; Cognition plans a gated weight-sharing program for industry partners and research labs.
Official exact-value snapshot from Cognition's November 4, 2026 SWE-2 Flash launch post. BenchLM maps only the SWE-2 Flash column of Cognition's benchmark table onto existing lanes: FrontierCode 1.1 Main, DeepSWE 1.1, Terminal-Bench 2.1, Terminal-Bench 4.0, SWE-Atlas QA (the Codebase QnA lane) and ExploitBench. FrontierCode 2.0 comes from Cognition's FrontierCode 2.0 leaderboard data. The post gives no harness or trial details, so every launch row is provider-run and the model stays out of weighted ranking. SWE-Bench Pro v2 Hard (89.9) has no BenchLM lane and stays unmapped. The context window is inherited from the Inkling-Small base model because Cognition does not publish a separate figure. Cognition also reports the gains over the base model: Terminal-Bench 4.0 from 0.9% to 20.0%, FrontierCode 2.0 from 3.2% to 24.0%, and ExploitBench from 11.3% to 34.8%.
SWE-2 Flash sits in the SWE family with SWE-1.7, SWE-2. 8 of 802 tracked benchmark slots currently have displayable evidence. Missing categories stay blank.
Last updated November 5, 2026. Runtime fields remain blank until a sourced snapshot exists.
Questions
How does SWE-2 Flash perform overall in AI benchmarks?
SWE-2 Flash has 8 source-displayable benchmark rows, but it does not qualify for a public overall rank. The available rows remain visible by category without being converted into a site-wide score. Missing evidence stays blank instead of being estimated from an earlier model.
Is SWE-2 Flash good for coding and programming?
SWE-2 Flash has source-displayable benchmark coverage for coding and programming, but the public category table does not assign it a rank there. The individual rows remain available for inspection. A missing category position means the evidence threshold was not met; it does not convert the model's unmeasured work into a zero.
Is SWE-2 Flash good for agentic tool use and computer tasks?
SWE-2 Flash has source-displayable benchmark coverage for agentic tool use and computer tasks, but the public category table does not assign it a rank there. The individual rows remain available for inspection. A missing category position means the evidence threshold was not met; it does not convert the model's unmeasured work into a zero.
Which sibling models are related to SWE-2 Flash?
SWE-2 Flash belongs to the SWE family. Related tracked variants include SWE-1.7, SWE-2. A sibling link indicates shared lineage or a documented configuration relationship; it does not mean the variants have identical pricing, context limits, benchmark evidence, or deployment behavior. Compare before switching.
Does SWE-2 Flash have full benchmark coverage on BenchLM?
No. SWE-2 Flash currently has 8 source-displayable rows. The profile exposes published, non-generated evidence and leaves missing categories blank until an exact evaluation is available. Coverage describes how much was measured; it is not a penalty added to an individual benchmark result.
What is the context window size of SWE-2 Flash?
SWE-2 Flash has a documented context window of 1M. That figure is the maximum combined prompt and retained-conversation space reported for this exact model; it is not the maximum output length. The profile keeps output limits separate because providers often publish those limits independently.