AI Models News:
Frontier Release Tracker & 2026 Matrix

May 2026: three frontier launches in four weeks

The last 35 days delivered three frontier releases: Claude Opus 4.7 on April 16, GPT-5.5 on April 23, and Qwen 3.7-Max-Preview on May 20. That cadence — roughly one frontier launch every 10–14 days across the eight main labs — is the new normal. The Stanford HAI 2026 AI Index reports training compute now doubles every five months, datasets every eight months, and roughly 90% of notable models come from industry — up from ~60% in 2023. Read this section first; the dated matrix below is the long form.

Every frontier release, dated and sourced

Twelve dated releases from December 2024 through May 2026 — the full frontier wire across OpenAI, Anthropic, Google DeepMind, Meta, xAI, Alibaba, DeepSeek, and Mistral. Each row links to the primary source (provider announcement, release blog, or tracker). Reasoning and coding bars are normalized composite scores against published benchmark families (MMLU, GPQA, SWE-bench, HumanEval, MATH 500) — useful for scanning, never a substitute for a workload-specific eval.

Eight labs. Eight release cadences.

Each provider has carved out a distinctive release pattern. OpenAI ships step-changes; Anthropic prefers narrow lifts on Opus; Google maintains a Pro/Flash split with rapid Flash refreshes; Meta and DeepSeek anchor the open-weight frontier; Alibaba runs quarterly increments capped with API-only previews; xAI iterates in tenths; Mistral stays the European data-residency story. Spotlight summaries below, each with a primary source.

GPT-5 arrived August 2025; GPT-5.5 dropped April 2026 with incremental reasoning and code gains while keeping the same 400K context window and agent/tool-use scaffolding. Tooling remains the strongest of any closed lab.

Claude Opus 4.7 (April 2026) cites a 13% lift on coding evals and resolves roughly 3× more production tasks than its predecessor, with a 1M-token context window and explicit pricing of $5 / $25 per million input/output tokens.

Gemini 2.5 Pro and Flash reached general availability in June 2025 with thinking mode, ~2× faster inference than Gemini 2.0, and the largest mainstream context window at 2M tokens. May 2026 brought further Ultra-tier and Spark-Omni rollouts on the consumer side.

Llama 4 (Scout, Maverick) shipped April 2025 as the first MoE Llama family. Maverick claimed wins over GPT-4o and Gemini 2.0 Flash on multiple benchmarks; Scout headlined a 10M-token context window, the largest open-weight context published to date.

Qwen 3 (April 2025) shipped six dense and two MoE variants under Apache 2.0. Qwen 3.5 followed in February 2026, and Qwen 3.7-Max-Preview was announced at the Alibaba Cloud Summit on May 20, 2026 — API-only on Alibaba Cloud, no open weights yet.

DeepSeek V3 (December 2024) reset the conversation around training economics — frontier-grade performance at a claimed $5.6M training cost using a 671B-parameter MoE with 37B active per token. The model continues to drive open-weight cost benchmarks into 2026.

Grok 4 shipped July 2025; Grok 4.1 followed in November 2025 with incremental reasoning and coding gains. The 4.3 rollout in May 2026 widened availability across xAI's app and API surfaces.

Mistral Large 3 landed in December 2025 as the most capable model in the family to date, retaining the European data-residency narrative that drives much of its EU deployment story. Mid-tier weights continue to ship under permissive licenses.

The numbers behind the wire

Frontier benchmark scores are converging — every flagship now clears 88% on MMLU and 93%+ on HumanEval, so legacy multiple-choice tests no longer discriminate. The real contest has shifted to expert-reasoning evals (GPQA Diamond, Humanity's Last Exam) and agentic engineering benchmarks (SWE-bench, MLe-bench). Seven of the most-cited 2025–2026 numbers, each with a primary source.

Every model in the matrix, one subscription

ZeroTwo unifies more than 60 frontier models in a single chat. The full matrix above — GPT-5.5, Claude Opus 4.7, Gemini 2.5 Pro, Grok 4.1, Llama 4 (Scout / Maverick), Qwen 3 and 3.5, DeepSeek V3, and Mistral Large 3 — is accessible on one account. Switch models turn-by-turn inside the same thread; the conversation history travels with you. For context on how that compares vs single-vendor stacks, see our best AI platforms 2026 scorecard and the head-to-head AI models for text generation comparison. For the full tool catalog including image, video, research, and code, browse our AI tools directory.

• Rotating access to frontier models
• Daily message limits
• Multi-model picker

• 60+ frontier models
• Image, video, deep research
• All tools included

• 2× request limits vs Pro
• Higher-volume agent workflows
• Multi-model side-by-side runs

• Highest request limits
• Team and research workloads
• Priority access to new models

Mid-2026 outlook: reasoning, MoE economics, calibration

Three patterns will define the next 90 days. First, reasoning models that scale inference-time compute — Claude Mythos Preview already leads GPQA Diamond at 94.6%, and expect more o-series-style architectures from every major lab. Second, MoE cost economics will keep compressing the open-weight frontier behind the DeepSeek V3 cost curve — training compute is doubling every five months, but per-token inference cost is falling. Third, benchmark saturation forces the field toward harder evals: Humanity's Last Exam, agentic SWE-bench, and MLe-bench replace MMLU as the meaningful signal. The risk Edge AI Vision flags is calibration: even top models still report confident-sounding wrong answers.