SWE-bench / Morph
Public MindPattern findings, entities, and graph evidence that cite this source.
Findings
15
All-time hits
15
High value
6
Last seen
2026-06-30
Connected entities
MorphLLM — Context Engineering: Why More Tokens Makes Agents WorseContext Window Performance Degrades at 1M Tokens Regardless of Window Size — .claudeignore Alone AchMorph LLMSWE-bench Verified Contaminated: OpenAI Stops Reporting After Finding All Frontier Models Trained onMorphLLM — Claude Code Best PracticesClaude Code .claudeignore + Proactive Compact at 70% Context ThresholdMorph — Claude Code Context Window Management GuideClaude Code 70% Context Threshold: Proactive /compact Trigger Prevents 25% Output Quality DropMorphLLMCursor 2.5 Async Subagents: 35% of Cursor's Own Merged PRs Now Written by AI — GitHub Copilot MarketSWE-Bench Pro Exposes Benchmark Inflation: GPT-5.4 Leads at 57.7% vs 80%+ on Contaminated Verified
Related findings
- 2026-06-30 / AGENTSSWE-bench leaderboards hit new coding-agent state of the art as of June 29As of June 29, 2026, SWE-bench leaderboards show fresh state-of-the-art coding-agent results, with Claude's latest models leading — Opus 4.8 topping the active SWE-bench Pro board around 69.2%, and Fable/Mythos 5 leading the Verified board. The gains confirm continued rapid progress on real-world software-engineering tasks. Builders should compare agents on leaderboard deltas, not raw headline percentages across different test sets.
- 2026-06-29 / TOOLSExport-Control Directive Keeps Frontier Coding Models (Mythos 5, Fable 5) Offline; US Access Expected ~July 1A US export-control directive (effective June 12) required suspending access for foreign nationals, and because access can't be gated by nationality in real time, the top SWE-bench Verified coding models — Claude Mythos 5 (95.5%) and Fable 5 (95%) — went offline for all users, with US-based access expected back around July 1, 2026. For builders this means the leaderboard-topping coding models you read about are currently unusable, leaving Opus 4.8 (88.6% Verified) as the practical frontier. Plan tool/model fallbacks accordingly.
- 2026-06-28 / TOOLSOpen-Source Coding-Agent Landscape Reshuffles: Goose Joins the Linux Foundation, OpenCode Moves to anomalyco, Codex+GPT-5.5 Tops Terminal-Bench at 83.4%Governance moves are consolidating the CLI-agent field: Goose has moved to the Linux Foundation and OpenCode relocated to the anomalyco/opencode org, now citing 160K+ GitHub stars, 900+ contributors, and 7.5M monthly developers. On capability, Codex CLI paired with GPT-5.5 still leads Terminal-Bench at 83.4%, while Google retired Gemini CLI (consumer access ended June 18) and folded it into the Antigravity agent platform. For builders, the open-source tier is stabilizing around a few foundation-backed harnesses rather than fragmenting further.
- 2026-06-24 / TOOLSResource: Morph Publishes a Scored June 2026 Leaderboard of AI Coding AgentsMorph's 'Best AI Coding Agents (June 2026)' publishes a scored leaderboard ranking current coding agents on capability and price — a neutral reference point as Claude Code, Cursor, Devin Desktop, Codex, Kiro, Antigravity, and Copilot all ship competing agent modes. It's a useful starting point for builders deciding which agent to route a given task to, though as a vendor-adjacent ranking it should be sanity-checked against your own task mix.
- 2026-06-23 / DISPATCHMiniMax-M3 Tops Vendor-Reported Open-Weight SWE-Bench Pro at 59.0% — But Scale's Standardized Harness Tells a Very Different StoryMiniMax-M3 is being reported atop the open-weight SWE-Bench Pro at 59.0%, edging Kimi K2.6's 58.6%, but those figures come from vendor-tuned agent harnesses. On Scale AI's standardized leaderboard — identical scaffolding for every model — the top open-weights entry is qwen3-coder-480b-a35b at just 38.7%, a 10-to-30-point gap. The takeaway for builders: most of that delta is context-retrieval and tool-use quality in the harness, not raw model capability, so headline open-weight coding scores should be read against the scaffolding that produced them.
- 2026-06-19 / RESEARCHTerminal-Bench 2.1: Codex/GPT-5.5 Leads Coding Agents at 83.4%, Edging Claude CodeNewly posted Terminal-Bench 2.1 results (entries dated June 17) put Codex CLI on GPT-5.5 first at 83.4%, Claude Code on Fable 5 second at 83.1%, and Claude Code on Opus 4.8 at 78.9%. Notably, Fable 5/Mythos 5 remain export-suspended since June 12, so most users cannot run the top Claude configuration today — making Codex/GPT-5.5 and Opus 4.8 the practically available leaders. Useful real-world signal for builders selecting a coding agent.
- 2026-06-19 / PROJECTSOpenCode Crosses 172K Stars and Tops June 2026 Power Rankings, Displacing Cursor From #1OpenCode, the MIT-licensed, model-agnostic terminal coding agent, has crossed roughly 172K GitHub stars and about 7.5M monthly active developers, and topped LogRocket's June 2026 AI dev-tool power rankings — displacing Cursor from the #1 spot. It supports 75+ providers (Claude, GPT-5.5, Gemini, DeepSeek, Grok, and local models via Ollama) from a single CLI, treating the model as a pluggable dependency rather than the product. Its rise tracks the broader 2026 shift toward provider-agnostic harnesses.
- 2026-06-14 / TOOLSPattern: The CLI Coding-Agent Field Is Consolidating Around a Few WinnersJune 2026 head-to-heads show a clear shape: OpenCode has crossed ~150K–172K GitHub stars and ~6.5M monthly active developers to become the de facto open-source choice, Codex CLI on GPT-5.5 takes the benchmark performance lead, and Aider — still strong for Git-native pairing — is visibly slowing (last repo push 2026-05-22 vs daily pushes from OpenCode and Cline). For builders choosing a CLI agent, momentum and release cadence are now meaningful differentiators, not just raw capability.
- 2026-06-13 / PROJECTSGoogle's Antigravity 2.0 Harness Splits Into a Desktop App and Standalone CLIAntigravity 2.0, announced at Google I/O on May 19, 2026, restructured Google's agent harness into a redesigned desktop app plus a new standalone CLI, adding parallel-task subagents, cross-platform terminal sandboxing, credential masking, and hardened Git policies. In early June 2026, Google reset Gemini quota counters to zero for all free and paid users and shipped a refreshed Gemini 3.5 Flash build to fix post-launch issues. The sandboxing and credential-masking additions reflect harnesses maturing toward production-grade safety defaults.
- 2026-05-26 / SKILLSContext Window Performance Degrades at 1M Tokens Regardless of Window Size — .claudeignore Alone Achieves 80%+ Context Reduction on Typical ProjectsMorphLLM's analysis confirms models hit a clear performance ceiling around 1M tokens where irrelevant information actively worsens hallucinations. Three concrete mitigations: (1) use .claudeignore to exclude node_modules, build artifacts, lock files, and generated code — this alone achieves 80%+ context reduction; (2) use hierarchical CLAUDE.md files with subdirectory-level files for module-specific context instead of polluting project-wide context; (3) implement tool lazy loading to reduce context by up to 95% by loading tool definitions only when needed. Claude Code uses 5.5x fewer tokens than Cursor for equivalent tasks through better context management.
- 2026-04-14 / AGENTSSWE-bench Verified Contaminated: OpenAI Stops Reporting After Finding All Frontier Models Trained on DatasetOpenAI has stopped reporting SWE-bench Verified scores after confirming training data contamination across every frontier model. Claude Mythos Preview leads Verified at 93.9% but on the uncontaminated SWE-bench Pro benchmark, the best score drops to 57% (GPT-5.3-Codex) and 45.9% (Claude Opus 4.5 under SEAL standardized scaffolding). The 35-point gap between Verified and Pro scores exposes how contamination inflated the primary coding agent benchmark — SWE-bench Pro is now the reliable measure of actual agent coding capability.
- 2026-03-16 / SKILLSClaude Code .claudeignore + Proactive Compact at 70% Context ThresholdThree complementary token management tactics: create `.claudeignore` (gitignore syntax) to exclude build artifacts before they consume context, trigger `/compact` with a focus area at 70% context capacity rather than waiting for 100% to preserve coherent working state instead of forcing a cold restart, and delegate codebase exploration to subagents that return only relevant line ranges (saves 40%+ input tokens). Proactive compaction at 70% is the critical insight — it preserves essential state while summarizing noise at the point of maximum context coherence, not at the point of overflow.
- 2026-03-15 / SKILLSClaude Code 70% Context Threshold: Proactive /compact Trigger Prevents 25% Output Quality DropAnthropic's internal testing found Claude Code response quality degrades measurably at ~70% context utilization (~140K of 200K tokens), not at the advertised limit—with an average 25% quality drop beyond that threshold. The production practice is to run /compact at 65–70% fill (or every 60–90 minutes of active development), not waiting for the auto-compact trigger near 90% that lets quality degrade first. Claude Code v2.1.32+ automatically writes discovered project patterns to MEMORY.md during sessions, enabling higher-fidelity context rehydration after compaction by re-reading this file at session start.
- 2026-03-15 / MARKETSCursor 2.5 Async Subagents: 35% of Cursor's Own Merged PRs Now Written by AI — GitHub Copilot Market Share Collapses 42% to 25%Cursor's February 2026 release introduced async subagents that spawn nested sub-agents in isolated cloud environments — producing merge-ready PRs with no human intervention. Cursor's own engineering team reports 35% of merged PRs now come from these agents. GitHub Copilot's market share has collapsed from 42% to 25% as Cursor crossed $2B ARR, and Copilot responded by allowing paid users to run Claude, Codex, and Copilot simultaneously on the same issue to pick the best output.
- 2026-03-15 / AGENTSSWE-Bench Pro Exposes Benchmark Inflation: GPT-5.4 Leads at 57.7% vs 80%+ on Contaminated VerifiedSWE-Bench Pro (1,865 multi-language, uncontaminated tasks) reveals a dramatic performance gap vs SWE-Bench Verified (500 Python-only, contaminated). GPT-5.4 leads Pro at 57.7%, while top models score 80%+ on Verified. A separate analysis found that changing the evaluation harness moves benchmark scores by 22% independently of model choice, making harness selection a hidden variable that can eclipse model selection decisions for coding workloads.