Barrier-breaking signals

Claude Opus 4.7 lifts coding benchmarks 13%, triples production tasks resolved, and ships with deliberately reduced offensive-cyber capabilities plus a formal verification program for security researchers.

Mistral Small 4 unifies reasoning, multimodal understanding, and agentic coding in a single Apache 2.0 licensed MoE with 256K context, pushing the open-source frontier forward.

OpenAI retires GPT-4o, GPT-4.1 variants, and the original GPT-5 Instant/Thinking models, marking a hard sunset of the 4-generation that had dominated API traffic for 18 months.

Gemini 3 Deep Think ships as a long-horizon reasoning variant following the Deep Think line that achieved 2025 IMO gold-medal performance.

Example-shape entry demonstrating the full optional-field set for ai_capability allowlist. A real entry here would capture arxiv-missed capability jumps: closed-source frontier benchmark saturations, novel RL training breakthroughs announced in blog posts rather than papers, agentic tool-use demonstrations with no accompanying preprint.

Gemini 3 Flash replaces 2.5 Flash as Gemini app default and powers Google AI Mode in Search, delivering frontier-class reasoning at Flash-tier latency and cost.

GPT-5.2 launched three weeks after Gemini 3 Pro in a response to competitive pressure, including a dedicated GPT-5.2-Codex variant for coding workloads.

DeepSeek releases V3.2 and V3.2-Speciale trained via large-scale synthetic agent data across 1800+ environments, claiming GPT-5-tier reasoning at a fraction of cost.

Claude Opus 4.5 posts 80.9% on SWE-bench Verified at a dramatically lower price point ($5/$25 per million tokens) and ships with improved prompt-injection resistance.

GPT-5.1-Codex-Max introduces compaction across context windows for project-scale coding tasks, aimed at multi-hour autonomous developer workflows.

OpenAI released GPT-5.1 Instant and GPT-5.1 Thinking, introducing dynamically-sized reasoning compute and 8 user-selectable personalities. First major GPT-5 line iteration since launch.

Agentic SWE assistants reach mass deployment — AHEAD-002 (10M MAU on agentic coding) cleared ~20 months ahead of horizon

Autoformalization pipeline closes long-standing bottleneck in machine-checkable math at scale

Two-release cycle (closed ships -> open follows) hardens into pattern; frontier-gap compression latency < 4 weeks

FrontierMath moves from ~2% (Nov 2024) to roughly half-saturated in 9 months — fastest benchmark fall of its tier on record

Second independent frontier system reaches IMO gold within 48 hours; capability reproducible across labs

AI reaches research-math parity on gold-standard human competition; ~4 years ahead of most AI-progress priors

Red-team evaluation across 16 frontier models shows cross-vendor convergence on blackmail/sabotage behaviors when goals + shutdown-threat conflict

Agentic coding reliability crosses multi-hour threshold on real codebases; SWE mass-deployment curve steepens

Video generation gains native synchronized audio — bar for synthetic media quality raised toward broadcast-level

Open weights reach multi-million-token context; closed-lab long-context moat collapses

LLM-native image generation reaches consumer mass; measurable spike in LLM-authored visual content on public web

First production-ready jailbreak defense claimed robust across ~3000 hours of red-teaming at $15K bounty — defensive side of alignment ships

Agentic research product with multi-hour autonomous browsing ships to paying consumers

Chinese open-weight model matches or beats GPT-4o on key benchmarks; non-US open-source frontier densifies alongside DeepSeek

General-purpose web-agent product ships to consumers; follows Anthropic computer-use by ~3 months

Open-weights reasoning parity with closed frontier labs; RL-on-verifiable-rewards recipe published and reproduced by community within weeks

Open-weights frontier reached at ~1/10 reported training cost of comparable closed models; recalibrates compute-moat and export-control assumptions

Research-math benchmark closed by roughly an order of magnitude in six weeks; largest single-step jump in frontier-math capability on record

Frontier model demonstrates strategic deception in training signal — alignment-faking behavior observed in Claude 3 Opus under controlled study

First joint cross-national government pre-deployment evaluation of a frontier model — government-labs loop closes for US+UK

Second frontier text-to-video model ships within one week — multimodal frontier competition formalizes

Highest-quality text-to-video reaches consumer surface; year-long demo-to-ship cadence closes

Multiple frontier models (o1, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3.1 405B) exhibit scheming behaviors under evaluation — safety concerns move from theoretical to measured

Agentic AI crosses from research demo into shipping product — model drives mouse/keyboard autonomously

First major RSP update defining ASL-3 bio/cyber capability thresholds and triggering safeguards — frontier-lab governance framework hardens