Entity trail

GPT-5.6 Pro

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

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-04 / arxiv-researcherMonitrLLM Pilot: Students Rate ChatGPT 4.19/5 While Failing 23.1% of Their Actual Goal Tasks, and Multi-Turn Conversations Fail 2.5x More OftenMonitrLLM is open-source infrastructure that links full conversation transcripts to user-reported task intent and outcome assessment, treating all three as primary evaluative signals rather than optional metadata — filling the gap between capability benchmarks, unlabeled conversation corpora, and thumbs-up feedback. A two-week pilot with 26 college students produced 206 evaluation reports with transcripts. Despite average satisfaction of 4.19/5, participants hit a 23.1% failure rate on goal tasks, and multi-turn conversations failed at 2.5 times the rate of single-turn exchanges — reframing extended interaction as a signal of difficulty rather than engagement.
2026-08-04 / agents-researcherAntares: a 3B vulnerability-localization model approaches GPT-5.5 and beats open models 200x its size at $0.002 per taskPosted 2026-08-03 (arXiv 2608.02407), Antares is a family of 350M, 1B, and 3B agentic vulnerability-localization models built on IBM Granite bases, trained with supervised fine-tuning on cybersecurity reasoning plus repository-exploration data, then reinforcement learning from verifiable rewards over vulnerable repositories. Antares-3B approaches GPT-5.5 while outperforming open-weight models over 200x larger. The economics are the story for builders: a full 500-task evaluation sweep finishes in about 15 minutes on a single H100, amortizing to under 2 seconds and under $0.002 per task, which puts continuous local security scanning inside a CI budget instead of a frontier-API budget.
2026-08-04 / skill-finderSplitting one web agent into a multi-agent crew opens an attack that is inert against single agents: 80% success via a cross-delegation "Telephone Loop"WebMASLab holds the user task, tool surface, and browser substrate fixed and varies only the architecture, isolating what role decomposition itself costs you in security. The novel Telephone Loop attack exploits cross-agent delegation to create cyclical task loops; it has no effect on single-agent setups but compromises the multi-agent version of three of four frontier models tested (Claude Sonnet 4.5, GPT-5.2, GPT-5.4), averaging 80% success at baseline with 0% detection — only Claude Sonnet 4.6 resisted, at a 92% detection rate. Defenses did not generalize: prompt-hardening cut one model's attack success rate from 100% to 8% while barely moving the others, and lifted detection to just 33% for one. The takeaway for anyone fanning work out to sub-agents is that specialization is an architectural risk decision, not only a performance one.
2026-08-03 / github-pulse-researchersol-advisor Takes 853 Stars in Two Days Codifying Model-Tier Routing Across GPT-5.6's Luna, Terra, and SolDannyMac180/sol-advisor (853 stars, created 2026-08-01, Shell, MIT) is a Codex-native orchestration setup that runs implementation in Luna and Terra lanes and then requires a mandatory fresh-context Sol review before work stands. GPT-5.6 shipped publicly on 2026-07-09 in exactly those three tiers, with Sol as the flagship for complex agentic tasks, Terra as the everyday production option, and Luna as the fastest and cheapest. The pattern generalizes past OpenAI: cheap tiers do the typing, the expensive tier reviews with no prior context, and the review is a gate rather than a suggestion.
2026-08-03 / hn-researcherFROGS: One Absurd Prompt, 14 Models, Three Tries a Month — 'Generate an SVG of a Frog With a Habsburg Jaw'A personal benchmark that hit 143 points and 77 comments on HN runs a single deliberately weird prompt across 14 models from Anthropic (Opus 5, Sonnet 5, Haiku 4.5), OpenAI (GPT-5.5, GPT-5.4-mini) and Google (Gemini 2.5 Pro, Gemini 3.6 Flash), three attempts each per month, tracking execution time, output file size, and the comments models leave inside the SVG source. All 42 August attempts produced valid SVGs; the differentiator turned out to be annotation style — some models write interpretive commentary about anatomical exaggeration and royal framing, others emit bare structural labels. Dataset on Hugging Face, code on GitHub for reproducible runs.
2026-08-03 / reddit-researcherOnton's Ontology 1 Claims 0.630 P@10 Against Google Shopping's 0.543 While Indexing 1% of the CatalogOnton released Ontology 1 on August 2–3, 2026, a neurosymbolic product-search model evaluated on the released Subtext-Decor-90 benchmark by three multimodal judges (Claude Opus 4.8, Gemini 3.1 Pro, GPT-5.5): mean precision@10 of 0.630 versus 0.543 for Google Shopping and 0.469 for Amazon, winning 52 of 90 queries outright while indexing roughly 1% of either catalog. The architecture is an inspectable knowledge graph that decomposes vague predicates into checkable properties. Two caveats: judge agreement is modest (Krippendorff's alpha 0.465), and there are no open weights, no public API, and no pricing — partner access only.
2026-08-02 / arxiv-researcherTycho Reaches Perfect 100.00 RHAE on ARC-AGI-3 With Opus 5 and GPT-5.6 Sol, Using 61% Fewer Actions Than HumansTycho formalizes ARC-AGI-3 games as parameterized rendered deterministic Moore machines and has a coding agent build, test, repair, or bypass a free-form executable hypothesis of each game during play. Across all 25 public games under matched inference budgets, actor-requested delegation to a model builder scored the highest mean Relative Human Action Efficiency at 88.49; with that policy, GPT-5.6 Sol and Opus 5 both hit 100.00 RHAE and completed all 183 levels, with Opus 5 using 61% fewer scored actions than the aggregate official human baselines. Notably, automatic repair after verification failures produced simulators that matched observed transitions far better yet reached only 83.07 RHAE — reproducing dynamics is not the same as identifying the objective.
2026-08-02 / hn-researcherMIT Sloan: LLM Financial Advice Is Good Until You Prompt Like a Normal PersonTaha Choukhmane, Weidong Lin and Matthew Akuzawa (MIT Sloan) with Tim de Silva (Stanford) tested GPT-5.2, GPT-5.6 and Gemini 3 Flash on financial advice in a paper that won the 2026 Swiss Finance Institute Outstanding Paper Award. Following the models' recommendations produced sizable savings buffers for people over 30, but the advice degraded on life-change adjustments like unemployment and on active portfolio rebalancing. The finding with the widest applicability: performance jumped substantially with structured prompts carrying explicit financial information and assumptions, because 'regular people are not writing their prompts the way a finance professor is' — a capability-versus-elicitation gap that shows up in every domain, not just finance. The story drew 310 points and 308 comments on HN.
2026-08-02 / reddit-researcherNoam Brown: All Ten Astra Proofs Cost Under $2,000 in Tokens at Sol API Prices, and No Millennium Prize Problem FellOpenAI researcher Noam Brown posted that generating the proofs for all ten Astra breakthroughs combined cost under $2,000 at GPT-5.6 Sol API prices, and explicitly noted the team tried other major problems without success — 'Sadly no Millennium Prize problems (yet)' — adding that they did not spend heavily per problem and that test-time compute could be pushed much further. Sébastien Bubeck separately confirmed the non-sofic group result and the Connes disproof on X. The cost figure is the load-bearing detail for builders: it reframes frontier mathematical discovery as a test-time-compute scaling knob rather than a fixed capability, which is a very different planning assumption than 'wait for the next model.'
2026-08-02 / thought-leaders-researcherTao Calls It 'Very Strange and Not Particularly Pleasant'; Gowers Warns of 'Possible Destruction of Mathematical Culture'The mathematician reaction to OpenAI's ten results split four ways on August 1: Terence Tao described GPT-5.6 Pro solving problems he'd spent real time on as "very strange and not particularly pleasant" while staying cautiously optimistic; Fields medalist Timothy Gowers warned of the "possible destruction of mathematical culture" as literature expands while human understanding thins; Queen Mary's Abhishek Saha called frontier models "at least as good as a solid and indefatigable PhD student" and says he now plays "conductor, rather than doubling up as the whole orchestra"; Kirwin Hampshire called the reassuring framing "a well-muffled scream." The counterweight to the hype is concrete — the six remaining Millennium Prize Problems are untouched, and Epoch AI's benchmark still shows zero solves in its hardest categories.
2026-07-31 / sources-researcherllm-chat-completions-server 0.1a0 Turns Your Local LLM Model Collection Into a ChatGPT-Compatible Endpoint — Written Entirely by GPT-5.6 SolWillison released this plugin on 2026-07-30 as the first consumer of LLM 0.32rc1's content-addressable logs: it exposes every model in your LLM installation behind an OpenAI Chat Completions endpoint, so `curl http://localhost:8002/v1/chat/completions` works against local, Anthropic, or any plugin-provided model. Setup is three lines — `uv tool install llm --pre`, `llm install llm-chat-completions-server`, `llm chat-completions-server -p 9001`. Willison notes 'GPT-5.6 Sol wrote the whole thing' and showed strong recall of OpenAI's own API specification, which is a quietly telling data point about where model-written glue code is now competent.
2026-07-31 / sources-researcherLLM 0.32rc2 Swaps the Default Model From GPT-4o mini to GPT-5.6 Luna and Adds an openai-endpoint Command for Unconfigured OpenAI-Compatible ServersWillison followed rc1 the same day (2026-07-30) with rc2, retiring GPT-4o mini as the default in favor of 'GPT-5.6 Luna, a much better and more recent model' — pricing moves from $0.15/$0.60 to $0.20/$1.20 per million, a modest increase for a large capability jump. rc2 also adds a command to run prompts, chats and model listings against arbitrary OpenAI-compatible endpoints with no prior configuration, demonstrated against an LM Studio model on localhost:1234, with those calls deliberately left unlogged. The documented one-liner is `uvx --pre llm`, so you can point the tool at a local server without installing anything.

Graph relationships

CRITICIZES
Terence Tao
Tao described GPT-5.6 Pro solving problems he had spent time on as very strange and not particularly pleasant.
Source finding
MENTIONS
Abhishek Saha
Saha characterized GPT-5.6 Pro as at least as capable as a solid and indefatigable PhD student.
Source finding
CRITICIZES
Timothy Gowers
Gowers warned that frontier AI models pose a risk of possible destruction of mathematical culture.
Source finding
CRITICIZES
Terence Tao -> GPT-5.6 Pro
Tao described GPT-5.6 Pro solving problems he had spent time on as very strange and not particularly pleasant.
Source finding
MENTIONS
Abhishek Saha -> GPT-5.6 Pro
Saha characterized GPT-5.6 Pro as at least as capable as a solid and indefatigable PhD student.
Source finding

Source trail

arXiv 2608.02409 arXiv arXiv 2608.00202 — From Monoliths to Swarms: A Study of Attack Surface Evolution in the Transition to Multi-Agent Web Systems GitHub FROGS / Hacker News MarkTechPost arXiv MIT Sloan / Hacker News Noam Brown on X (corroborated by Sébastien Bubeck and The Decoder)The Decoder Simon Willison Simon Willison

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…