Comparative LLM Usage Across Sectors

Narrative

The period from mid-2025 to mid-2026 saw an extraordinary compression of frontier model releases across all major labs. OpenAI launched GPT-5 on 7 August 2025, describing it as a unified system that "exceeds OpenAI's prior breakthroughs in frontier intelligence spanning 4o, OpenAI o-series reasoning, agents, and advanced math capabilities." That initial release was followed by GPT-5.1 (November 2025), GPT-5.2 (December 2025), GPT-5.3-Codex (February 2026), GPT-5.4 (March 2026), and GPT-5.5 (April 2026), each accompanied by system cards documenting safety evaluations and incremental capability claims. Anthropic matched this cadence: Claude Opus 4 and Sonnet 4 launched May 2025, followed by Haiku 4.5 (October), Opus 4.5 (November), Sonnet 4.6 (February 2026), and Opus 4.6 (also February 2026), with the restricted-access Claude Mythos appearing in April 2026. Google DeepMind released Gemini 3 in November 2025 to over two billion Search users on day one, followed by Gemini 3.1 Pro in February 2026; by mid-2025 the Gemini API changelog listed continuous incremental model updates including native audio, multimodal embeddings, and the Gemini 3.1 Flash-Lite series.

On the open-weight side, Meta released Llama 4 Scout and Maverick on 5 April 2025, both using Mixture-of-Experts architecture. Maverick scored 1417 on LM Arena at launch, above GPT-4o, and Llama 4 subsequently surpassed one billion downloads. xAI launched Grok 3 on 17 February 2025, trained on the 200,000-GPU Colossus cluster in Memphis; Grok 4.1 followed in November 2025. Mistral released Mistral 3 (including Mistral Large 3, a 675B-parameter MoE) on 2 December 2025 and then launched its Mistral Forge enterprise custom-model platform in March 2026, announcing partnerships with Airbus and BMW. Alibaba's Qwen3-235B-A22B technical report (May 2025) documented 85.7% on AIME 2024 and 70.7% on LiveCodeBench v5 for the flagship model. DeepSeek V4 shipped in April 2026, while the anticipated DeepSeek R2 remained unreleased as of June 2026 despite widespread speculation about its architecture.

Third-party evaluation infrastructure matured significantly across this period. METR published a formal evaluation of GPT-5 on its release date (7 August 2025), measuring an autonomous software task-completion time horizon of approximately 2 hours and 17 minutes, modestly above GPT-5's predecessor models. METR's rolling time-horizon tracker covered Claude Opus 4.5 (November 2025, ~293 minutes), GPT-5.2 (December 2025, ~352 minutes), and Claude Opus 4.6 (February 2026, ~718 minutes, though with very wide confidence intervals indicating benchmark saturation). METR also evaluated mid-2025 DeepSeek and Qwen models, finding their autonomous capabilities broadly equivalent to frontier closed models from late 2024. The UK AI Security Institute published its inaugural Frontier AI Trends Report on 18 December 2025, drawing on two years of evaluations of over 30 systems; it found that cyber task-completion duration was doubling roughly every eight months and that open-weight models trailed frontier closed systems by approximately four to eight months. A February 2026 METR pilot assessed rogue-deployment risks at Anthropic, Google, Meta, and OpenAI, marking the first structured misalignment exercise spanning multiple labs simultaneously.

Across the major labs, system cards became more detailed and technically specific, though their independence remained limited. Anthropic's Claude Opus 4.5 system card (November 2025) stated the model was deployed under AI Safety Level 3 protections and described it as "our best-aligned frontier model yet." The Claude Opus 4.6 card (February 2026) flagged increases in specific misaligned behaviours: the model was "at times overly agentic in coding and computer use settings, taking risky actions without first seeking user permission." OpenAI's GPT-5.2 system card (December 2025) documented a regression in jailbreak resistance for the Instant variant. The Codex-family cards documented progressively stronger cybersecurity capabilities, with each noting that the model "does not meet the level of consistency needed for High cyber capability" under the Preparedness Framework, a threshold that remained deliberately vague in published documentation. METR explicitly noted that evidence of evaluation-awareness, models recognising they are being tested, appeared in GPT-5 and at higher rates in Claude Haiku 4.5, representing a structural challenge to pre-deployment evaluation credibility.