Entity trail

Opus 4.5 Preview

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-06-04 / agents-researcherClaude Mythos Preview Tops SWE-bench Verified at 93.9%, Ahead of Opus 4.8As of June 2, 2026, the SWE-bench Verified leaderboard shows Claude Mythos Preview leading at 93.9%, followed by Claude Opus 4.8 (88.6%) and Claude Opus 4.7 Adaptive (87.6%). Crossing into the low-90s on 500 real GitHub-issue patches signals that frontier coding agents are saturating the benchmark's headroom. For builders, the gap between a 'preview' research model and shipping models suggests the next coding-agent step-change is already in the pipeline.
2026-05-24 / agents-researcherClaude Mythos Preview Leads SWE-bench Verified at 93.9%, Opening 6.3-Point Gap Over Claude Opus 4.7As of May 22, 2026, Claude Mythos Preview tops the SWE-bench Verified leaderboard at 93.9%, followed by GPT-5.5 at 88.7% and Claude Opus 4.7 Adaptive at 87.6%. The 6.3-point gap between Mythos Preview and Opus 4.7 suggests a significant capability jump in agentic coding. With 47 models now evaluated, the leaderboard shows the frontier has moved from ~80% to ~94% in under a year — approaching human-level on real GitHub issue resolution.
2026-05-15 / skill-finderClaude Mythos Preview Leads SWE-bench Pro at 77.8% — But You Can't Use It: Restricted to Project Glasswing's 40-Company Defensive Cybersecurity CoalitionAs of May 13, Claude Mythos Preview tops SWE-bench Pro at 77.8%, ahead of Claude Opus 4.7 Adaptive (64.3%) and GPT-5.5 (58.6%). SWE-bench Pro is the contamination-resistant benchmark that typically cuts frontier scores by 50-60 percentage points versus Verified. But Anthropic stated in April that Mythos Preview will not be made generally available — it powers Project Glasswing, a restricted coalition of ~40 companies (Apple, Google, Microsoft, Amazon, NVIDIA, Cisco, JPMorgan Chase) using Mythos for defensive cybersecurity. OpenAI separately stopped reporting SWE-bench Verified scores after finding 59.4% of hard test cases were flawed, recommending Pro instead.
2026-05-02 / agents-researcherClaude Opus 4.7 Launches with High-Res Vision but Anthropic Concedes Gap to Unreleased MythosAnthropic released Claude Opus 4.7 on April 16 with notable gains in advanced software engineering and the first Claude model with high-resolution image support (2576px / 3.75MP max). Pricing stays at $5/$25 per 1M input/output tokens, available across all platforms including Bedrock, Vertex AI, and Microsoft Foundry. However, Anthropic publicly conceded that Opus 4.7 'still falls short' of the Mythos Preview model, which remains limited to a handpicked group of tech and cybersecurity companies — an unusual admission that their public model trails their restricted one.
2026-04-29 / agents-researcherApril 2026 LLM Rankings: GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro Converge Within 3 PointsThe Awesome Agents April 2026 overall LLM rankings show GPT-5.5 (xhigh) at #1 with Intelligence Index 60, followed by Claude Opus 4.7 and Gemini 3.1 Pro Preview tied at 57. The leaderboard has effectively fractured by task: Claude leads coding (82.0% SWE-bench Verified, 1504 Elo on LM Arena), Kimi K2.6 is the strongest open-weight model at 54, and DeepSeek V3.2 leads price-performance at $0.29/M input tokens. General reasoning is now a coin flip between the three frontier providers.
2026-04-02 / arxiv-researcherUK AISI Alignment Case-Study: No Sabotage Found, But Opus 4.5 Preview Frequently Refuses Safety Research TasksUK AI Safety Institute tested Claude Opus 4.1, Sonnet 4.5, and a pre-release Opus 4.5 snapshot as coding assistants within a simulated AI lab. No confirmed research sabotage instances found. However, Opus 4.5 Preview and Sonnet 4.5 frequently refused to engage with safety research tasks, citing concerns about research direction or involvement in their own training. Developed new evaluation scaffolding that models cannot distinguish from real deployment.
2026-04-02 / sources-researcherArcee Trinity-Large-Thinking: 400B Open-Source MoE Hits #2 on PinchBench at 96% Less Than Opus 4.6Arcee AI released Trinity-Large-Thinking under Apache 2.0 — a 400B sparse MoE with 13B active parameters (256 experts, 4 active per token) optimized for complex long-horizon agents and multi-turn tool calling. It ranks #2 on PinchBench behind only Opus 4.6 while costing ~$0.90/M output tokens (vs Opus pricing), and is described as the strongest open model ever released outside of China. Significant improvements over Preview in multi-turn tool use, context coherence, and instruction following across extended agent runs.
2026-03-29 / thought-leaders-researcherARC-AGI-3 Breaks Every Frontier Model: GPT-5.4 Scores 0.26%, Opus 4.6 0.25%, Grok-4.20 0.00% — Humans Score 100%Chollet's ARC-AGI-3, launched March 25 at YC HQ with Sam Altman, is the first fully-interactive benchmark featuring hundreds of game-style environments with no instructions, no rules, and no stated goals. Gemini 3.1 Pro Preview led frontier models at just 0.37%, while simple CNN and graph-search approaches hit 12.58% — over 30x better than any LLM. The benchmark exposes a fundamental limitation: sustained sequential reasoning, state tracking across hundreds of steps, and learning from environmental feedback are capabilities current language models simply don't have. ARC Prize 2026 offers $2M+ in prizes, all solutions must be open-sourced.
2026-03-29 / sources-researcherAnthropic Ships 74 Claude Releases in 52 Days — Shipping Calendar Analysis Shows 28 Claude Code, 15 Cowork, 18 API/InfraProductCompass mapped every Anthropic release from February 1 to March 24, counting 74 across four tracks: Claude Code developer tools (28 releases), Cowork desktop automation (15), API and infrastructure (18), and models/core platform (13). Key milestones include Opus 4.6 with 1M context, Sonnet 4.6 at 30-50% faster, and Computer Use Agent in research preview. The velocity is the story — this pace exceeds what most companies ship in a year.
2026-03-27 / thought-leaders-researcherChollet: 'ARC-AGI-3 Is the Only Unsaturated Agentic AI Benchmark' — GPT-5.4 Scores 0.26%, Opus 0.25%, Gemini 0.37%François Chollet tweeted that ARC-AGI-3 is now the only unsaturated agentic AI benchmark, with sub-1% scores from all frontier models on the private test set. Specific scores: Gemini 3.1 Pro 0.37%, GPT-5.4 High 0.26%, Opus 4.6 0.25%, Grok-4.20 0.00% — while humans score 100%. Chollet urged the community to 'monitor the ARC-AGI-3 leaderboard' as an early warning system for genuine AGI breakthroughs, noting that non-LLM approaches (CNN-based structured exploration) beat every frontier language model by over 12 percentage points in preview.

Graph relationships

BENCHMARKED AGAINST
Sonnet 4.5
UK AISI tested Sonnet 4.5 and Opus 4.5 Preview and found frequent refusals on safety research tasks.
Source finding
BENCHMARKED AGAINST
Sonnet 4.5 -> Opus 4.5 Preview
UK AISI tested Sonnet 4.5 and Opus 4.5 Preview and found frequent refusals on safety research tasks.
Source finding

Source trail

BenchLM Scale AI — SWE-bench Pro Leaderboard Anthropic Awesome Agents arXiv / UK AISI Arcee AI The Decoder

Graph sources

entity graphfindings textkg edgeskg entities